# Azgard — complete site content

> Azgard is a Sydney-based AI implementation agency. It builds production AI systems into the way a business actually runs — and leaves the team owning the system, not renting it. This document is the entire public website (https://www.azgard.tech) as one markdown file, for LLM consumption. The short index is at https://www.azgard.tech/llms.txt.

## About Azgard

**Positioning:** practitioners, not theorists. Azgard was founded to close the gap between organisations that talk about AI and organisations that actually use it.

**Founder:** Angus McDonald. 15 years in SaaS operations; HubSpot President's Club twice; former COO of a 50-person company; self-taught TypeScript, Python, LangGraph, and MCP; four production AI systems shipped with paying customers. Every engagement is delivered by Angus personally — no handoff to a junior.

**The Company Brain framework:** Azgard's core idea is that a business needs one structured, plain-text workspace holding what it needs to know, do, reach, and remember — a "Company Brain" that both the team and their AI tools work from. Azgard runs its own operations on exactly this system: it drafts proposals, runs end-of-week client communications, and briefs the business every morning. Everything Azgard sells either teaches this framework, builds it with you, or builds on top of it.

**Ownership philosophy:** every build is fixed-scope and milestone-priced (typically 30% on signing, 30% at mid-point, 30% on delivery, 10% held 30 days). Everything is documented, plain files, owned outright by the client. No platform lock-in, no dependency — the point is that it keeps working after Azgard leaves.

**Named customers:**

- Apex Signage
- Home Grown Electrical
- Inner Game Basketball

**Shipped portfolio products (proof of build capability):**

- GUS: public-sector agent stack handling 30k+ queries/month
- Grassmaster Gus: AI lawn-care coach — personalised turf advice from the Grassmaster Gus persona
- Euclid: deterministic compute tools powering LLM math
- FairyTaleFactory: personalised kids' stories, read aloud
- Tatara: markdown-native workspace where operators run AI agents against their own documents

**The ladder (how the offerings fit together):** start with the AI Company Brain Webinar (education, $149) → have the Kickstart build your Brain with you ($1,950) → run the one-day AI Readiness Audit to find where AI pays off → ship a Pilot Build (one workflow, 4–8 weeks) → scale to a Custom AI System (3–12 months) → keep it sharp with the Embed Retainer.

## Products (buy online at https://www.azgard.tech/shop)

### AI Company Brain Webinar

URL: https://www.azgard.tech/shop/ai-fundamentals
Format: Live · 2 hours · online
Price: $149 AUD (buy online; checkout localises to AUD/USD/GBP/CAD/NZD)

Two hours that turn “we should be using AI” into “here's exactly where we start.” The same grounding we run at the top of every Azgard engagement, now standalone, and a fraction of the price.

**Outcome:** the real picture, plus the skill that builds your Company Brain.

**Who it's for:** You're the person everyone asks about AI, and you're tired of guessing. Your team uses ChatGPT in scattered tabs with no system behind it. You want the real picture (what's hype, what isn't, and where it actually pays off in your business) without committing to a $20k engagement to find out.

**what we cover**

- The 5 levels of AI: an honest map, and where your team really sits today (most guess too high).
- The Company Brain: one structured place your team and your AI can both work from: what your business needs to know, do, reach, and remember. The same structure Azgard runs on.
- Where it pays off first: the highest-leverage jobs in your industry, not generic use cases.

**what you walk away with**

- scaffold-brain: the skill that builds your whole Company Brain for you. The same one Azgard runs on.
- A starter Skills pack you can use Monday morning.
- Your roadmap to Level 5: where you sit now, and how to get there.
- The recording and every template, yours to keep.

**What the price includes**

- 2 hours live with Angus
- scaffold-brain, the skill that builds your Company Brain
- Starter Skills pack + templates
- The recording, yours to keep

Everything above for $149. One payment.

**Refund policy:** Can't make it live? You still get the recording and the Starter Skills pack. Register anyway. All sales final.

**Why trust this:** Azgard runs on this exact system, the Company Brain that drafts our proposals, runs our end-of-week client communications, and briefs us every morning. On the call we open ours and show you the real thing. This isn't theory off a slide. It's the system we use to run the business, and you leave with the skill that builds your own.

**Your guide:** Angus McDonald, founder of Azgard. An operations leader who builds AI systems for the businesses he used to run by hand. The frameworks here are the ones he uses every day, not theory off a shelf.

**FAQ**

Q: Do I need to know how to code?
A: No. You're writing plain English. If you can write a document and follow a guide, you can do this.

Q: Is it live or recorded?
A: Live, so you can ask about your actual business. Everyone gets the recording and the Skills pack afterwards, so register even if you can't make it live.

Q: What tools do we use?
A: The principles transfer to any AI tool. We'll show you the exact stack we run on and what it costs (it's cheaper than you think).

Q: Can I send my team?
A: Yes. Each ticket is per seat. Bringing four or more from one business? Email us first.

Q: What's your refund policy?
A: All sales are final. We don't offer refunds. But if you can't make the live session, you're covered: we send you the full recording and the Starter Skills pack regardless, so you keep everything you paid for.

Q: What happens after?
A: You'll run scaffold-brain and stand up your own Company Brain scaffold, then start filling it. When you want it built with you (your context, your skills written against your real jobs, working by the end) that's the Kickstart. No pressure. Most people start with the skill and come back when they want it done with them.

---

### AI Company Brain Kickstart

URL: https://www.azgard.tech/shop/ai-brain-kickstart
Format: Done-with-you · 2 sessions · workspace built with you
Price: $1,950 AUD (buy online; checkout localises to AUD/USD/GBP/CAD/NZD)

A done-with-you build. Two sessions with Angus. We build your Company Brain together (context, skills, and the workspace that runs them), and you leave with a working system, not a list of homework.

**Outcome:** your Company Brain, built with you. working by the end.

**Who it's for:** You've seen the framework (the webinar, or you already get it) and you want it built. Not by you, alone, over six weeks of evenings. With Angus, in two sessions, working by the end.

**How it runs**

1. Pre-session research + questionnaire: You fill a short form. We research your business before we meet.
2. 90-minute build session: We build the workspace live: your context drawer (company, team, services, tone), your first skills, the folder structure your team can plug into.
3. 5-7 days of real use: You run the Brain on real work. We're on email if you get stuck.
4. 60-minute review session: Refine what's working, fix what isn't, ship the next few skills.
5. 30 days of email support: Questions and refinement as you keep building.

**what you walk away with**

- A custom Company Brain workspace, built and handed over, yours to own and extend.
- Your context drawer, built live (company brief, team, services, tone).
- A tailored Skills pack: your first 3-5 skills, authored against the real jobs in your business.
- The folder structure that lets your team plug in.
- 30 days of email support.

**What the price includes**

- 90-minute build session with Angus
- 60-minute review session
- Pre-session research + questionnaire
- Custom workspace, built and handed over
- Tailored Skills pack (3-5 skills)
- 30 days email support

Everything above for $1,950. One payment.

**Why trust this:** This is the same system Azgard runs itself on, the one that drafts our proposals and runs our weekly client comms. We're not handing you a template and wishing you luck. We build yours, with you, against your jobs.

**Your guide:** Angus McDonald, founder of Azgard. Operations leader turned AI builder. Every Kickstart is delivered by Angus personally, no handoff to a junior.

**FAQ**

Q: Do I need to have done the webinar first?
A: Strongly recommended. The Kickstart applies the framework the AI Company Brain Webinar teaches, but if you already know it, come straight in.

Q: What's the gap between the two sessions?
A: Usually 5-7 days. You use the gap to run your new workspace for real, then we review and refine in session 2.

Q: Do I need to code?
A: No. It's plain-English files the AI reads and follows. We set up the tools with you on the call.

Q: Who owns what we build?
A: You do. Completely. Plain files, no lock-in. The whole point is that it keeps working after we're done.

---

### AI Readiness Audit

URL: https://www.azgard.tech/shop/ai-readiness-audit
Format: Full-day intensive · on-site or remote
Price: scoped and quoted on a call (book via https://www.azgard.tech/contact)

Before you spend on a build, know exactly where AI pays off, and where it won't. One full day in your business, end to end: we map your real workflows, score what's ready, and hand you a 90-day plan you could start Monday. The fastest way to turn “we should do something with AI” into a costed, prioritised list.

**Outcome:** one day in your business. leave with the costed roadmap.

**Who it's for:** You're past the hype and into the where. You suspect AI could take real cost out of your operation, but you don't want to bet $50k on a hunch. You want a clear-eyed read on what's worth doing, in what order, with the numbers attached.

**what's included**

- AI Maturity Scorecard: where your team and tooling honestly sit today.
- Prioritised Opportunity Map: every candidate, plotted effort vs. impact, with dollar estimates.
- Technology & Data Readiness Assessment: what's ready to build on, what needs work first.
- 90-Day Action Plan: the sequence: what to pilot first, and why.
- Executive Summary Deck: the one-pager your board or ELT signs off.

**How it runs**

1. Before the day: We scope the audit and get read access to the workflows that matter.
2. The day: discovery: Interviews, process walk-throughs, a look at the real work, not the org chart.
3. The day: analysis + scoring: We model the opportunities and the ROI, and pressure-test the data.
4. The readout: We present the roadmap live, then hand over every deliverable.

**what you walk away with**

- A costed, prioritised roadmap: the single most useful artefact for deciding what to build.
- An honest readiness read, including the things not to do yet.
- A Skills pack seeded from the audit: quick wins your team can run while the roadmap is decided.
- A clear, no-obligation path into a Pilot if the numbers justify it.

**What you're paying for:** A senior, hands-on diagnostic that pays for itself by stopping you building the wrong thing. The clarity of a Big-4 or Accenture-equivalent diagnostic, without the seven-person team: strategic depth delivered hands-on, by the person who'll build what comes next.

**Why trust this:** The Audit is how Azgard finds the 70%-cost-reduction jobs hiding in a business: the manual process nobody questions because “that's just how we do it.” We build Azgard's own roadmaps the same way.

**Your guide:** Angus McDonald, founder of Azgard. Operations leader turned AI builder: he reads a business from the workflow up, not the slide down, because he's run the operations he now automates.

**FAQ**

Q: How is this different from a strategy deck?
A: A deck tells you AI is important. The Audit tells you which three workflows, in what order, for what return, with the readiness to back it.

Q: Will you just recommend your own build?
A: Only if the numbers say so. Part of the value is an honest “don't build that yet.” The roadmap is yours either way. No obligation to engage us.

Q: How long until we see it?
A: It's one intensive day, with the readout live at the end of it.

Q: What do you need from us?
A: A few hours of the right people's time, and read access to the workflows in scope. We do the rest.

## Services (scoped on a call — https://www.azgard.tech/work)

### Pilot Build

URL: https://www.azgard.tech/work/pilot-build
Format: Build · 4-8 weeks · one workflow, shipped
Price: scoped and quoted on a call (book via https://www.azgard.tech/contact)

One workflow. One measurable outcome. Built, shipped, and owned by your team in weeks, not a proof-of-concept that dies in a sandbox. The fastest way to turn a line on your Audit roadmap into cost off your P&L.

**Outcome:** one workflow, shipped & owned.

**Who it's for:** You know the workflow that's bleeding time, the one everyone complains about, and you want it fixed, properly, in production. You're done with pilots that demo well and then quietly disappear. You want a working system your team runs after we leave.

**what we build**

- One real workflow: integrated with the tools you already use.
- In production: with your team using it on real work, not a demo.
- A Skills pack: your team owns and can extend to the next workflow.
- The measurement: to prove the outcome, so the next build is an easy yes.

**Example builds** — The kind of pilot we build, and the outcome each one targets (illustrative, not a claim of past work):

- Tender / bid response assistant: aims to cut bid prep from ~2 weeks to ~3 days.
- Defects / snagging classifier from site photos: automated first-pass QA review.
- Contract & variation clause extractor: targets ~60% less legal review time.
- Email-to-project-system auto-filer: zero manual filing.
- RFI triage & drafting assistant: faster, more consistent turnaround.

**How it runs**

1. Scope + fixed fee: We agree the one outcome and the price up front. No hourly meter.
2. Build in the open: Weekly working sessions; you see it take shape, not a big reveal at the end.
3. Ship + adopt: It goes live, your team uses it, we sit with the rough edges.
4. Handover: Documented, owned, and yours, with a 30-day warranty.

**What you're paying for:** A fixed-fee, milestone-based build: typically 30% on signing, 30% at mid-point, 30% on delivery, 10% held 30 days. You're paying for a system that earns its keep, not a retainer that never ends. We scope it to one outcome and show you the payback maths on the call, before you commit.

**the terms:** Fixed fee, fixed scope, clear milestones. Change requests are quoted, never surprise-invoiced. 30-day warranty on delivery. Everything built is documented and owned by you.

**Why trust this:** Every system Azgard runs on itself started as exactly this: one workflow, built to actually work, then extended. We build yours the same way: ship one thing that earns its keep, then let the wins fund the next.

**Your guide:** Angus McDonald, founder of Azgard. He scopes to one measurable outcome because he's been the operator holding the P&L, not just the consultant holding the deck.

**FAQ**

Q: Why not just buy an off-the-shelf AI tool?
A: Generic tools make you bend your process to fit them. We build the system that fits the process you already run, and you own it.

Q: What if we haven't done an Audit?
A: Fine if you already know the workflow and the outcome. If you don't, the Audit is the lower-risk way to find the right first build.

Q: Who owns and maintains it?
A: You own it outright: plain, documented files, no lock-in. Want us to keep optimising it? That's the Embed retainer, entirely optional.

Q: How fast?
A: 4-8 weeks from signing for a single workflow.

---

### Custom AI System

URL: https://www.azgard.tech/work/custom-ai-system
Format: Build · 3-12 months · multi-workflow system, wired into your business
Price: scoped and quoted on a call (book via https://www.azgard.tech/contact)

When one workflow isn't the answer: when it's the way a whole department runs. We build the AI system that connects two, three, a dozen workflows, wired into your ERP, CRM, and project tools, on a data foundation that holds. The “we'll build it for you” end of the ladder: your operating system, owned by you.

**Outcome:** your operating system. built for you, owned by you.

**Who it's for:** A single pilot won't move the needle. You need a department, or the business, to run differently. You have real systems to integrate (ERP, CRM, project management) and real data to build on. You want a partner who'll architect the whole thing, not a vendor who'll sell you a seat licence.

**what we build**

- Multiple integrated workflows: that hand off to each other, not point solutions in isolation.
- Real system integration: wired into your ERP, CRM, project, and finance tools.
- A data foundation: the system runs on and your team can trust.
- A growing Skills library: your team owns, governs, and extends across the business.
- The operating model: to run it: who owns what, how it's measured, how it stays current.

**How it runs**

1. Architecture + roadmap: We design the whole system and phase the build so value lands early, not at the end.
2. Phased delivery: Each phase ships a working capability into production. You're never waiting 12 months for a big bang.
3. Integration + hardening: Real connections, real data, real edge cases: tested, monitored, documented.
4. Capability handover: Your team is trained to run and extend it, because dependency isn't the product.

**What you're paying for:** A senior architecture-led build, fixed-fee and milestone-based (30 / 30 / 30 / 10). Scope runs from a single integrated multi-workflow system to a platform-scale, multi-department rollout on a custom data foundation. Where the Big-4 and global SIs staff this five-deep and bill by the hour, we deliver it hands-on and you own the result outright. We architect, scope, and cost it with you on a call.

**the terms:** Fixed fee, phased milestones, value in production early. Everything architected, documented, and owned by you: no platform lock-in, no dependency on us to keep it running. A defined handover is written into the engagement.

**Why trust this:** Azgard is itself a custom AI system: a business run on context, skills, and automation rather than headcount. We build at the scale we operate at: the same architecture, applied to your operation instead of ours.

**Your guide:** Angus McDonald, founder of Azgard. He architects from the operations up: the system has to survive contact with how the business actually works, because he's run businesses that lived or died on exactly that.

**FAQ**

Q: Is this a retainer?
A: No. It's a fixed-scope build with a defined end. Want ongoing optimisation after? That's the Embed retainer, and it's optional.

Q: Do we need a finished Audit first?
A: For a build this size, yes. The Audit (or equivalent scoping) is how we architect the right system and phase it sensibly.

Q: Will our team be locked out of it?
A: The opposite. Plain, documented, owned by you, with your people trained to run and extend it. The whole point is that it keeps working after we leave.

Q: How do you de-risk something this big?
A: Phased delivery: each phase ships real capability into production, so you're proving value continuously, not betting it all on a final reveal.

---

### Embed Retainer

URL: https://www.azgard.tech/work/embed-retainer
Format: Retainer · monthly · ongoing optimisation & a fractional AI lead
Price: scoped and quoted on a call (book via https://www.azgard.tech/contact)

The build ships, but the work doesn't stop: models change, your business changes, and the system should keep getting better. Embed is the ongoing partnership that keeps your Company Brain sharp: monitored, tuned, and growing. At the top tier, a fractional AI lead embedded in your team. This is what makes “AI-native” durable instead of a one-off.

**Outcome:** durable capability, not dependency.

**Who it's for:** You've built something that works and you don't want it to rot. New jobs keep surfacing that the system could take on. You'd rather have a senior AI partner on tap than hire a full-time AI lead you can't yet justify, or watch a great build quietly fall behind.

**the tiers**

- Guardian (8-12 hrs/mo): System monitoring, prompt tuning, minor fixes, and a monthly performance report. Keeps what you've built running clean.
- Partner (15-25 hrs/mo): Everything in Guardian, plus strategic advisory, quarterly roadmap reviews, and priority support. Keeps the system improving, not just running.
- Embedded (30-50 hrs/mo): A fractional AI lead inside your team: drives the roadmap, attends your standups, builds hands-on. The capability of a senior AI hire, without the hire.

**what you walk away with (every month)**

- A system that stays current as models and your business change.
- A growing Skills pack: new jobs automated as they surface.
- A monthly performance report: what changed, what it saved, what's next.
- A senior AI partner who already knows your operation.

**What you're paying for:** Durable capability, not dependency. How deep we go scales with your needs: from light-touch Guardian to a near-full-time Embedded lead. Terms: minimum 3-month commitment; scope flexes up or down on 30 days' notice, and longer commitments earn better rates. We size it with you on a call.

**the terms:** Minimum 3 months, then month-to-month. Unused hours don't roll over, but utilisation is reported monthly. You always see what you're getting. Scope flexes with 30 days' notice. No lock-in beyond the initial term.

**Why trust this:** Azgard's own Company Brain is never “finished”: it gets a new skill the week a new job shows up, and a tune-up the week a new model lands. Embed gives your system the same standing care, from the people who built it.

**Your guide:** Angus McDonald, founder of Azgard. As your Embedded lead he works as part of your team, not a ticket queue: he knows the system because he built it, and he knows the operation because he takes the time to.

**FAQ**

Q: Do I need a retainer after a build?
A: No. Every build is owned and documented so it runs without us. Embed is for businesses that want it to keep improving, not just keep working.

Q: What's the difference between Partner and Embedded?
A: Partner keeps the system sharp and advises the roadmap. Embedded is a team member: in your standups, driving and building.

Q: Can we change tiers?
A: Yes. Flex up or down on 30 days' notice as your needs change.

Q: Is this just support hours?
A: No. It's a senior partner who knows your system and your operation, focused on making both better month over month.

## Learn — articles (https://www.azgard.tech/learn)

Field notes on production AI: strategy, agents, operations and security, written by practitioners who ship and run real systems.

### Is ChatGPT safe for company data in Australia? The account decides

URL: https://www.azgard.tech/learn/is-chatgpt-safe-for-company-data-au
Published: 2026-08-25
Topic: security · Tags: security, privacy, governance, tooling · ~12 min read

> Is ChatGPT safe for company data? The account tier, not the model, decides whether your prompts train it, how long they are kept and who can read them.

**TL;DR**

- The Azgard AI Data Boundary sets four classes of company data - public, internal operational, personal information about customers or staff, and client-confidential or regulated - against three account tiers: personal consumer, paid business subscription, and contracted workspace or API.
- ChatGPT Plus and Pro are paid consumer subscriptions, so paying for ChatGPT does not stop OpenAI using what staff paste to train its models; the no-training commitment starts at the business tiers, not at the paywall.
- If nobody in your business can name the account tier a staff member is using an AI tool on, that staff member is on the consumer tier.
- Australian privacy law does not require personal information to stay in Australia, but APP 8 and section 16C of the Privacy Act 1988 leave the Australian business liable for an overseas recipient's breach even after it has taken reasonable steps.
- AI data residency in Australia covers storage at rest only: Australia is not on OpenAI's inference-residency list, so an Australian ChatGPT Enterprise workspace stores conversations here while the model itself runs offshore.

ChatGPT is safe for some company data and not for other company data, and the account
tier decides which, not the model. On a personal consumer account, free or Plus or Pro,
OpenAI [may use what you paste to train its models](https://help.openai.com/en/articles/5722486-how-your-data-is-used-to-improve-model-performance).
On a business or contracted account it does not. The Azgard AI Data Boundary sets four
classes of company data against those three tiers, so you can tell before you paste.

#### Why "is ChatGPT safe for company data" is the wrong question

"Is ChatGPT safe for company data" asks about the model, and the model is identical in
every account. What differs is the contract wrapped around it.

> The rule behind the Azgard AI Data Boundary: the model does not decide what happens to
> your data. The account does. Name the tier you are on, read three lines of that tier's
> published terms - training, retention, human review - and match them against the most
> sensitive class of data you intend to paste. If you cannot name the tier, you are on
> the consumer one.

There is a real objection to that. A free ChatGPT user can turn training off in Settings
under [Data Controls](https://help.openai.com/en/articles/7730893-data-controls-faq), so the
consumer account is not unconditionally training on your work. But that toggle is one
person's preference. Nobody else in the business can see whether they flipped it, or prove
to a client it stayed flipped.

#### The three AI account tiers, and what each one's published terms say about your data

Three AI account tiers exist across every major vendor, and the paywall is not the line
between them. OpenAI's commitment that ["by default, we do not train on any inputs or
outputs from our products for business users"](https://help.openai.com/en/articles/5722486-how-your-data-is-used-to-improve-model-performance)
names ChatGPT Business, Enterprise and the API. Plus and Pro are consumer products.

| Account tier | Trains on your content | Retention | Admin visibility | Processor agreement |
|---|---|---|---|---|
| Personal consumer: free, Plus, Pro | Yes, unless the user switches it off | Deleted chats gone [within 30 days](https://openai.com/index/response-to-nyt-data-demands/) | None, and no audit trail | No |
| Paid business subscription: ChatGPT Business | No, by default | [Admins control how long data is retained](https://openai.com/enterprise-privacy/) | Admins read, export and delete staff chats | Yes, the DPA applies |
| Contracted workspace or API: Enterprise, Edu, API | No, unless you agree in writing | [Admin-set; deleted conversations gone within 30 days, API logs after 30 days](https://openai.com/enterprise-privacy/) | [Compliance API audit log](https://openai.com/enterprise-privacy/) | Yes, plus confidentiality terms |

At the contracted tier the promise becomes a clause. OpenAI's
[Services Agreement](https://openai.com/policies/services-agreement/) says it "will not use
Customer Content to develop or improve the Services, unless Customer explicitly agrees",
treats that content as Confidential Information, and deletes all of it within thirty days of
termination. Its [data processing addendum](https://openai.com/policies/data-processing-addendum/)
is incorporated rather than sold separately.

Admin visibility is the variable owners miss. On ChatGPT Business, workspace admins can
["view, access, export, and delete end user conversations"](https://openai.com/enterprise-privacy/),
and Microsoft 365 Copilot works the same way: prompts and responses
[are stored and searchable through Purview](https://learn.microsoft.com/en-us/copilot/microsoft-365/microsoft-365-copilot-privacy)
while never training the models. The safe tier also buys a manager the ability to read what
staff typed.

#### The Azgard AI Data Boundary: which class of company data may cross which account tier

The Azgard AI Data Boundary is a grid: four classes of company data down the side, three
account tiers across the top. Every cell answers one question. May this class of data cross
into this tier?

| Class of company data (Azgard AI Data Boundary) | Personal consumer account | Paid business subscription | Contracted workspace or API |
|---|---|---|---|
| Public or already published | Yes | Yes | Yes |
| Internal operational: your own pricing, process, drafts | No | Yes | Yes |
| Personal information about customers or staff | No | Only with the DPA signed and your APP 8 steps documented | Yes, under the agreement |
| Client-confidential (NDA work, unreleased client material) or regulated by statute (health, legal, financial) | No | Contractual only, and only where the client contract allows a subcontracted processor under confidentiality. Statutory: no | Both: contractual under the agreement; statutory only where the vendor signs the instrument the obligation names |

The fourth class covers two sources of obligation over the same data, and the source changes
the answer. A statutory duty waits for the contracted tier, because no setting and no
subscription can consent on a regulator's behalf: the vendor has to sign the specific
instrument your obligation names. A contractual duty is governed by whatever that contract
permits, and where it allows a subcontracted processor bound to confidentiality, the business
tier's agreement can satisfy it. Read the NDA before assuming either answer.

The Azgard AI Data Boundary covers files, not only typing. OpenAI's training commitment runs
over "content", and its residency scope names
["Files (e.g., uploaded images, documents)"](https://help.openai.com/en/articles/9903489-data-residency-and-inference-residency-for-chatgpt)
alongside conversations. Uploading a client contract is the same decision as pasting a
paragraph out of it.

Two vendor exceptions keep the Azgard AI Data Boundary honest, because an account tier sets
the default and defaults have exceptions. OpenAI says a conversation you give feedback on
["may be used to train our models"](https://help.openai.com/en/articles/5722486-how-your-data-is-used-to-improve-model-performance)
even after you have opted out, so a thumbs-up counts as consent. And Anthropic's business
associate agreement
[does not cover web search](https://privacy.claude.com/en/articles/8956058-i-need-to-sign-a-baa-with-anthropic-what-should-i-do):
a feature inside a covered product can sit outside the cover.

#### What actually happens to a prompt you paste into a consumer AI account

A prompt pasted into a consumer AI account does three things the person who pasted it
cannot see. It becomes eligible for training, because OpenAI says that when you use its
services for individuals ["we may use your content to train our
models"](https://help.openai.com/en/articles/5722486-how-your-data-is-used-to-improve-model-performance).
It persists: deleted conversations and Temporary Chats
[go within 30 days](https://openai.com/index/response-to-nyt-data-demands/),
and Anthropic
[may retain consumer chats de-identified for up to five
years](https://privacy.claude.com/en/articles/10023548-how-do-you-use-personal-data-in-model-training)
where it may use them for model improvement. And it can be produced to somebody else.

That third one is not hypothetical. In the New York Times copyright litigation, OpenAI's
own account of who was affected named
[ChatGPT Free, Plus, Pro and Team plus API use without a zero-data-retention agreement,
while stating Enterprise and Edu were not impacted](https://openai.com/index/response-to-nyt-data-demands/).
The American Bar Association Journal reported that in early January 2026 a federal judge
[affirmed an order to produce 20 million de-identified chat
logs](https://www.abajournal.com/news/article/chatgpt-creator-must-turn-over-20m-chat-logs-in-copyright-litigation-federal-judge-says).
Same model, one tier inside the order and one outside it.

#### Shadow AI: staff are already pasting company data into personal AI accounts

Shadow AI is the gap between the AI account a business bought and the AI accounts its staff
actually use. Nobody can tell you your number, so count it yourself. List everyone on your
workspace admin console, then list everyone who writes to customers. Anyone on the second
list but not the first is on a personal account.

The exposure is easier to source than the prevalence. Only
[12% of Australian businesses used AI in 2024-25](https://www.abs.gov.au/media-centre/media-releases/business-adoption-artificial-intelligence-accelerates-2024-25),
around 11% of small and micro businesses, on ABS fieldwork across nearly 7,000 businesses.
Staff adoption ran well ahead of that, on accounts nobody procured. In the OAIC's 2026
attitudes survey of 1,504 adults on a probability panel,
[93% said using personal information to train AI models or products is not fair and
reasonable](https://www.oaic.gov.au/engage-with-us/research-and-training-resources/research/australian-community-attitudes-to-privacy-survey/australian-community-attitudes-to-privacy-survey2026/overview)
and trust in AI companies sat at 4%. The OAIC tells staff to
[refrain from entering personal information into publicly available
tools](https://www.oaic.gov.au/news/blog/GenAI-tools-in-the-workplace-balancing-protection-of-personal-information-and-business-efficiency).

#### Client-confidential and regulated data: what Australian privacy law actually requires

Australian privacy law handles client-confidential and regulated data through APP 8 and
section 16C of the Privacy Act 1988. Before disclosing personal information to an overseas
recipient, APP 8.1 requires you to
[take such steps as are reasonable in the circumstances to ensure that recipient does not
breach the APPs](https://www.oaic.gov.au/privacy/australian-privacy-principles/australian-privacy-principles-guidelines/chapter-8-app-8-cross-border-disclosure-of-personal-information),
and section 16C keeps you liable if it breaches anyway, even after you took those steps.
You cannot outsource the exposure. This is general information, not legal advice.

Australian privacy law splits two events that look identical from the keyboard, and APP 8
only bites on one of them. The OAIC's APP 8 guidance treats giving personal information to
an overseas provider as
[a use rather than a disclosure where a binding contract limits that provider to storing and
enabling access, binds its subcontractors identically, and gives the entity effective control
of how the personal information is handled](https://www.oaic.gov.au/privacy/australian-privacy-principles/australian-privacy-principles-guidelines/chapter-8-app-8-cross-border-disclosure-of-personal-information).
Same model, same servers, different legal event, and the contract is what changes it.

The account tier's power stops there. A tier settles the facts, but what those facts make
you legally is a separate assessment your contract does not decide.

Three Australian specifics change who is covered.

- **The small business exemption is A$3 million and still in force.** The OAIC's
  [small business page](https://www.oaic.gov.au/privacy/privacy-guidance-for-organisations-and-government-agencies/organisations/small-business)
  states the threshold with no repeal notice attached. Removing it is a government commitment
  under review, not enacted law, so treat any blog handing you a repeal date as wrong until
  the regulator's page changes.
- **From 1 July 2026, tranche 2 brought whole professions inside.** Lawyers, accountants,
  conveyancers, real estate professionals, trust and company service providers and dealers
  in precious stones and metals became reporting entities under the AML/CTF Act, and the
  OAIC states such small businesses
  [must comply with the Privacy Act for their AML/CTF activities](https://www.oaic.gov.au/privacy/privacy-guidance-for-organisations-and-government-agencies/organisations/privacy-guidance-for-reporting-entities-under-the-anti-money-laundering-and-counter-terrorism-financing-act)
  while staying outside it for the rest of the business. The accountant pasting client
  onboarding files into free ChatGPT was outside the Act in June and inside it now.
- **Automated decision transparency commences 10 December 2026.** New APP 1.7 to 1.9 require
  a privacy policy to disclose
  [automated decisions that could reasonably be expected to significantly affect an
  individual's rights or interests](https://www.oaic.gov.au/privacy/australian-privacy-principles/australian-privacy-principles-guidelines/chapter-1-app-1-open-and-transparent-management-of-personal-information).
  That covers decisions about people, not AI use generally: screening job applicants is
  caught, drafting an email is not.

#### AI data residency in Australia: do you need a private LLM, or is a contracted tier enough?

AI data residency in Australia is the feature buyers reach for first and should ask about
second. Australian law does not require it: the OAIC states that
[the APPs do not prevent an organisation or agency from sending personal information
overseas](https://www.oaic.gov.au/privacy/privacy-guidance-for-organisations-and-government-agencies/handling-personal-information/sending-personal-information-overseas).
APP 8 governs how you do it, not whether you may.

The feature also does less than its name suggests. Australia is on OpenAI's data-residency
list for storage at rest and
[is not on the inference-residency list](https://help.openai.com/en/articles/9903489-data-residency-and-inference-residency-for-chatgpt),
which covers the United States, Europe and the UAE. Conversations, files and memory sit in
Australia while the GPUs answering them
[run offshore](https://openai.com/business-data/). Residency is also offered to eligible API
customers and new ChatGPT Enterprise and Edu workspaces only, never to free, Plus or ChatGPT
Business, and cannot be retrofitted to a workspace you already have. Account data, billing,
workspace metadata and anything sent to an external integration sit outside the region
regardless.

"Private AI versus public AI" is how the residency choice usually gets framed, and it is the
wrong axis. The same public model on a contracted account gives you the written terms APP 8 asks
you to rely on, while a private deployment gives you control over a question Australian law
is not asking. So a contracted tier with an enforceable contract covers what Australian law
asks for, and a private LLM buys you something the law does not require. The honest
counterpoint: an
APRA-regulated entity carries
[CPS 234](https://www.apra.gov.au/standards/cps-234) obligations to oversee third parties
handling its information security, subcontractors included, and that assessment lands
somewhere different.

#### How to move your team onto a safe AI tier in a week, while the policy gets written

Moving a team onto a safe AI tier takes about a week, and it goes first because it protects
data while you are still writing the policy. It does not replace the policy. The OAIC
advises organisations to write generative AI policies, communicate them and run refresher
training, and the Australian Government publishes a free
[AI policy guide and template](https://www.industry.gov.au/publications/guidance-for-ai-adoption/ai-policy-guide-and-template)
alongside the National AI Centre's
[six essential practices](https://www.ai.gov.au/staying-safe-and-responsible/essential-ai-practices/guidance-ai-adoption-foundations).

1. **Name the tier on every account.** Anyone who cannot name theirs is on the consumer tier.
2. **Buy the business tier for everyone who touches customer information**, on the company
   domain, with single sign-on.
3. **Set the two settings that matter.** Training off at the workspace level, retention set
   to a period you can defend.
4. **Sign the addendum.** At the business and contracted tiers the DPA is part of the
   agreement, so this is a signature rather than a negotiation.
5. **Publish the boundary.** One page, four data classes, three tiers. Then write the policy.

One qualifier, because it is where this advice breaks. A purchase plus two settings covers
training and retention. It does not cover the contract-level instruments regulated work
needs: a healthcare addendum with OpenAI is an email and a case-by-case review that can be
declined, and Google states its business associate agreement
["is not subject to modification"](https://cloud.google.com/security/compliance/hipaa-compliance).
Those are requests, not toggles, and they take longer than a week.

#### What Azgard does with a client's commercial data, and what it will not promise

Azgard is [one engineer in Sydney](https://www.azgard.tech/ai-consultant/au), on-site around
[Sydney](https://www.azgard.tech/ai-consultant/au/sydney) and remote elsewhere, so the data answer is short: builds
land in the client's own accounts and tenancy, documented in plain files the client's team
can read and change. The Azgard AI Data Boundary gets written down for the specific business
before anything is built, because deciding which documents may go into a system is the same
decision as [how to train AI on your company data](https://www.azgard.tech/learn/how-to-train-ai-on-your-company-data).

In practice: Apex Signage's quoting runs through a system Azgard built and Apex's own staff
operate, with quoting time down from about three and a half hours to just over two hours per
quote. Underneath it sits a supplier cost snapshot of 15,758 rows, built from the client's
own files and held in the client's own storage. Home Grown Electrical and Inner Game
Basketball are the other two named Australian customers.

Now the part most consultancies leave out. Azgard holds no security certification, publishes
no audit report and has no penetration test to hand you. It is one person, so there is no
separate security team and no bench to absorb a bad month. That is a genuine ceiling on what
Azgard can take on. Azgard does not publish which vendor tiers it holds either, so apply
the Azgard AI Data Boundary's own rule to Azgard: ask which account tier your work will
sit on, before you
sign. Costs are in
[what an AI consultant costs in Australia](https://www.azgard.tech/learn/how-much-does-an-ai-consultant-cost-au).

**FAQ**

Q: Is ChatGPT safe for company data?
A: It depends on the account, not on ChatGPT. On a personal consumer account, free or Plus or Pro, OpenAI says it may use your content to train its models unless you turn that off in Data Controls. On ChatGPT Business, Enterprise, Edu and the API, OpenAI says it does not train on business inputs or outputs by default, and its Services Agreement makes that a contract term rather than a setting. So the same question has three different answers depending on which account the prompt was typed into.

Q: Does ChatGPT Plus train on my data?
A: Yes by default. ChatGPT Plus and Pro are paid consumer subscriptions, and OpenAI's no-training commitment covers its business products, not its individual ones. You can switch training off yourself in Settings under Data Controls, but that is a per-person preference nobody else in the business can see or audit. It is the commonest gap in a small business: someone expensed a Plus subscription and everyone assumed paying for it changed the terms.

Q: Can I put customer personal information into ChatGPT under Australian privacy law?
A: The Privacy Act 1988 does not forbid it, and the OAIC states plainly that the Australian Privacy Principles do not prevent an organisation from sending personal information overseas. But APP 8 requires you to take reasonable steps to ensure the overseas recipient does not breach the APPs, and section 16C keeps you liable if it does, even after you took those steps. As a matter of best practice the OAIC recommends organisations do not enter personal information, and particularly sensitive information, into publicly available generative AI tools. This is general information, not legal advice.

Q: Is my ChatGPT data stored in Australia?
A: Only on some accounts, and only for storage. Australia is on OpenAI's data-residency list for data at rest, so conversations, files and memory can be stored here. Australia is not on its inference-residency list, which currently covers the United States, Europe and the UAE, so the model itself still runs offshore. Residency is offered to eligible API customers and new ChatGPT Enterprise and Edu workspaces, not to free, Plus or ChatGPT Business accounts, and it cannot be added to an existing workspace later.

Q: Is Claude safe for business data?
A: The same tier logic applies to Claude as to ChatGPT: the account you reach the model through decides what happens to what you send, not the model. On Anthropic's consumer side, chats are used to improve Claude only if you allow it, and where you do, Anthropic says it may retain that data in a de-identified format for up to five years. Zero data retention is a separate contracted arrangement, available on eligible Anthropic APIs, on Commercial organization API keys including Claude Code via the API, and on Claude Code Enterprise plans. One limit survives even there: Anthropic's business associate agreement does not cover web search.

Q: Does the small business exemption mean the Privacy Act does not apply to my AI use?
A: Not necessarily, and the answer changed for several professions this year. The exemption still applies to most businesses with annual turnover of A$3 million or less, and the OAIC's small business page states that threshold with no repeal notice attached. But from 1 July 2026 lawyers, accountants, conveyancers, real estate professionals, trust and company service providers and dealers in precious stones and metals became reporting entities under the AML/CTF Act, which brings them under the Privacy Act for those activities. The carve-out matters: it covers the AML/CTF work, not the rest of the business.

---

### How to train AI on your company data, in the order that works

URL: https://www.azgard.tech/learn/how-to-train-ai-on-your-company-data
Published: 2026-08-25
Topic: operations · Tags: company-brain, context, knowledge-base, sops, adoption · ~10 min read

> To train AI on your company data you choose files, not weights. The Company Brain Load Order sets what goes in first, what goes last, and what stays out.

**TL;DR**

- At small-business scale, nobody trains a model on their own company data: the model reads your files at question time, and choosing the files is the whole job.
- The Azgard Company Brain Load Order loads three to five finished jobs first, the numbers and thresholds that decided them second, and finished templates last or not at all.
- Ask any AI vendor what they want from you first: if the answer is 'your documents', they are building a search index over your filing, and if it is 'three finished jobs and the rules you used to price them', they are building a brain.
- A company file an AI can use carries no bare numbers: every value states its unit and the rule that produced it.
- Commercial fine-tuning APIs are poor at learning new facts from a small corpus, averaging 37% generalisation accuracy on new knowledge and 19% on updated knowledge in Stanford's FineTuneBench.

Training an AI on your own company data is not training, at small-business scale. Nobody
adjusts a model's weights. The model reads your files at the moment you ask the question,
so the work is choosing which files, in which order. The Azgard Company Brain Load
Order sets that sequence: three to five finished jobs first, then the numbers and rules
that decided them, then finished templates last or not at all. Drawings, contracts and
correspondence stay out.

#### What "training AI on your company data" means, and why you are not training anything

Two different things get called training AI on your company data. Fine-tuning changes the
model's weights. Retrieval leaves the model alone and puts your files in front of it at
question time. At small-business scale it is the second one, every time.

| Approach | What it changes | What the evidence shows | When it applies |
| --- | --- | --- | --- |
| Fine-tuning through a commercial API | The model's weights | [37% generalisation accuracy on new knowledge, 19% on updated](https://arxiv.org/abs/2411.05059) | Teaching a format or an output shape |
| Retrieval at question time | Nothing. Your files sit in the prompt | [Mistral-7B 0.481 to 0.875](https://arxiv.org/abs/2312.05934), against 0.504 fine-tuned and 0.588 paraphrase-augmented; Llama2-7B fine-tuned 0.219, under its 0.353 baseline | Facts that changed last week, rules only you hold |
| Synthetic continued pretraining | The weights, after synthesising a much larger corpus | The model answers without the documents, and [compounds with retrieval rather than replacing it](https://arxiv.org/abs/2409.07431) | Research pipelines, not services you can buy |

FineTuneBench ran 1,075 test questions across five models and two providers at default
hyperparameters; Microsoft's comparison ran three 7B models on multiple-choice current
events, unsupervised fine-tuning only.

Stanford's FineTuneBench found the tuned models memorised their training pairs and failed
the same facts rephrased: the model learns that the first-pitch temperature at Dodger
Stadium on 8 September 2024 was 103°F, then answers 103 for 2030.

Learning a fact into weights is data-inefficient: it takes hundreds to thousands of diverse
representations of that fact, which breaks down on a small corpus
["where each fact may appear rarely or only once"](https://arxiv.org/abs/2409.07431). Your
margin rule appears once.

The FineTuneBench authors make two arguments against retrieval, and they are not equal. The
first is that retrieval does not scale with the size of the corpus, on cost and on context
limits. That one is scale-bound, and at small-business scale
[the whole corpus fits in the prompt](https://www.anthropic.com/engineering/contextual-retrieval).
The second is that retrieval often fails even when the answer is already in the context,
and that holds at any size. Their authority is ClashEval, the paper this page returns to
on messy documents. Concede it: the fix is fewer contradictions in your files, not fewer
files.

#### What a company brain is, and how it differs from a knowledge base or an internal wiki

A knowledge base is written by people for people to read. A company brain, also called
an AI second brain for business, is written for a machine to apply, and most of it is
pulled out of work already done rather than composed from scratch. The difference is
authorship, not storage.

Point an AI at your shared drive and you get a good search engine over your own filing. It
will find the quote you sent in March. It cannot price the next job the way your best
estimator would, because the rules that produced that quote were never written down as
rules. They exist in the finished output and in the maker's head, which is why
[you cannot buy the part that was never written down](https://www.azgard.tech/learn/you-cannot-buy-your-company-brain).
You build a brain by extraction: take work the business has already done and pull the
decisions out of it.

#### The Azgard Company Brain Load Order: what to feed it, in what order

The Azgard Company Brain Load Order is three tiers and a never list, and the order is the
point. It comes from Azgard's setup guide for the Apex Signage estimating
build: "put in what helps Claude understand how Apex
prices, and hold back anything that would anchor the team onto a finished structure before
they've described their own".

| Tier | What goes in | Why it sits there |
| --- | --- | --- |
| 1. Finished jobs | Three to five past quotes, estimates or job files, complete | Real rates and hours live in outputs, not policy |
| 2. Decision numbers | The rates, margins and thresholds behind tier one, however rough | Turns worked examples into something repeatable |
| 3. Templates | Your template, last or not at all, labelled reference only | Pre-decides a structure nobody has described yet |
| Never | Drawings, tender packs, contracts, correspondence | Dilutes what the model pays attention to |

Azgard's Apex setup guide calls those three to five past quotes "the
single most valuable input. This is where Apex's real rates and labour hours actually
live." Tier two is the raw cost side of every row, allowed to be scrappy: at Apex, a
supplier cost snapshot of 15,758 rows.

A portable test: ask any AI vendor what they want from you first. If the
answer is "your documents", they are building a search index over your filing. If the
answer is "three finished jobs and the rules you used to price them", they are building a
brain.

#### Why your finished templates should go in last, or not at all

A finished template is the most tempting file in the building and the worst one to load
first. Azgard's Apex setup guide is blunt about why: dropping a finished 20-column template
in front of the team "anchors them onto someone else's structure before they've described
their own", and "its margin machinery quietly pre-decides the cost-vs-sell question". You
wanted the business's pricing logic. You got the template author's.

The template does get back in. Add it near the end, once the team has shaped its own
structure, as a mapping check labelled reference only, do not fill in.

#### How to write an SOP an AI can use: units, rules, and no bare numbers

An SOP an AI can use states the unit and the rule behind every value, so no number stands
on its own and nothing rests on "obviously". A number with no rule attached is a fact about
one job in the past. A number with its rule attached is a decision the model can make
again.

Here is the pattern on an invented line. Before:

> Callout: 1 hour.

After:

> Callout: 1 hour on site per visit, one technician, inside 20 km of the depot. Add 30
> minutes each way beyond 20 km. Anything outside the metro area is quoted separately,
> never estimated from this line.

Same number. The second version survives a job it did not come from, and it tells the model
where to stop and ask. The dangerous failure is the model that confidently applies a metro
callout rate to a job three hours out of town.

#### What files your AI should have access to, and what to keep out

Capacity is no longer what decides which files your AI should read. Anthropic's engineering
guidance states that
["if your knowledge base is smaller than 200,000 tokens (about 500 pages of material), you can just include the entire knowledge base in the prompt"](https://www.anthropic.com/engineering/contextual-retrieval).
Few small businesses hold 500 pages of decision rules. The same page names its own
limit: "more information can be distracting for models so there's a limit to this."

Knowledge management for a small business is now that question: not what fits, but what
earns its place. Azgard's Apex
guide draws the line at whether a document is about how the business decides. Everything
else stays out: "Drawings, tender packs, contracts, correspondence, or anything not about
how Apex prices. It dilutes what Claude pays attention to. Keep the folder tight."

A tight folder also makes the next job survivable: deciding
[who owns each file and when it expires](https://www.azgard.tech/learn/your-company-brain-needs-an-owner-and-an-expiry-date).
Twelve files can each have an owner. Four thousand cannot.

#### Getting tribal knowledge out of people's heads before they leave

Tribal knowledge does not come out of people by asking them to write documentation. Ask
your best estimator to document how they price and you get three bullet points and a
promise to finish it next week.

Ask them to narrate one real job instead, while the AI writes it down. Open a quote they
sent last month, walk it line by line, and at every number ask why that number and what
would have changed it. The model drafts the rule as they talk. They correct it out loud as
it appears, which in our sessions is the editing step people actually do.

An AI Company Brain Kickstart ($1,950) runs that way: a 90-minute build session that authors
your first skills against real jobs, five to seven days of running them, then a 60-minute
review that ships the next few. Three to five skills in total.

Two hours of narration captures the rules behind the jobs you walked, not
everything in somebody's head. Nothing captures all of it.

#### Can AI use messy documents, or do you have to clean everything first?

Messy documents are fine. Contradictory ones are not, and that distinction is most of the
answer. Azgard's Apex guide asks for the cost side "however rough", because "even a messy
internal cheat-sheet is gold".

Conflicting numbers are the problem. Deakin University's catalogue of retrieval failure
points names it: "the answer is present in the context, but the large language model
failed to extract out the correct answer... typically this occurs when there is too much
noise or contradicting information in the context." [That study](https://arxiv.org/abs/2401.05856),
which ran 1,000 question-and-answer pairs on GPT-4, catalogues seven failure points, so
retrieval is not flawless either. At small-business scale the constraint is not retrieval, it is what your files disagree
about.

The sharper finding is what happens when your file contradicts what the model knows. In
ClashEval, six leading models including GPT-4o
[overrode their own correct prior knowledge more than 60% of the time](https://arxiv.org/abs/2404.10198)
when a retrieved document disagreed with it. ClashEval adds a caveat: "the more
unrealistic the retrieved content is... the less likely the model is to adopt it." A wildly
wrong file gets caught. A plausibly stale price does not. That is a company brain's failure
mode exactly: reconcile the numbers that disagree, so every number that matters has a
single source of truth for AI to find.

#### What a company brain looks like ninety days in

Ninety days in, a company brain is a small set of files people argue about. That is the
healthy version. At Apex Signage, the estimating build moved quoting time from about 3.5
hours to just over 2 hours per quote, and the arguments were worth more than the minutes:
the rules it got wrong were mostly rules nobody had ever agreed out loud.

Azgard's AI Baseline Card has six
fields, and [the baseline you take before anyone builds anything](https://www.azgard.tech/learn/how-to-measure-ai-roi)
is only evidence when all six are filled. On the Apex estimating work, fields four and
five, the named measurer and the date, are blank. We captured the process in detail and
never logged the minutes or the total live. So that figure is an honest
before-and-after from the people doing the work, not an auditable measurement. Weight it
accordingly.

Do better than we did: take the count and time per unit for one process, put a name and
a date against them, then load the first three finished jobs.

**FAQ**

Q: Do I need to fine-tune a model to train AI on my company data?
A: Almost certainly not. At small-business scale the model reads your files at the moment you ask a question, and nothing about the model changes. Stanford's FineTuneBench found commercial fine-tuning APIs averaged 37% generalisation accuracy on new knowledge and 19% on updated knowledge, memorising training pairs and then failing the same fact rephrased. Fine-tuning is a good way to teach a format and a poor way to teach facts.

Q: What files should I give an AI access to?
A: Three to five complete finished jobs first, then the rates, margins and thresholds that decided them, then your templates last or not at all. Keep out drawings, tender packs, contracts and correspondence, because they are not about how the business decides and they dilute what the model pays attention to. Anthropic's engineering guidance notes that a knowledge base under 200,000 tokens, roughly 500 pages, fits in a prompt whole, so capacity is not your constraint. Attention is.

Q: Can AI use messy documents, or do I need to clean everything up first?
A: Messy is fine. Contradictory is not. Scrappy formatting, a spreadsheet with three tabs and a cheat-sheet in someone's handwriting all work. Two files stating different prices for the same thing do not, because a model given conflicting context frequently picks the wrong one. Resolve the numbers that disagree, then stop cleaning.

Q: What is a company brain, and how is it different from a knowledge base?
A: A knowledge base is written by people for people to read, and it records what happened. A company brain is written for a machine to apply, and it records what decides: the rates, thresholds and exceptions that produced the work. Most of a company brain is pulled out of finished jobs rather than composed from scratch, because the rules that made those jobs were never written down as rules.

Q: How do I capture what's in a long-serving employee's head before they leave?
A: Do not ask them to write documentation. Ask them to narrate one real job while the AI writes it down. Open a quote they sent last month, walk it line by line, and at every number ask why that number and what would have changed it. They correct the draft out loud as it appears.

---

### You cannot buy your company brain, and this week proved it

URL: https://www.azgard.tech/learn/you-cannot-buy-your-company-brain
Published: 2026-08-20
Topic: perspectives · Tags: company-brain, context, agents, adoption · ~9 min read

> The AI context layer is now something you can buy. It indexes what your business recorded, not the company brain rules that decide whether an agent is useful.

**TL;DR**

- A context layer indexes knowledge a business has already recorded, and the operating rules that make an AI agent useful in a 15-person company were never recorded anywhere.
- In The Modern Data Company's interim 2026 survey of 540+ data leaders, 60.9% called a reliable context layer necessary while only 16.0% deliberately engineer context as a product.
- Anthropic reported that its analytics agents did not exceed 21% accuracy on its own internal evals without skills and ran consistently above 95% with them, where a skill is procedural knowledge written into a markdown file.
- Enterprise context infrastructure from AWS and Salesforce is real and earns its money at scale, but it moves recorded knowledge around rather than manufacturing knowledge nobody wrote down.
- The cheapest context layer a small business can build is a couple of pages naming its margin floor, its approval thresholds and its exceptions, with one person's name against it and a date to review it.

AWS started this in June. The rest of the market caught up this week: three announcements
in three days, and somewhere in the middle of them the "context layer" finished its
journey from a phrase engineers use in architecture diagrams to a category you can put on
a purchase order. Salesforce shipped a way into its version. A London company that had
gone nine years without outside money took $22M to build one, and a startup founded by
three ex-AWS engineers went generally available with a seed round behind it.

The category is real, and I want to say that up front because most of what follows is an
argument against buying it. But the conclusion a small business is being invited to draw
from this week is wrong, and it will cost people money.

#### What an AI context layer for business actually is

Strip out the vocabulary and every one of these products does the same job. It points at
systems you already run - the CRM, the ticketing queue, the warehouse, the job management
platform - reads what is in them, works out how the records relate, and hands your agent a
governed view of it at the moment it needs one.

AWS calls its version Context. It was
[announced at the New York summit in June](https://aws.amazon.com/blogs/aws/top-announcements-of-the-aws-summit-in-new-york-2026/),
and the pitch is that it builds the knowledge graph automatically from your existing
organisational data rather than making somebody curate it, so you stop hand-rolling
retrieval pipelines. VentureBeat called that
[entering "the context layer race"](https://venturebeat.com/data/aws-enters-the-context-layer-race-with-a-graph-that-learns-from-agents-not-manual-curation).
The rest of the field spent this week proving the race exists.

Yesterday Salesforce
[expanded Headless Data 360 for MCP](https://siliconangle.com/2026/08/19/salesforce-expands-headless-data-360-mcp-developers-can-bring-insights-agents/),
exposing 200-odd existing APIs as programmable endpoints so an agent can pull governed
customer context with no UI in the way. The same day Prevalent AI
[raised $22M from Integrity Growth Partners](https://www.unite.ai/prevalent-ai-raises-22m-to-scale-trusted-enterprise-context-for-ai-systems/)
to scale trusted enterprise context, and the detail worth noticing is that this was the
first primary outside investment in the company's nine-year history. A security data
fabric business repositioned around context and the cheque arrived. Two days before that,
[xpander went GA on a $7.5M seed](https://venturebeat.com/orchestration/as-enterprises-confront-ai-agent-sprawl-xpander-wants-them-to-own-their-own-control-and-context-layer),
selling a vendor-neutral control plane to enterprises drowning in agents.

Different companies, one sentence: point this at your systems and your agents will finally
understand your company.

That sentence has a hole in it, and the hole gets wider the smaller you are. A graph is an
index of what you already recorded, and it is very good at that. But in a fifteen-person
business, the knowledge that decides whether an agent is useful or dangerous is the
knowledge nobody ever recorded.

#### The 45-point gap nobody can sell you out of

The Modern Data Company released interim findings from its third annual survey yesterday,
and the numbers accidentally make my case better than I can. Across
[540-plus qualified responses from 66 countries](https://salestechstar.com/predictive-ai-artificial-intelligence/ai-agents-are-advancing-faster-than-enterprise-data-is-ready-for-them/),
60.9% consider a reliable context layer necessary. Only 16.0% deliberately engineer
context as a product.

Hold the caveats where you can see them. This is interim, it is self-selected, the
respondents are data leaders and practitioners who chose to answer, and it skews
enterprise. These are not small businesses. They are the professionals whose job this is.

Which is what makes the gap interesting. Forty-five points separate the people who believe
they need a context layer from the people who are actually building one, among a sample
that could sign a vendor contract this afternoon. Budget is not the binding constraint.
Neither is belief. Every one of those 60.9% could buy AWS Context or Salesforce Data 360
by Friday and the gap would still be 44.9 points wide on Monday, because engineering
context means deciding what your own terms mean and writing that down. Nobody sells
that.

The rest of the survey reads the same way. 8.4% say the data feeding their AI is
production-ready. 63.5% name missing context and lineage as an obstacle. 75.9% rank data
quality and trust among their top three barriers. Meanwhile 57.3% are already piloting or
running agents, with 23.5% in production. Saurabh Gupta, the chief executive, puts it as
"the harder question is whether those agents have the trusted data and business context
they need to operate reliably."

Agents in production, data nobody trusts. That combination is the whole reason this week
happened.

#### Why AI agents give confidently wrong answers about your business

Sit with an estimator for an afternoon and you can watch this happen.

Your job management system records that you quoted job 4821 at a particular number. It
does not record the margin you will not go below. It has nothing to say about the two
customers who get a different number because of what happened in 2019. It has never heard
of the loading your best estimator applies to anything involving a two-storey install,
which they have never written down because to them it is not a rule, it is just how you do
the job.

Index all of that and you get a beautiful graph of outcomes with the reasoning stripped
out. An agent reading it can tell you what you charged. Ask it what you should charge and
it will produce a number, confidently, because it has 4,000 examples of numbers and no
statement of the rule that produced them. The honest response would be "nobody has told
me", and that is not what these systems do.

This is where the reflex kicks in, and it is the same one I keep writing about: a
knowledge problem gets met with
[a purchase order](https://www.azgard.tech/learn/stop-buying-ai-tools-before-you-map-the-work). It feels like
progress, and it is a great deal faster than sitting down with your estimator for a
morning. It also leaves the gap exactly where it was.

#### Where buying an AI context layer does pay off

There is a real case on the other side of all this, and it starts at a scale most people
reading this will never reach. At 5,000 staff and 400 systems, writing things down is not
your constraint. Your constraint is that "customer" means four incompatible things across
four platforms and no human alive can hold the relationships in their head. TechTarget's Scott Thompson framed
it this week as
["the bottleneck isn't volume, but context"](https://www.techtarget.com/data-technologies/feature/Why-more-data-will-not-deliver-AI-data-readiness),
and quotes Collibra's chief executive Felix van de Maele on the "hallucination tax"
enterprises pay when agents run without governed context. Thompson's own read on what that
tax buys you: a human-in-the-loop verification bottleneck that "erodes the productivity
gains that justified the AI investment." A real cost with a real remedy, and the remedy is
infrastructure.

Scale makes it worse fast. Gartner estimates, as reported by VentureBeat, that the average
Fortune 500 company will run more than 150,000 AI agents by 2028, up from fewer than 15
last year, while only 13% of organisations think they have the right governance today. If
that is your 2028, buy the control plane. Buy two.

So no, the vendors are not selling snake oil, and I have no interest in the cheap version
of this argument. AWS Context building a graph automatically instead of by hand is a
genuine improvement on the previous state of the art. Salesforce exposing 200 APIs so
agents get governed customer data is straightforwardly good engineering.

My argument is about category, not quality. Infrastructure moves recorded knowledge around
and makes it reachable. It cannot manufacture unrecorded knowledge, and no amount of
funding changes that, because the missing input is a judgement somebody in your business
has to make out loud for the first time. The coverage of the Prevalent AI round gets
within one word of it: the challenge is shifting
["toward whether it has access to accurate information about the organization in which it is operating"](https://www.unite.ai/prevalent-ai-raises-22m-to-scale-trusted-enterprise-context-for-ai-systems/).
Access to accurate information. Access assumes the information exists.

#### Do I need a knowledge graph for AI agents, or do I need to write things down?

Anthropic published a number in June that I have quoted in client rooms more than any
other this year. On its own internal evals, for analytics questions specifically, Claude's
accuracy
["didn't exceed 21%"](https://claude.com/blog/how-anthropic-enables-self-service-data-analytics-with-claude)
without skills. With them, "consistently above 95%". Skills, in that post, are defined as
procedural knowledge encoded in markdown files.

Take the caveats seriously. Those are Anthropic's evals, on Anthropic's data model,
measuring one narrow class of question. Do not carry 21-to-95 into a board paper as a
general business benchmark, because it is not one.

The shape of it is still the point. Anthropic has as much data infrastructure available to
it as anyone on earth, and what closed the gap between useless and reliable was somebody
writing out how the question gets answered. Not a new retrieval architecture. Files.

That is the same argument I made from the other direction a week ago, when OpenAI's
telemetry showed frontier firms
[pulling 8.3 times the output per user out of identical models](https://www.azgard.tech/learn/same-model-8x-the-output-the-difference-is-written-down).
That post was about what the winners write. This week is about what the market would like
to sell you instead.

#### What I would actually do with a Thursday afternoon

If you run a business under a hundred people and this week has you worried, the useful
response is smaller and less satisfying than a procurement process.

Pick the workflow where a wrong answer costs real money. Usually that is quoting. Sit
with whoever is best at it and get the rules out of their head and into a file: the floor,
the thresholds, the exceptions, and an honest note about which situations the rules do not
cover. Two pages is a good first target. Then connect the one or two systems that hold the
facts, so the agent stops guessing at what it cannot see.

That is a context layer. It cost you an afternoon and it is worth more to a small business
than a knowledge graph over records that never explained themselves.

The catch is that written knowledge rots faster than infrastructure does, so once you have
written it,
[somebody has to own it and it has to carry an expiry date](https://www.azgard.tech/learn/your-company-brain-needs-an-owner-and-an-expiry-date).
A file with no name against it is a liability with better formatting.

The vendors will keep raising, and enterprises will keep buying, and both of those things
are fine. Just do not read this week as permission to skip the part nobody can do for you.

**FAQ**

Q: What is an AI context layer for business?
A: It is infrastructure that connects an AI agent to the systems where your company data already lives, works out how the records relate to each other, and hands the agent a governed view of that at the moment it answers a question. AWS Context, Salesforce Data 360 and the data fabric vendors all sell versions of it. It indexes what you have recorded; it does not create knowledge that was never written down.

Q: Why do AI agents give confidently wrong answers about my company?
A: Because they can see your records and not your reasoning. Your CRM stores the number you quoted, not why you quoted it. When an agent has the outcome but none of the rules that produced it, the honest answer would be to say it does not know, and models are not built to do that. It generalises from the pattern in front of it, which is exactly how you get a plausible number nobody in the business would have signed off.

Q: Do I need a knowledge graph for AI agents?
A: Not first, and probably not at all under a few hundred staff. A knowledge graph pays off when you have so many systems with conflicting definitions that no human can hold the relationships in their head. A 15-person business has the opposite problem: few systems, and almost nothing written down about how decisions get made. Write the rules down and connect the two or three systems that hold the facts, then reassess.

---

### How to choose an AI consultant: the show-me list

URL: https://www.azgard.tech/learn/how-to-choose-an-ai-consultant-au
Published: 2026-08-20
Topic: strategy · Tags: vendor-evaluation, hiring, due-diligence, procurement · ~8 min read

> How to choose an AI consultant: ask for four things before you pay - a client you can phone, a system in production, a named builder, exit terms.

**TL;DR**

- The Azgard Show-Me List is four things you ask an AI consultant to produce before you pay: a named client you can contact, a system in production operated by someone who does not work for them, the name of the person who will do your work, and a written answer to what you keep if you stop paying.
- The portable rule is to ask for the thing, not the answer about the thing, because anything a consultant can only produce after you pay is a promise rather than evidence.
- Across nine Sydney AI consulting firms whose case-study and client pages were checked on 20 August 2026, seven publish no client name at all and two do, and not one of the nine publishes contact details for a referee.
- An Australian buyer can verify an AI consultant's business for free in about ten minutes using ABN Lookup, which shows the date GST registration took effect, plus ASIC Connect, AusTender and the buy.nsw supplier search.
- Australian registers like ABN Lookup, ASIC Connect and AusTender prove an AI consultant's business exists and how long it has traded, but none of them proves the consultant is any good at the work.

The Azgard Show-Me List is four things to ask an AI consultant to produce before you pay
them: a client you can phone by name, a system in production you can watch a non-employee
use, the name of the person who will do your work, and a written answer to what you keep if
you stop paying. Answers about methodology are free. Only the four artefacts are checkable,
and a consultant who has none is billing you to learn.

#### Why most checklists for choosing an AI consultant test the wrong thing

Most checklists for choosing an AI consultant test what the consultant says: methodology,
frameworks, change management. Those are answers, and answers are free. A market that
selects on answers selects for whoever answers most fluently - the one skill every
consultant can borrow from the product they sell.

The questions that work are the ones whose answers existed before you asked. An Australian
buyer in Azgard's harvest named the fear: "Biggest fear is paying for fancy roadmap deck or
some 'agent' nobody uses after 2 months."

#### The Azgard Show-Me List: four things to ask an AI consultant to produce

The Azgard Show-Me List runs in the order below. Each item exists or it does not.

| Ask them to produce | What it proves | A weak answer |
|---|---|---|
| 1. A named client you can contact | The work happened, and an outsider says so | "A leading Australian manufacturer" |
| 2. A production system on screen-share, driven by a non-employee | It survived real staff | A recorded demo |
| 3. The name and calendar of whoever builds it | You know who builds, and if they are free | "Our delivery team" |
| 4. A written answer to what you keep at the end | You are buying an asset, not attention | "We'll sort that out later" |

The portable rule underneath all four: ask for the thing, not the answer about the thing.
Anything an AI vendor can only produce after you pay is a promise you cannot check.

#### Ask for a named client you can contact, not a logo wall

A named client you can contact is the first Azgard Show-Me List item. Across nine Sydney AI
consulting firms whose case-study and client pages were checked on 20 August 2026, seven
publish no client name at all. Two name clients:
[Team 400](https://team400.ai/case-studies) lists Coast Smoke Alarms, and
[AI Consulting Group](https://aiconsultinggroup.com.au/case_studies/axify_voice_ai/) lists
Momento Hospitality.

None of those nine Sydney firms publishes contact details for a referee. A name on a
case-study page is not a reachable referee, and several named entries are products, not
companies: Cargo Guardian, Springboards AI, Houdini AI Platform.

Some firms genuinely cannot publish names:
[Kernel Flow's terms](https://kernelflow.io/terms) reserve the right to use anonymised case
studies. That is why the ask is a contact rather than a logo - confidentiality
still lets a consultant ring a client and ask if they will take a call.

#### Watch someone who is not the consultant use the system in production

Item two of the Azgard Show-Me List is a screen-share of a system in production, driven by
someone who does not work for the consultant. The idea is not new. Sydney firm
[Kernel Flow](https://kernelflow.io/blog/best-ai-consultants-sydney) puts it as a rule: "the
key is understanding if they build running systems or just write roadmap plans." Team 400,
another Sydney consultancy, puts it as a count - how many AI systems does the firm have
running in production right now - recorded in Azgard's competitor survey.

What the Show-Me List adds is the operator. A demo the consultant drives proves only that the
consultant can drive it. Watch a client's staff member instead: where they hesitate, what they
stopped using. An AI consultant needs no computer science degree, but
whoever writes your build must have shipped something you can watch a stranger use.

#### Ask what you keep the day you stop paying

The fourth Azgard Show-Me List item is a written answer to what you keep the day you stop
paying an AI consultant. Put it in the proposal: who holds the API keys and accounts, whose
tenancy the automations run in, whether your documents sit in files you can open. A system
you cannot administer is a subscription to their attention - the subject of
[you cannot buy your company brain](https://www.azgard.tech/learn/you-cannot-buy-your-company-brain).

Measurement nests inside this item: agree which number moves and who reads it, the job
of the [Azgard Three-Question Vendor Filter](https://www.azgard.tech/learn/how-to-measure-ai-roi). Azgard publishes
its exit answer, so a buyer can hold it to this: "every build is owned and documented so
it runs without us."

#### How to verify an AI consultant's business is real in Australia

An Australian buyer can verify an AI consultant's business free, in ten minutes. Australian
AI adoption went from
[1% of businesses in 2022-23 to 12% in 2024-25](https://www.abs.gov.au/statistics/industry/technology-and-innovation/characteristics-australian-business/latest-release),
so the demand is new, and registration dates show whether the supplier is too.

| Register | What it shows free | Paid tier |
|---|---|---|
| [ABN Lookup](https://abr.business.gov.au/) | Entity name and type, ABN and GST registration each with its effective date, state and postcode | Nothing to pay |
| [ASIC Connect company register](https://connectonline.asic.gov.au/RegistrySearch/faces/landing/SearchRegisters.jspx) | Company type, ACN, registration and review dates, registered office suburb, documents lodged | $10 current extract; $20 with directors and history |
| [ASIC business names register](https://connectonline.asic.gov.au/RegistrySearch/faces/landing/bn/SearchBnRegisters.jspx) | Who holds a trading name, and their ABN | Extracts paid |
| [buy.nsw supplier search](https://buy.nsw.gov.au/supplier/search?schemes=20) | NSW Government ICT prequalification, small-business and NSW-based flags | Free, no login |
| [AusTender](https://www.tenders.gov.au/cn/search) | Federal contracts at or above $10,000 inc GST: agency, value, period, consultancy flag | Free, no login |

The GST date is a longevity signal almost nobody uses. GST registration is compulsory
[once turnover reaches $75,000](https://www.ato.gov.au/businesses-and-organisations/gst-excise-and-indirect-taxes/gst/registering-for-gst),
so its effective date on an [ABN Lookup record](https://abr.business.gov.au/) roughly marks
when the business became a trading concern. ABN Lookup carries no address, email or
phone.

AusTender lists every Australian Government contract at or above
[$10,000 including GST](https://www.tenders.gov.au/cn/search) for most agencies, showing
whether an agency has paid this consultant. buy.nsw's
[SCM0020 ICT Services scheme](https://buy.nsw.gov.au/scheme/6B1F91AE-B3CC-0F0E-17267D5EF7BCEB9E)
carried 6,899 suppliers on 20 August 2026,
[4,947 publicly listed](https://buy.nsw.gov.au/supplier/search?schemes=20), and its
Advanced tier requires two referee reports for work in the last three years -
somebody else already ran the reference check.

buy.nsw marks displayed supplier information unverified, so membership is the signal and
the profile blurb carries no weight.

#### AI consultant red flags, and the ones that are not red flags

The AI consultant red flags worth acting on are refusals to produce something that exists.
Refusing to name who does the work is the most common - one buyer called agencies "very
secretive of who actually works there". That is the bait and switch: whoever sells is not
whoever staffs it. A five-figure discovery phase as their smallest sellable unit is another,
as is a proposal naming no deliverable you keep, or a certification presented as a
qualification: no licence is required to call yourself an AI consultant here,
and
[ASIC's professional registers](https://www.asic.gov.au/online-services/search-asic-registers/professional-registers-search/)
cover financial services licensees, auditors and liquidators - consultants are not on them.

Two things look like red flags here and are not. A sole trader or a two-person Pty Ltd is
fine in Australia; a hidden structure is what should worry you. Absence from buy.nsw or the
[AI Directory](https://aidirectory.industry.gov.au/) means nothing either:
listing is voluntary, and NSW buyers can go direct to a small business
[up to $250,000 excluding GST](https://buy.nsw.gov.au/scheme/6B1F91AE-B3CC-0F0E-17267D5EF7BCEB9E).

Australian law helps less than buyers hope. Since 9 November 2023 the unfair contract terms
protections cover a small business with
[fewer than 100 employees or under $10 million in annual turnover](https://www.accc.gov.au/business/selling-products-and-services/contracts)
and no contract-value cap, replacing the old 20-employee and $300,000 tests. The maximum
corporate penalty for conduct on or after 28 March 2026 is
[the greater of $100,000,000, three times the benefit obtained, or 30% of adjusted turnover](https://www.accc.gov.au/business/compliance-and-enforcement/fines-and-penalties)
- the [$50 million figure from 2022](https://www.accc.gov.au/media-release/accc-welcomes-new-penalties-and-expansion-of-the-unfair-contract-terms-laws)
is superseded. But only a court can void a term, and the ACCC does not investigate individual
complaints. Start with the
[Australian Small Business and Family Enterprise Ombudsman](https://www.asbfeo.gov.au/dispute-support),
which mediates such disputes cheaply. None of it gets your project back - which is why the
four demands come first.

#### How Azgard answers its own four questions

Azgard has to answer its own Show-Me List. It is
[one engineer in Sydney](https://www.azgard.tech/ai-consultant/au), Angus McDonald - on-site around
[Sydney](https://www.azgard.tech/ai-consultant/au/sydney), remote everywhere else.

Item one, a named client you can contact: three Australian ones - Apex Signage, Home Grown
Electrical and Inner Game Basketball. Three, not thirty, and a buyer can ask for any by name.
Item two, a system run by someone else: Apex Signage's quoting runs through a system Azgard
built and Apex's own staff operate, with quoting time down from around three and a half hours
to just over two per quote. Item three, who does the work: always Angus, a capacity
ceiling rather than a feature - there is no bench to absorb a bad month. Item four, what you
keep: every build documented in plain files your team can change without Azgard on retainer.

Two admissions that cost something. The Apex figure was never baselined the way
[Azgard's own AI Baseline Card](https://www.azgard.tech/learn/how-to-measure-ai-roi) demands - it misses two of the
card's five fields - and is published anyway with that caveat. And two of Azgard's six
offerings publish a price while four are scoped on a call, the opacity this page criticises;
bands are in
[what an AI consultant costs in Australia](https://www.azgard.tech/learn/how-much-does-an-ai-consultant-cost-au).
Run Azgard through the registers first.

**FAQ**

Q: How do I check an AI consultant's business is real in Australia?
A: Start with ABN Lookup, which is free and shows the entity name, the date the ABN became active and the date GST registration took effect. Add ASIC Connect for the company record, where a $20 current and historical extract lists directors and prior company names. AusTender shows any Australian Government contract awarded to them at or above $10,000 including GST, and the buy.nsw supplier search shows whether they hold NSW ICT prequalification. All four are free to search and take about ten minutes together.

Q: Do AI consultants need a licence in Australia?
A: No. There is no licence, registration or certification required to call yourself an AI consultant in Australia, and ASIC's professional registers cover financial services licensees, credit licensees, auditors and liquidators rather than management or IT consultants. Anyone can print certified AI consultant on a business card here. The registers that do exist confirm a business exists and how long it has traded, not that it is any good.

Q: How many clients should an AI consultant be able to name?
A: One is enough if you can ring them. Across nine Sydney AI consulting firms checked on 20 August 2026, seven published no client name at all and two named clients, and not one of the nine published contact details for a referee. A consultant bound by a client's confidentiality can still ask that client whether they will take a call from you, so a confidentiality policy is not the same as having nobody to offer.

Q: What should I ask when I call an AI consultant's reference?
A: Ask what broke, who fixed it, how long the handover took, and whether the system is still running today. Ask who on their team uses it and who quietly stopped. A referee who can only tell you the consultant was pleasant to deal with never operated the thing that was built.

Q: Am I protected if an AI consulting contract turns out to be unfair?
A: Partly. Australia's unfair contract terms protections cover a small business with fewer than 100 employees or under $10 million in annual turnover, for standard form contracts made or varied since 9 November 2023, with no contract-value cap. Only a court can declare a term unfair and void it, and the ACCC states it does not investigate individual complaints or provide legal advice. The practical first step in a dispute is the Australian Small Business and Family Enterprise Ombudsman, which offers low-cost mediation for business-to-business contract disputes.

---

### How to measure AI ROI when hours saved are not money

URL: https://www.azgard.tech/learn/how-to-measure-ai-roi
Published: 2026-08-18
Topic: strategy · Tags: roi, measurement, adoption, process · ~18 min read

> Measure AI ROI by baselining the process first: count, time per unit, a named owner, a date, and the one downstream change that turns hours into money.

**TL;DR**

- Hours saved by an AI system become money only when something downstream changes: a role you did not backfill, work you took on that you would have turned away, overtime you stopped paying, or a deadline you stopped missing.
- The Azgard AI Baseline Card records six fields for one process before anything is built - the process, the count over one week, the time per unit, the named person who measured it, the date, and the one downstream change that converts the saving into money - and the same person re-takes it at 30 and 90 days.
- Self-reported AI time savings are unreliable evidence: METR's randomised trial of 16 experienced open-source developers on 246 real tasks found they estimated AI had made them 20% faster while measurement showed tasks took 19% longer.
- AI time savings and AI net outcomes are different numbers: Workday's vendor-sponsored research across 3,200 employees at large enterprises found 85% saved one to seven hours a week while only 14% consistently got clear, positive net outcomes.
- An AI ROI calculator returns the number you fed it, because hours saved multiplied by an hourly rate is an assumption about redeployment, not a measurement of it.

Measure AI ROI on one process before you build, using an Azgard AI Baseline Card: how
often it runs in a week, how long one run takes, who measured it and when. Re-take at 30
and 90 days. Then check the one downstream thing that had to change for the saving to
become money - a role you did not backfill, a deadline you stopped missing. If nothing
downstream moved, the hours are real and the money is imaginary.

This page assumes you have already
[mapped the work](https://www.azgard.tech/learn/stop-buying-ai-tools-before-you-map-the-work), which is where a
defensible baseline comes from. A defensible baseline still is not money.

#### Why AI ROI is hard to measure in a small business

Measuring AI ROI is hard in a small business for three reasons, none of them about AI:
attribution, missing baselines, and no cost centre.

Attribution first. Deloitte surveyed 1,854 executives across Europe and the Middle East
between 15 August and 5 September 2025, and quoted one saying they
["only managed to get a ballpark estimate of the benefits because it was hard to separate the gains from AI initiatives from those of other initiatives"](https://www.deloitte.com/global/en/issues/ai/ai-roi-the-paradox-of-rising-investment-and-elusive-returns.html).
Enterprise, two regions, not a small-business sample - and the problem is worse in a
ten-person business, because you changed three things that quarter and the same two
people did all of them.

Then the missing baseline. Most small businesses cannot say what a process cost before
the AI arrived, so there is nothing to compare against. IBM's study with Oxford
Economics, covering 2,000 CEOs in 33 countries between February and April 2025, found
[only 25% of AI initiatives have delivered expected ROI over the last few years](https://newsroom.ibm.com/2025-05-06-ibm-study-ceos-double-down-on-ai-while-navigating-enterprise-hurdles):
enterprise again, and as much a verdict on measurement discipline as on technology. MIT
Sloan Management Review, on interviews with more than 30 leaders, names the mechanism:
["few companies apply the same financial discipline to artificial intelligence as they would to a new factory or piece of machinery"](https://sloanreview.mit.edu/article/three-approaches-to-measuring-and-managing-ai-roi/).

And the cost centre. A van has a purchase price, a depreciation schedule and a job it
does. An AI subscription has a monthly charge spread across five people who use it
differently and none of whom bills time to it. The spend is visible; the unit it attaches
to is not.

#### Hours saved are not money saved until something changes downstream

Hours saved by an AI system are not money saved until something downstream of those hours
changes: a role you did not backfill, work you took on that you would have turned away,
overtime you stopped paying, a deadline you stopped missing. If none of those move, the
hours are real and the money is imaginary.

Not a new observation, and this page will not pretend otherwise. CIO ran the argument in
March 2026 under the headline
["why tools don't pay off until work changes"](https://www.cio.com/article/4145044/the-davos-reality-check-on-ai-roi-why-tools-dont-pay-off-until-work-changes.html).
The consensus is right. What it lacks is an instrument: the downstream test is nearly
always applied afterwards, as a post-mortem, rather than written down in advance as a
commitment. Meanwhile every failing ROI method fails at the same joint: it multiplies
hours saved by an hourly rate and calls the product money. That multiplication is an
assumption, not a measurement.

| What you measured | What has to change downstream for it to be money | How you check that it did |
|---|---|---|
| Quoting time fell from 3.5 hours to just over 2 | More quotes go out, or the second estimator is never hired | Quotes issued per week; next year's headcount line |
| Two hours a week off bookkeeping | Contracted hours are cut, or that person moves onto collections | Invoiced bookkeeping hours; debtor days |
| Proposal first drafts come back in minutes | Proposals go out same day, and deal count or win rate moves | Enquiry-to-proposal days; proposals per month |
| An assistant drafts 40% of support replies | The support backfill is never hired, or churn moves | Open roles not filled; first-response time; cancellations |
| Fewer hours on the monthly report | The report lands early enough to change a decision | The date it is issued; one decision that moved |

The hours themselves are usually self-reported, and self-reports about AI are measurably
unreliable. METR ran a randomised controlled trial: 16 experienced open-source
developers, 246 real tasks in repositories they already maintained, Cursor Pro with
Claude 3.5 and 3.7 Sonnet.
["Before starting tasks, developers forecast that allowing AI will reduce completion time by 24%. After completing the study, developers estimate that allowing AI reduced completion time by 20%. Surprisingly, we find that allowing AI actually increases completion time by 19%"](https://arxiv.org/abs/2507.09089).
METR later quantified the gap: people
["overestimated AI's effect on their time spent on tasks by 40 percentage points on average"](https://metr.org/blog/2026-05-11-ai-usage-survey/).
Its 2026 survey of 349 technical workers recorded a median self-reported speed change of
3x, with the write-up conceding that "survey results are not necessarily grounded in
reality".

METR's limits deserve as much weight as its headline: 16 developers, all experienced, all
on codebases they knew well, and the paper states plainly that it does not claim its
["developers or repositories represent a majority or plurality of software development work"](https://arxiv.org/abs/2507.09089).
What it measures is self-report, not AI.

Workday's "Beyond Productivity" research, run by Hanover Research across 3,200 employees
at large enterprises and announced on 14 January 2026, puts the gap closer to the P&L:
[85% of respondents saved one to seven hours a week, and only 14% consistently got clear, positive net outcomes from AI](https://newsroom.workday.com/2026-01-14-New-Workday-Research-Companies-Are-Leaving-AI-Gains-on-the-Table).
The hours were real; the money mostly was not. Two caveats: Workday sells AI-powered HR
software, so this is vendor-sponsored research, and the sample is large-enterprise
employees, not small businesses. Neither caveat means the gains are fake. They are real,
and the best-measured ones are counted per unit of work.

Azgard's own number sits at the same joint between hours and money. At Apex Signage, a
Sydney signage manufacturer, quoting went from about 3.5 hours per quote to a little over
two, against a baseline of roughly 12 quotes a week. The arithmetic is sitting right there
and we are not going to do it: 1.5 hours times 12 quotes times somebody's notion of an
estimator's hourly cost is a number nobody measured. The hours are real and checkable in
Apex's own job records. What they become in money depends on the downstream field, and
that figure belongs on Apex's books rather than the marketing page of
[an AI consultancy in Sydney](https://www.azgard.tech/ai-consultant/au).

That number also fails two fields of our own Card.
[Mapping the work](https://www.azgard.tech/learn/stop-buying-ai-tools-before-you-map-the-work) established the
baseline, which is where 3.5 hours came from, but nobody wrote down who took the timings
or on what date. The measurer and the date, fields four and five, are blank. A page
arguing that an unattributed baseline is not evidence has to say that about its own best
number: the Apex figure is a real measurement with no measurer and no date attached to it.

#### The Azgard Three-Question Vendor Filter: what to ask before you sign

The Azgard Three-Question Vendor Filter is three questions to put to any AI vendor,
including this one, before you sign anything: **ask what number they will measure before
they build, who takes the measurement, and the date it gets taken again.** If they can
only tell you the number afterwards, they are selling a story. It is the same discipline
Azgard applies to its own builds, the six-field Azgard AI Baseline Card, turned around and
pointed at the person selling to you, which is why a vendor who cannot answer it is
telling you their own work will not be measurable.

#### The Azgard AI Baseline Card: six fields to fill in before anyone builds

The Azgard AI Baseline Card is a six-field record, filled in for one process, before
anything gets built. One card per process. It exists because "we will see if it works" is
not a measurement plan, and because the field you cannot fill in is the one that predicts
the project.

1. **The process.** One named workflow with a start and an end, not a department and not
   a tool. "Producing a quote from an enquiry", not "estimating".
2. **The count.** How many times it ran over one week, from records you already keep.
3. **The time per unit.** How long one run takes, measured on real runs, recorded as a
   range, not a tidy average.
4. **The measurer.** The name of the person who took the count and the timings. Not a
   team, not a system. A person.
5. **The date.** The day it was taken, so the later comparison is between two dated
   numbers rather than a number and a memory.
6. **The downstream change.** The one thing that has to change for the saving to become
   money, written as a sentence somebody can be held to. "We do not replace the estimator
   who leaves in March." "We stop declining Friday jobs."

The same person re-takes the same six fields at 30 days and again at 90 days.

Field six is the whole argument. Its retrospective version is already common advice:
after the project, ask whether the time got redeployed. Asking afterwards is a
post-mortem. Writing it down beforehand, with a name against it and a date to check it,
is a commitment, and commitments change what gets built. Field six kills projects at the
whiteboard, the cheapest place to kill them. If nobody in the room can say what
would change downstream, the build is a preference rather than an investment.

Fields four and five are the same discipline as
[an owner and an expiry date on a company brain](https://www.azgard.tech/learn/your-company-brain-needs-an-owner-and-an-expiry-date):
a measurement with no name against it decays exactly like a document with no owner.

#### How to baseline an AI project in one week without new tooling

Baselining an AI project takes about a week and needs no new tooling, because the counts
already sit in systems you pay for. Instrument nothing; count what is already there.

- **Volume, from records you already keep.** Export the last eight weeks of jobs,
  tickets, quotes or invoices and count the rows. Eight weeks absorbs one quiet week
  without becoming a research project.
- **Time per unit, from a stopwatch, five times.** Ask the person who does the work to
  time five real runs on their phone. The range matters more than the average, because
  the slow runs are the ones AI changes most.
- **Queue time, from timestamps you already have.** The gap between an enquiry arriving
  and someone logging it is two date fields in an inbox or a CRM. In most small
  businesses the queue is longer than the task.
- **Rework, from the revision trail.** Count how many outputs came back for a second pass
  over those same eight weeks. Without the before number you cannot tell later whether AI
  reduced rework or manufactured it.

Raid your own systems before buying anything that promises measurement. Job management
tools, invoicing systems and shared inboxes already log every transaction they process,
and in most small businesses the count you need is one export away.

If the AI is already running, none of this is closed to you. The same records give a
retrospective baseline: count the eight weeks before you switched it on, take the timings
now, and record on the card that the before-number was reconstructed rather than observed.
That is weaker evidence than a measurement taken in advance, and it is the honest version
of what most businesses actually have.

Rough but honest beats precise and absent. A count taken by hand from eight weeks of job
records, with a name and a date on it, survives a room full of sceptics. A dashboard
figure nobody can trace does not.

#### How to measure AI time savings you can actually defend

AI time savings are defensible when measured per unit of work and fragile when reported
as hours per week. Hours per week moves with volume, staffing and season, so a busy
quarter buries a real improvement and a quiet one invents a fake improvement. Time per
unit holds still while everything around it moves.

Per unit is what the best-run studies measure. Brynjolfsson, Li and Raymond studied 5,179
customer support agents and found AI assistance
["increases productivity, as measured by issues resolved per hour, by 14% on average, including a 34% improvement for novice and low-skilled workers but with minimal impact on experienced and highly skilled workers"](https://www.nber.org/papers/w31161).
Note the unit, and note the distribution: the gain went to the least experienced people,
which changes who you should be measuring.

The preregistered Harvard Business School study of 758 Boston Consulting Group
consultants by Dell'Acqua and colleagues found the same shape with a sharper edge. Inside
the AI's capability frontier, consultants completed 12.2% more tasks, 25.1% faster, at
higher quality. On a task outside it,
["subjects using AI were 19% less likely to produce correct solutions"](https://www.hbs.edu/faculty/Pages/item.aspx?num=64700).
Same tool, same people, and the sign flips with task selection, which is why the Baseline
Card covers one named process rather than a tool. ROI is not a property of the tool.

The practical version is unglamorous. Same person, same kind of work, five timed runs
before and five at each re-take, range recorded alongside the average. If the measurer
changes between the baseline and the 30-day check, you have two numbers rather than a
comparison.

#### The AI KPIs worth tracking, and the soft benefits worth counting but not pricing

Four numbers are enough to track an AI project in a small business: volume, time per
unit, rework rate, and the one named downstream metric from the Baseline Card.

- **Volume.** How many times the process ran. If time per unit falls and volume falls
  with it, you found a slow month, not a productivity gain.
- **Time per unit.** The defensible measure of the saving, taken the same way each time.
- **Rework rate.** The share of outputs needing a second pass. Workday's research found
  [nearly 40% of AI time savings are lost to rework](https://newsroom.workday.com/2026-01-14-New-Workday-Research-Companies-Are-Leaving-AI-Gains-on-the-Table),
  which is vendor-sponsored and drawn from large enterprises, so read it as a warning
  rather than a benchmark for your own business. Rework is the one KPI that can invert a
  project's result, and the one nobody baselines.
- **The named downstream metric.** Whatever field six says: quotes issued, roles not
  filled, overtime hours, debtor days, jobs declined.

Two numbers people ask for that this method deliberately will not produce. **Cost per
task** is time per unit multiplied by an hourly rate, which is the same multiplication
that turns real hours into imaginary money: fine for comparing two ways of doing one job,
worthless as evidence of return, because it inherits the redeployment assumption whole.
And **the ROI of an AI automation** is not a separate calculation from the ROI of anything
else. An automation moves time per unit and volume, so it goes on a Baseline Card like any
other change, and it still has to name the one thing downstream that turns the saving into
money.

IBM splits AI KPIs into hard ROI (labour cost, operational efficiency, conversion,
revenue) and soft ROI (satisfaction, retention, decision making), and says the soft ones
["are often measured with surveys and qualitative research initiatives"](https://www.ibm.com/think/insights/ai-roi).
Deloitte says why that resists a fix: AI
["frequently delivers outcomes that matter"](https://www.deloitte.com/global/en/issues/ai/ai-roi-the-paradox-of-rising-investment-and-elusive-returns.html)
but they "are hard to monetize".

So count the soft event and never price it. "The owner stopped doing quotes on Sunday" is
a countable event with a date on it. Attaching an hourly rate to those Sundays and adding
the total to a business case is where a defensible measurement turns into a sales deck.
Soft benefits are often the real reason a business keeps a system running. They are still
not evidence of return.

#### Why an AI ROI calculator returns the number you fed it

An AI ROI calculator returns the number you fed it, because every input except the
licence fee is an assumption. Hours saved is self-reported, the hourly rate is a choice
and adoption is a guess. The redeployment of the saved hours, the step that makes the
arithmetic true, is not an input at all.

Think Technology Australia, an Australian IT provider, states the consensus method
clearly:
["Before deploying any AI tool, record how long the target task currently takes per week and the hourly cost of the person doing it. Review those same numbers at 90 days."](https://www.thinktechnology.com.au/blog/ai-roi-small-business/)
The instruction is fine. It is the worked example that gives the game away. Their
[published example](https://www.thinktechnology.com.au/blog/ai-roi-small-business/) has a
tool saving five hours a week for one staff member at AU$45 an hour, for an annual saving
of AU$11,700 against AU$1,800 of tool cost. They publish the formula as well: "ROI:
($11,700 minus $1,800) divided by $1,800, times 100 equals 483%."

Run that division and it returns 550%. Nor is it a rounding slip, because 483% would need
a saving of AU$10,500 or a tool cost of about AU$2,006, different numbers entirely. This
is not a gotcha: the author did the harder thing by publishing the formula, and without it
there would be nothing to check at all. It is simply a worked ROI example that does not
survive its own arithmetic. These figures were checked on 18 August 2026; if the page is
corrected upstream, this paragraph is the stale one.

Elsewhere on the same page, 7.5 hours a week recovered is
["roughly $15,000 per year in capacity that can be redeployed or avoided as a future hire"](https://www.thinktechnology.com.au/blog/ai-roi-small-business/),
in Australian dollars. "Can be redeployed" is the joint: the whole return sits inside a
modal verb, stated honestly and never tested. Field six of the Baseline Card exists to
turn that into "will be redeployed, here is who says so, and here is the date we check".

CIO named the same failure and called out calculators
["that treat 'hours saved' as 'money saved' without tying into staffing or throughput plan"](https://www.cio.com/article/4145044/the-davos-reality-check-on-ai-roi-why-tools-dont-pay-off-until-work-changes.html).
Its diagnosis lists the downstream changes that do exist: AI "can shave minutes off
tasks, but P&L only moves when that time converts into throughput, reduced overtime,
fewer contractors, higher conversion or shorter cycle time with the same headcount.
Without an operating plan, saved minutes can simply dissolve into the calendar."

Calculator defaults are old, too. Gartner's July 2024 release is the source of three figures
still seeded into 2026 business cases - 15.8% revenue increase, 15.2% cost savings, 22.6%
productivity improvement - from
[a survey of 822 leaders conducted between September and November 2023](https://www.gartner.com/en/newsroom/press-releases/2024-07-29-gartner-predicts-30-percent-of-generative-ai-projects-will-be-abandoned-after-proof-of-concept-by-end-of-2025).
The same release predicted "at least 30% of generative AI projects will be abandoned
after proof of concept by the end of 2025". A 2023 survey should not be the default value
in a 2026 business case.

Payback runs longer than most calculators assume. Deloitte found
["most respondents reported achieving satisfactory ROI on a typical AI use case within two to four years"](https://www.deloitte.com/global/en/issues/ai/ai-roi-the-paradox-of-rising-investment-and-elusive-returns.html),
against the
[seven to 12 months](https://aws.amazon.com/executive-insights/content/evaluating-the-roi-of-major-tech-investments/)
typically expected of a technology investment, with only 6% reporting payback in under a
year and 13% of even the most successful projects seeing returns within 12 months. That
is enterprise scale in Europe and the Middle East, and one fixed-price build on one
small-business process is a different animal. But if your model shows five months, you
are more optimistic than the median large enterprise.

#### What the "95% of AI projects fail" statistic actually measures

The most-quoted number in AI ROI, the 95% figure, has come loose from the thing that was
measured. Here is the sentence it comes from, in the MIT NANDA report "The GenAI Divide:
State of AI in Business 2025":
["Despite $30-40 billion in enterprise investment into GenAI, this report uncovers a surprising result in that 95% of organizations are getting zero return."](https://mlq.ai/media/quarterly_decks/v0.1_State_of_AI_in_Business_2025_Report.pdf)
Organisations. The same report separately finds that 5% of integrated pilots are
extracting millions, a different population pointing the other way. Elsewhere it calls the
figure a "95% failure rate for enterprise AI solutions". IBM's own AI ROI page renders it as
["95% of generative AI pilots are failing"](https://www.ibm.com/think/insights/ai-roi),
footnoted to that same PDF. One statistic, at least three denominators: organisations,
solutions, pilots. The measured one is the first. Open the PDF and you do not catch
anybody out - you watch the drift happen inside a single document.

The NANDA report also states its own limits: 52 interviews, 153 surveyed leaders,
directionally accurate on individual interviews rather than official company reporting,
and it "may not represent broader market patterns". It is an industry report, not a
peer-reviewed paper.

Two findings in that report beat its own headline. Top performers moved from pilot to
implementation in around 90 days, against nine months or more at large enterprises. And
where returns showed up, they came from reduced external spend: eliminating BPO
contracts, cutting agency fees. Cancelled invoices, in other words. That is the
downstream change, and it sits inside the most-cited AI ROI report there is.

**FAQ**

Q: How do I measure the ROI of AI in a small business?
A: Pick one process and record six things before anything is built: the process, how many times it ran in a week, how long one run took, who took the measurement, the date, and the one downstream change that has to happen for the saving to become money. Have the same person re-take those numbers at 30 and 90 days. The counts come out of records you already keep, so it costs about a week and no new software. If the downstream field is blank, you have a preference, not an investment.

Q: What should I measure before we start an AI project?
A: Volume and time per unit for the specific process you are changing, plus the rework rate and one named downstream metric. Volume is how many times the process runs in a week, taken from job records, tickets or invoices you already have. Time per unit is a stopwatch on five real runs, written down as a range, not a single average. Record who took the measurement and when, because an unattributed baseline is not evidence.

Q: Does AI actually save money, or just time?
A: It saves time first, and that time becomes money only if something downstream changes: a role you do not backfill, work you take on that you would have declined, overtime you stop paying, or a deadline you stop missing. Workday's research found 85% of respondents saved one to seven hours a week while only 14% consistently got clear, positive net outcomes, though that study was vendor-sponsored and covered large enterprises, not small businesses. Measured productivity gains are real, including 14% more issues resolved per hour in an NBER study of 5,179 support agents, but converting them to money is a separate step you have to plan.

Q: How long before an AI project pays for itself?
A: Longer than most business cases assume. Deloitte's 2025 survey of executives in Europe and the Middle East found most reported satisfactory ROI on a typical AI use case within two to four years, against the seven to 12 months that AWS Executive Insights gives as typical for a technology investment, with only 6% reporting payback in under a year. That is enterprise data from two regions, not small business data. One fixed-price build on one process can pay back faster, but if your model shows five months, name the downstream change that makes it true.

Q: Is an AI ROI calculator worth using?
A: Only as a way to write down your assumptions, never as evidence. Every input except the licence fee is an assumption: hours saved is self-reported, the hourly rate is a choice you make, and the redeployment of the saved hours is usually not an input at all. Some calculators are also seeded with defaults from a Gartner survey of 822 leaders conducted in late 2023. Use one to state what you believe, then go and measure the process.

Q: How do I prove AI value to a sceptical business owner?
A: Show a before number they can check in their own records, an after number taken the same way by the same person, and one line on the P&L or the roster that moved. Sceptical owners are rarely sceptical about the hours. They are sceptical about the multiplication. Do not multiply hours by an hourly rate in front of them. Name the downstream change, show the date it was re-checked, and let them decide what it is worth.

---

### How much does an AI consultant cost? The AI spend ladder

URL: https://www.azgard.tech/learn/how-much-does-an-ai-consultant-cost-au
Published: 2026-08-14
Topic: strategy · Tags: pricing, budgeting, vendor-evaluation · ~10 min read

> Expect an AI consultant in Australia to cost $150-$400 an hour or $950-$2,000 a day, and $1,500-$15,000 for a first fixed-price project.

**TL;DR**

- Australian AI consultants charge $150 to $400 an hour and $950 to $2,000 a day in 2026, and a first fixed-price project usually lands between $1,500 and $15,000.
- The Azgard AI Spend Ladder sorts AI consulting spend into five rungs by the unit you buy: an hour, a day, an owned document, a working system, or a month of capacity.
- Buy the lowest rung that produces something you can use without them, because on an hourly meter a vague problem is the buyer's cost and on a fixed price it is the vendor's.
- GST adds 10% to an Australian AI consulting invoice and a registered buyer claims it back, so the same fee costs two different amounts either side of the $75,000 registration threshold.
- No independent Australian benchmark for AI project pricing with a published methodology exists, so the available AUD evidence is vendors' own published prices, recruiter contractor surveys and government contract notices.

An AI consultant in Australia costs
[$150](https://freemeupai.com.au/insights/ai-consultant-cost-australia) to
[$400](https://team400.ai/blog/2026-04-03-ai-consulting-cost-australia) an hour,
[$950 to $2,000 a day](https://bigwavedigital.com.au/ai-salary-guide-sydney/) for a
contract specialist, and
[$1,500 to $15,000](https://freemeupai.com.au/insights/ai-consultant-cost-australia)
for a first fixed-price project. Mid-market production builds run
[$50,000 to $250,000](https://team400.ai/blog/2026-04-03-ai-consulting-cost-australia).
The rate matters less than the unit. An hour puts scope risk on you, a fixed price on the
vendor. Buy the lowest rung that gives you something you can use without them.

All figures are in Australian dollars.

#### What AI consulting costs in Australia: rates per hour, per day and per project in 2026

Australian AI consulting rates come from four kinds of evidence: vendors publishing their
own prices, recruiters surveying contractor day rates, government contract notices, and
nothing else.

| Unit bought | AUD band | Source |
|---|---|---|
| Hourly, independent consultant | $150-$350/hr + GST | [Free Me Up](https://freemeupai.com.au/insights/ai-consultant-cost-australia), Aug 2026 |
| Hourly, boutique AI firm | $250-$400/hr | [Team 400](https://team400.ai/blog/2026-04-03-ai-consulting-cost-australia), Apr 2026, its own tier |
| Hourly, large firm | $350-$600/hr | [Team 400](https://team400.ai/blog/2026-04-03-ai-consulting-cost-australia), Apr 2026, a boutique estimating rivals who disclose nothing |
| Day rate, contract AI/ML specialist | $950-$1,600/day | [Big Wave Digital](https://bigwavedigital.com.au/ai-salary-guide-sydney/), Jul 2026 |
| Day rate, AI architect or consultant | $1,400-$2,000+/day | [Big Wave Digital](https://bigwavedigital.com.au/ai-salary-guide-sydney/), Jul 2026 |
| Day rate, senior data contractor, Sydney | $850-$1,150/day ex GST | [Re:Sourced](https://www.resourced.com.au/articles/contractor-day-rates-australia-guide-2026), Jul 2026 |
| Day rate, named AI contract roles | Agent Architect $1,488, AI Principal Engineer $1,456 | [Talent International](https://www.talentinternational.com/blog/top-tech-contractor-day-rates-australia/), Mar 2026 |
| First diagnostic or roadmap | $495-$5,000 | [Aivy, surveying ten AU price pages](https://aivy.com.au/resources/ai-consulting-costs-australia/), Aug 2026 |
| First project, fixed price | $1,500-$15,000 | [Free Me Up](https://freemeupai.com.au/insights/ai-consultant-cost-australia), Aug 2026: packages from $1,500, structured projects $3,000-$15,000 |
| Mid-market production build | $50,000-$250,000 | [Team 400](https://team400.ai/blog/2026-04-03-ai-consulting-cost-australia), Apr 2026, first-party |

Two caveats decide whether that table helps. The day-rate rows are **contractor** rates,
one person carrying their own super, leave and downtime, so a consultancy sits above them
because it also carries margin and delivery risk. The hourly rows are **vendor-published
bands, not survey data**: prices firms chose to advertise.

**There is no independent Australian benchmark for AI project pricing with a published
methodology.** Not the ABS, not an industry body, not a consultancy research arm.
[Re:Sourced will not publish an AI/ML band at all](https://www.resourced.com.au/articles/contractor-day-rates-australia-guide-2026),
calling those rates "still forming as a distinct market". No large consultancy publishes
an AI rate here either, so the only large-firm evidence is a contract value: McKinsey
Pacific Rim billed Services Australia
[$1,683,000 for sixteen weeks](https://www.tenders.gov.au/Cn/Show/6bcabab9-5273-49d8-a248-1a0fa2f121a1)
of "Artificial Intelligence Design Services".

#### The Azgard AI Spend Ladder: five rungs and five units of purchase

The Azgard AI Spend Ladder sorts AI consulting spend by the unit you buy rather than the
rate you are quoted, because the unit decides who pays when discovery finds a different
problem than the one in the brief.

| Rung | Unit bought | What you own at the end | Who absorbs it if scope is wrong | AUD band |
|---|---|---|---|---|
| 1 | An hour | Notes, and someone else's memory | **You.** Every hour bills whether it finds the right problem or not | $150-$400/hr + GST, [independent](https://freemeupai.com.au/insights/ai-consultant-cost-australia) to [boutique](https://team400.ai/blog/2026-04-03-ai-consulting-cost-australia) |
| 2 | A day | A workshop output, usually a deck | **You**, at eight times the resolution per decision | [$950-$2,000+/day](https://bigwavedigital.com.au/ai-salary-guide-sydney/) |
| 3 | An owned document | A diagnostic, a map of the work, a roadmap you keep | **Shared.** You own the document either way | [$495-$5,000](https://aivy.com.au/resources/ai-consulting-costs-australia/), possibly $0 via an [AI Adopt Centre](https://business.gov.au/expertise-and-advice/ai-adopt-centres) if you qualify |
| 4 | A working system | Software running on real work, documented and handed over | **The vendor**, on a fixed price. Where risk crosses the table | [$50,000-$250,000](https://team400.ai/blog/2026-04-03-ai-consulting-cost-australia) mid-market |
| 5 | A month of capacity | Improvement, and a system that does not rot | **Shared**, renegotiated monthly | Structure, not dollars |

Rung 4 is where most AI implementation cost questions land, and the only published
Australian band for it is one boutique's own price list rather than a market rate.

Rung 5 is where cost quietly becomes permanent, so insist the vendor publishes the structure:
banded hours, a minimum term, the notice to flex scope, and whether unused hours roll over.
Azgard's retainer runs at 8-12, 15-25 or 30-50 hours a month, minimum three months, scope
flexes on 30 days' notice, and unused hours do not roll over. It matters as much that
[someone has to own it after handover](https://www.azgard.tech/learn/your-company-brain-needs-an-owner-and-an-expiry-date):
a retainer that exists because nobody internal owns the system is a subscription to your
own dependency.

#### Buy the lowest rung that produces something you can use without them

**Buy the lowest rung of the Azgard AI Spend Ladder that produces something you can use
without the consultant.** Ask any AI consultant the price of the smallest thing they will
sell you, and whether you still own the output if you never call them again. A vendor whose smallest sellable unit is a
five-figure discovery phase is asking you to fund their qualification process.

The failure is easy to describe: if the business problem is vague, an AI consultant just
turns confusion into an expensive project. What worries buyers is rarely the rate. It is
paying for a roadmap deck, or an agent nobody uses after two months. An hourly meter is
where that costs you most, because vagueness is billable. So
[map the work before you buy anything](https://www.azgard.tech/learn/stop-buying-ai-tools-before-you-map-the-work)
and buy the map as a thing you own.

NSW procurement law draws the line between advice and implementation, which is useful to
point at in a negotiation.
[NSW Procurement Board Direction 2023-05](https://www.info.buy.nsw.gov.au/__data/assets/pdf_file/0009/599814/ict-consulting-commercial-framework-ictc-cf_july2024.pdf)
defines consulting by output: work "where the output is report, advice, strategic plan
or audit as opposed to actual implementation, or 'doing'". A firm engaged to implement,
"supplying or writing the software and constructing the systems, would NOT be a
consultant".

#### What an AI consulting fee includes, and what sits outside it

An AI consulting fee normally covers discovery, build, integration with the tools you
already run, documentation, handover and a short warranty. Azgard's fixed-price Pilot
Build carries a 30-day warranty after handover. Ask which of those six sit inside your
quoted number, because "build" alone is roughly half a project.

Four costs sit outside the fee and outlive the engagement:

- **Model and tool licences.** [Microsoft 365 Copilot Business is AU$26.91 per user per month](https://www.microsoft.com/en-au/microsoft-365-copilot/business) on annual commitment, or AU$37.68 month to month. Microsoft's note on that page: "Prices shown here and on following pages do not include GST."
- **Token or usage spend**, which scales with how much your team actually uses the thing.
- **Your own team's hours** in discovery, testing and adoption, the biggest uncosted line in most AI projects.
- **Ownership after handover**: someone maintaining the system, or a retainer doing it.

#### What Azgard charges: two published prices and four scoped on a call

Azgard publishes two of its six prices. The AI Company Brain Webinar is $149 for a
two-hour live session. The AI Company Brain Kickstart is $1,950, a done-with-you build
across two sessions that ends with a working Company Brain your team owns. Both sit at
rung 3 of the Azgard AI Spend Ladder: a defined artefact you keep whether or not there is
a second engagement.

The other four are scoped on a call and carry no published price: the AI Readiness Audit,
the Pilot Build, the Custom AI System and the Embed Retainer. Two published and four not is
worth saying out loud rather than inventing six numbers, because those four turn on how many
workflows, which systems, and how much data needs untangling first. What is publishable is
their shape: the Embed Retainer, for instance, is sold in bands of 8-12, 15-25 or 30-50
hours a month on a three-month minimum.

What the spend buys, in one checkable case: Apex Signage, a Sydney signage manufacturer,
went from 3.5 hours per quote to just over 2 hours per quote. Home Grown Electrical and
Inner Game Basketball are the other named Australian clients. Azgard is
[an Australian AI consultancy](https://www.azgard.tech/ai-consultant/au) run by one engineer in
[Sydney](https://www.azgard.tech/ai-consultant/au/sydney) who delivers every engagement personally and works
on-site across greater Sydney when it helps.

#### GST, the R&D Tax Incentive and the grants that change what an Australian AI project costs

Three Australian mechanisms move the real cost of AI consulting away from the invoice
number: GST, the R&D Tax Incentive, and government adoption programs.

**GST.** GST is [a broad-based tax of 10%](https://www.ato.gov.au/businesses-and-organisations/gst-excise-and-indirect-taxes/gst/how-gst-works)
on most goods and services consumed in Australia, and consulting is a taxable supply. The
threshold makes it a decision rather than trivia: registration starts at a GST turnover of
[$75,000](https://www.ato.gov.au/businesses-and-organisations/gst-excise-and-indirect-taxes/gst/registering-for-gst),
$150,000 for non-profits. A registered buyer reclaims the 10% as a
[GST credit](https://www.ato.gov.au/businesses-and-organisations/gst-excise-and-indirect-taxes/gst/claiming-gst-credits),
so the fee's real cost is the ex-GST figure; an unregistered buyer under the threshold
pays it and keeps it. Same invoice, two prices. Two Australian quotes are not comparable
until you know which side of GST each one sits on.

**The R&D Tax Incentive, and the hedge it needs.** The rates are real: under $20 million
turnover the refundable offset is
[your company tax rate plus 18.5%](https://www.ato.gov.au/businesses-and-organisations/income-deductions-and-concessions/incentives-and-concessions/research-and-development-tax-incentive/r-d-tax-incentive-rates-and-entitlements/rates-of-r-d-tax-incentive-offset);
above it, the non-refundable offset runs at the tax rate plus 8.5% up to 2% R&D intensity
and plus 16.5% above. The
[gates](https://business.gov.au/grants-and-programs/research-and-development-tax-incentive/overview-of-rd-tax-incentive)
are narrow: an incorporated company, at least one core R&D activity, over $20,000 of
eligible spend, registration within 10 months of year end. A core activity needs an
outcome that could not have been known in advance, so configuring an existing model into
an existing workflow fails on its face, and no ATO or business.gov.au guidance says AI
implementation qualifies. Genuine experimental development is another matter: Australian
businesses put
[$668.3 million into AI R&D in 2023-24](https://www.abs.gov.au/media-centre/media-releases/ai-now-fastest-growing-area-business-rd),
up 142% from $276.3 million in 2021-22, though that total is enterprise-weighted. The rules
are also tightening: a reform
[announced on 12 May 2026](https://www.ato.gov.au/about-ato/new-legislation/in-detail/businesses/tax-reform-better-targeting-the-research-and-development-tax-incentive)
and "not yet law" lifts the minimum annual spend from $20,000 to $50,000 from 1 July 2028.

**Programs that can make rung 3 free.** Four AI Adopt Centres (SMEC AI, ARM Hub, SAAM and
ARAIN) are
[currently listed by business.gov.au](https://business.gov.au/expertise-and-advice/ai-adopt-centres)
as giving "free specialist services for eligible SMEs in National Reconstruction Fund
(NRF) priority sectors", including one-on-one consultations and roadmaps. The NRF limit is
the catch: it is not every Australian business.
[ARM Hub states it has worked with over 300 Australian SMEs](https://aiadopt.ai/) and
offers a free readiness assessment. The
[$17 million program](https://www.industry.gov.au/news/four-new-centres-help-australian-businesses-adopt-ai)
that established the centres is closed to new centre applications, not to businesses using
them. The
[National AI Centre](https://www.industry.gov.au/science-technology-and-innovation/technology/artificial-intelligence/national-ai-centre)
routes businesses to the centres but gives guidance, not money. Queensland's 2026-27 budget
[announced a $10 million program](https://statements.qld.gov.au/statements/105371) for
small and family businesses adopting AI; as at August 2026 it is absent from Business
Queensland's [grants schedule](https://www.business.qld.gov.au/running-business/support-services/financial/grants/schedule),
so treat it as announced rather than open if you are buying in
[Brisbane](https://www.azgard.tech/ai-consultant/au/brisbane).

**Terms worth borrowing.** The
[NSW ICT Consulting Commercial Framework](https://www.info.buy.nsw.gov.au/__data/assets/pdf_file/0009/599814/ict-consulting-commercial-framework-ictc-cf_july2024.pdf)
caps its suppliers' daily rates and publishes the rate table with every cell blank, marked
commercial in confidence. It does publish the terms, and a private buyer can ask for the
same ones: mandatory volume discounts of 2% above $250,000, 5% above $500,000 and 7% above
$1 million, and disbursements capped at 7% of engagement fees. If the largest ICT buyer in
the state writes that down, a mid-sized business can ask for it.

**FAQ**

Q: How much does an AI consultant cost for a small business?
A: In Australia, a small business should expect $1,500 to $15,000 for a first fixed-price piece of AI work, and $495 to $5,000 for a diagnostic or roadmap on its own. Hourly engagements run $150 to $400 an hour, which is the most expensive way to buy a problem nobody has scoped yet. If you are an SME in a National Reconstruction Fund priority sector, an AI Adopt Centre may do the readiness work for nothing.

Q: What kind of first project scope and pricing is reasonable?
A: One workflow, one owner, a fixed price, and a deliverable you keep. Quoting, scheduling, invoicing and client follow-up are the usual first candidates, because they repeat and they are already documented in your systems. If a vendor's smallest sellable unit is a five-figure discovery phase, you are funding their qualification process rather than buying a result.

Q: Why won't AI consultants publish their prices?
A: Because a published price commits them to a scope, and most AI consulting is sold before the scope is known. Australian boutiques increasingly do publish bands, but none of the largest consultancies publishes a rate or an indicative price for AI work here. The workaround is to ask for the price of the smallest thing they will sell you, which a vendor can answer without a discovery call.

Q: Do I pay for AI tool licences on top of the consulting fee?
A: Yes, almost always. Model and tool licences, token or usage spend, and your own team's hours sit outside the consulting fee and continue after the consultant leaves. Microsoft 365 Copilot Business, for example, is AU$26.91 per user per month paid yearly, and Microsoft states those prices do not include GST.

Q: Does GST apply to AI consulting fees in Australia?
A: Yes. AI consulting is a taxable supply, so a GST-registered Australian consultant adds 10% to the fee. If your business is GST-registered you claim that 10% back as a GST credit, so the real cost is the ex-GST number. If your GST turnover is under $75,000 and you are not registered, you pay the 10% and keep it as a cost.

---

### Same model, 8.3x the output. The gap is AI skills, written down

URL: https://www.azgard.tech/learn/same-model-8x-the-output-the-difference-is-written-down
Published: 2026-08-13
Topic: perspectives · Tags: skills, company-brain, adoption, context · ~10 min read

> Frontier firms get 8.3x the output per user from the same models. OpenAI's data points at written-down AI skills and connected company data.

**TL;DR**

- OpenAI's own telemetry shows frontier firms generating 8.3 times the output tokens per active user that typical firms do, up from 2.6 times in January, while both groups had access to the same models.
- OpenAI lists skills and plugins among the differentiators between frontier and typical firms, with 19% of weekly active users at frontier firms using skills against 3% at typical firms.
- Anthropic removed over 80% of Claude Code's system prompt with no measurable loss on its coding evaluations, which shows generic written instructions are waste while proprietary decision logic is not.
- MIT Technology Review and Google Cloud found only about 45% of company data is reachable by agentic AI on average, and every organisation that got past 70% reported trusting its agents' accuracy.
- Upwork found 74% of small-business leaders report productivity gains from AI while those gains stay under 25% for most, which is the signature of model access with nothing wired behind it.

OpenAI published its enterprise telemetry yesterday, and buried in it is the end of an
argument I have been having with business owners all year.

Firms it classes as "frontier" now generate 8.3 times the output tokens per active user
that typical firms do. In January that multiple was 2.6. The gap more than tripled in
five months, and it did that while both groups sat in front of the same models on the
same subscriptions. OpenAI puts it plainly: frontier firms
["have access to the same AI models as other enterprises, but they are deepening their use faster"](https://openai.com/signals/enterprise-data/).

So the question I get asked most often, which model should we standardise on, turns out
to be close to irrelevant. What matters is the fortnight after you pick one.

#### Why do some companies get so much more out of AI than others

OpenAI answers this in its own report, and the answer has nothing to do with how clever
the model is. Two features separate the groups. Skills, which are reusable written
instructions the model loads when it needs them. And plugins, which connect it to the
systems where company data lives.

At frontier firms, 19% of weekly active users use skills. At typical firms it is 3%.
Plugins run 21% against 9%.

That 3% is the number I would put on a wall. In the overwhelming majority of businesses
nobody has written down how the place makes its decisions in a form a model can pick up.
Not because it is hard. It has never been anybody's job.

The industry breakdown says the same thing sideways: information and technology at 11.7x,
manufacturing at 5.3x. Typical firms did not stand still either, growing 1.9x to 2.8x
across the year. Genuine improvement, in a straight line, while the top of the market
compounded.

Credit to OpenAI for putting the caveat in its own report: "Tokens are an imperfect
measure of business value: a short response can be highly valuable, while a long one may
add little." Worth holding onto. A firm burning eight times the tokens is not
automatically producing eight times the value, and I would not sign off on a business
case built on token counts. A 2.6x gap widening to 8.3x in five months is still not
noise.

#### The engineering head start was a filing head start

The same report explains why engineering ran ahead of every other function, and this is
the line I would print out. "Codebases give agents clear context, tests make outputs
easier to verify." Knowledge work lags because its tasks "provide limited context, can be
difficult to specify, and lack clear criteria for verifying the result."

Read that as a business diagnosis rather than a technical one. Engineering got there
first because a codebase is a business that already wrote itself down: one place, one
format, with an automated way to check whether the answer was right. Every other function
in your company runs on rules that were never written anywhere, producing outputs nobody
agreed how to grade.

Codex accounted for
[64% of combined Codex and ChatGPT output tokens](https://openai.com/index/how-enterprises-put-ai-to-work/)
among enterprise customers as of June. The interesting part is what has happened since
February. Enterprise Codex use grew 108x in
legal, 41x in sales, 41x in recruiting and 26x in marketing, against 5x in engineering.
None of those functions have a repo. They are getting there anyway, which is the
encouraging read: the context problem is solvable outside a codebase, it just has to be
solved deliberately.

#### Anthropic deleted 80% of its instructions and lost nothing

Now the argument being thrown at everything I have just written.

In a post from late July, Anthropic reported that it
["removed over 80% of Claude Code's system prompt"](https://claude.com/blog/the-new-rules-of-context-engineering-for-claude-5-generation-models)
for models like Claude Opus 5 and Claude Fable 5, with no measurable loss on its coding
evaluations. It has been circulating hard this week, and the take riding on it is blunt:
delete your instruction files, the model already knows.

The resolution sits in the same post. Skills earn their keep when they "encode particular
opinions, knowledge, or best practices that are particular to you, your team, or
product". Alongside it: "Avoid stating 'the obvious' things Claude should know by looking
at your file system or your repo."

The 80% Anthropic cut was the generic 80%. Instructions telling a very capable model
things it could work out by looking around.

Which means the frontier gap is not about volume of writing at all. It separates
businesses that have written down the decision logic nobody outside the building could
infer from businesses that have written nothing, or that have written forty pages of
generic AI advice the model never needed. A fat prompt full of common knowledge costs you
money and buys you nothing. Two pages naming your margin floor, your approval thresholds,
the exceptions your best estimator applies without thinking, and what you do with a job
you have never quoted before: that is an asset, and no vendor can hand it to you. It is
the same forty pages I was banging on about
[when Meta made the model free](https://www.azgard.tech/learn/meta-made-the-model-free-your-context-is-the-bill),
arrived at from the other direction.

#### What AI skills and plugins for business actually are

Strip the vendor vocabulary and there are two jobs.

A skill is a written instruction for a decision your business makes over and over. How a
quote gets built, or when a discount needs sign-off. It sits in a file, the agent loads
it when that situation comes up, and it improves every time somebody corrects it. The
hard part is not technical: writing one forces you to state a rule you have been carrying
around unstated, and occasionally to find that two people in your business hold different
versions of it.

A plugin, or app, or connector, whatever this month's word is, is a live line into a
system holding facts: your job management platform, your accounting file.

Both are things you build once and maintain, and neither shows up well in a demo.
That is most of the reason so few businesses have them: nobody sells a quarterly retainer
for a decision you wrote down yourself.

#### How to give AI agents access to company data safely

The access half of the gap picked up its own number this week. MIT Technology Review,
working with Google Cloud,
[surveyed 300 data and technology executives](https://www.nojitter.com/data-management/mit-agentic-ai-will-fail-without-a-solid-data-foundation)
and found that on average only about 45% of company data is reachable by agentic AI. The
organisations it classes as data leaders get past 70%. Everybody else sits at 30% or
below.

Then the finding I keep rereading. 100% of those data leaders say they trust the accuracy
and decisions of their agents. Across the full sample it is about half.

The obvious reading is that confidence follows competence. I think the causation runs the
other way. An agent working from 30% of your data is guessing at the rest, producing
plausible answers that are wrong in ways nobody can trace back to a missing file.
Distrust is the correct response to that agent. The leaders did not talk themselves into
trusting theirs. Their agents can see the file.

Which is where the safety question gets asked backwards. In
[Upwork's SMB research](https://www.upwork.com/resources/state-of-ai-in-smbs), data
security is the single biggest barrier to adoption at 27%, and I have never thought that
concern was silly. But the failure I actually get called about is not
a leak. It is an agent with access so thin that it improvises, and improvisation on a
customer quote costs more than a permissions review would have. Scope access per system
and per person, write down what the agent may not touch, and you are most of the way
there.

The rest of the MIT numbers describe the market I walk into every week: 98% of those
organisations use agentic AI, 73% on a limited number of use cases, and one in ten
widely. 66% say legacy systems are what stops them scaling.

#### The strongest argument against writing anything down

Skan AI [raised a $63M Series C this week](https://siliconangle.com/2026/08/12/skan-ai-raises-63m-give-ai-agents-map-enterprise-work/),
co-led by Cathay Innovation and Dell Technologies Capital, on the premise that enterprise
work context is becoming an infrastructure layer the way CRM became the system of record
for customer data. Its method is observation: it watches how work actually moves through
an organisation. At one large US bank it says it tracked 11.2 million context switches
across 1,500 finance staff and surfaced $37M in operational friction. Those figures are
company-supplied, so hold them loosely.

The line that should sting comes from the investor. Simon Wu at Cathay says Skan is the
only company he has seen building work context "from direct observation rather than from
documentation or system logs". And Avinash Misra, the chief executive: "Everyone is
obsessed with building a better car. We think the bigger opportunity is building a better
navigation system."

I take that seriously because I made half of the argument myself a couple of days ago.
Documentation lies. Your written policy and your practised policy drift apart, and
[the written one ages worse](https://www.azgard.tech/learn/your-company-brain-needs-an-owner-and-an-expiry-date).
Observation flatters nobody. It surfaces the workaround your team invented in March and
never mentioned upstairs.

Where I get off the bus is what observation can produce. It answers "what happens here".
An agent making a decision needs "what should happen here". Watch a sales team for a year
and you will learn that they discount inconsistently. You will not learn what the
discount should be, because that number is a judgement somebody has to make out loud.

Observation is a superb instrument for finding what needs writing down. Somebody still
has to write it.

#### Why your AI pilot is not delivering productivity gains

The encouraging thing for a small business is that none of this is capital-intensive.

Upwork's [research on AI in SMBs](https://www.upwork.com/resources/state-of-ai-in-smbs),
published in June, covers 195 leaders at firms of 10 to 99 people. 62% are very or
extremely confident handing high-stakes tasks to AI agents. 32% already describe agents
as mission-critical. 74% report improved productivity, and then the sting: those
improvements "remain under 25% for most".

Conviction is there. Access is there. Returns stalling in the low twenties inside a
business that already believes and already has the tools is the signature of model access
with nothing wired in behind it.

One more figure worth sitting with. Six months after adoption,
[OpenAI found](https://openai.com/index/how-enterprises-put-ai-to-work/) early-career
employees sending 13 more messages per week than executives. The people who know the
firm's rules are barely touching the thing, and the people using it hardest have not
learned the rules yet. A written skill is the only mechanism I know that closes that gap:
the senior person's judgement, in a file, available to the graduate at nine on a Tuesday
night.

#### What a small business needs before AI agents actually work

Roughly four things, none of which need a procurement process.

- One workflow where the decisions carry real money. Not the one that irritates you most.
- The rules that govern it, written where an agent can read them: current numbers, the
  exceptions, and an honest note on what the rules do not cover.
- A connection into whichever system holds the facts, so the agent stops guessing at the
  70% it cannot see.
- A person's name against it, and a date to check it again.

Then delete everything you wrote that a good model could have worked out on its own. That
step is what Anthropic's post is really about, and it is the difference between a skill
and a prompt full of throat-clearing.

The firms pulling away are not smarter and they are not better funded. They wrote it
down, and five months ago they were only 2.6x ahead. If you are waiting on the next model
release to change your position, you are watching the wrong number, and buying another
tool before you have [mapped the work](https://www.azgard.tech/learn/stop-buying-ai-tools-before-you-map-the-work)
will not move it either.

**FAQ**

Q: Do I need to write documentation for AI agents?
A: Not documentation in the usual sense. A capable model does not need explanations of things it can work out by looking at your files, which is why Anthropic could delete over 80% of Claude Code's system prompt without losing accuracy. What it cannot infer is your proprietary decision logic: margin floors, approval thresholds, and the exceptions your best estimator applies without thinking. Write that down and skip the rest.

Q: What are AI skills and plugins for business?
A: A skill is a reusable written instruction for a decision your business makes repeatedly, stored in a file the agent loads when that situation comes up. A plugin or app is a live connection to a system holding your facts, such as a job management platform or an accounting file. OpenAI's data shows 19% of weekly active users at frontier firms use skills, against 3% at typical firms.

Q: Why is our AI pilot not delivering productivity gains?
A: Usually because the agent has neither your written rules nor access to your systems, so it improvises. Upwork found 74% of small-business leaders report productivity improvements, but those improvements stay under 25% for most of them. Real usage plus stalled returns almost always means the model got switched on and nothing was wired in behind it.

---

### Meta made the model free. Your context is the bill.

URL: https://www.azgard.tech/learn/meta-made-the-model-free-your-context-is-the-bill
Published: 2026-08-12
Topic: perspectives · Tags: open-weights, context, company-brain, agents · ~9 min read

> Meta's Muse Glimmer runs a capable agent on one GPU. A 128K context window is rationed VRAM, so running AI agents locally rewards the tidy business.

**TL;DR**

- Meta's Muse Glimmer makes a capable agentic model free, private and local, which moves the binding constraint from the model to the quality of your written-down company context.
- On a local model the context window is rationed VRAM, because Meta's own figures put the 4-bit weights under 20 GB and leave the KV cache to share a 24 GB or 32 GB card with them.
- A 128,000-token context window holds roughly 96,000 words, about a 300-page book, which is more decision-bearing knowledge than most small businesses have ever written down.
- Oracle's agent-memory research beat a full-transcript baseline while using about ten times fewer input tokens, so curated context outperforms complete context even when the window is large.
- Redis found 97% of leaders believe context decides whether their agents work and only 4% have built for it.

Meta put Muse Glimmer out on Monday and my feed did the thing it does. Open weights,
Apache 2.0, thirty billion parameters, tuned for "always-on local agent workflows",
small enough to sit on one consumer graphics card. If you run a small business and you
have spent the past year being quoted per-seat pricing on cloud AI, that headline reads
like the bill just went to zero.

The bill didn't go to zero. It moved somewhere you can't pay it with a credit card.

#### A 128K context window is a budget, not a spec-sheet line

Read [Meta's launch post](https://research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model)
past the headline and the interesting part is the accounting. At full precision the model
"would require over 55 GB of memory". Quantised to roughly four bits, the language model
lands "under 20 GB", which Meta says "leaves enough headroom for the model's working
memory (its 'KV cache'), the perception encoder... and the speculative decoding drafter to
run simultaneously within a 24 GB or 32 GB envelope."

Headroom is the operative word. Your context window is whatever is left on the card once
the weights have taken their cut. Every token of company context you feed
the thing, a price list, a scope-of-works template, six months of job notes, is physical
memory you took away from something else.

None of that is your problem on a cloud frontier model. You get a million-token window,
you shovel a disorganised business into it, and you pay the difference in dollars. Crude,
but it works, and for plenty of SMBs it has quietly been the plan: skip the filing, rent
a bigger bucket.

Put the same agent on a card under your desk and the bucket is fixed at whatever you paid
for. The mess now has a hardware price.

#### The first thing practitioners argued about was filing

The [Hacker News thread](https://news.ycombinator.com/item?id=49241679) hit 1,182 points
in a day. Plenty of it is the usual release-day scrap over benchmarks and how it stacks up
against Qwen and Gemma. The complaint that kept surfacing underneath all that was the
context budget.

One commenter wrote off the 128K window as a complete non-starter for his team, because
their average context runs well above it. Another called it "the biggest apparent
limitation of this model". A third had already rope/yarn-scaled it to 256K and reported
running 216K tokens deep on a single Radeon R9700, which is a lovely hack and also tells
you precisely which problem he was solving.

Day one of the free local agent era, and the people planning to actually run the thing
were arguing about how much of their business fits in the window.

Redis published its
[State of Context Engineering 2026](https://redis.io/resources/state-of-context-engineering-2026/)
this month with the numbers that go alongside that: 97% of leaders believe context decides
whether their agents work, and 4% have built for it. Another 73% say their agents fail
more often from broken context than from broken models, and 81% are still sitting in the
two earliest maturity stages. The finding I keep coming back to is that 83% think fresh
context matters more than adding model parameters, which is a funny thing to believe in
the same week everyone queued up to download more parameters.

#### How much company context fits in a 128K context window

128,000 tokens is around 96,000 words. Roughly a 300-page book.

So: could a competent stranger run one day of your business from a 300-page book you
handed them? Not the whole business. One day. Quoting, scheduling, knowing what to charge,
knowing which supplier to ring and what your margin floor is before you walk away from a
job.

Most SMBs I have sat with do not have 300 pages of decision-bearing knowledge. They have
maybe forty pages of it. What they have is 40,000 files. The pricing logic is real and it
is perfectly legible to the person who invented it, it just lives across a spreadsheet,
three email threads and one estimator's memory.

Which is why I think the small window is good news. It sends you looking for the forty
pages that decide things and gets you to put them somewhere an agent can walk through.
That is work worth doing whether or not you ever run a model on your own hardware.

What I find persuasive is that the size a human can keep true and the size that fits on
the card come out about the same. Keeping a brain honest means
[holding the canon small enough that one person can reread all of it in an afternoon](https://www.azgard.tech/learn/your-company-brain-needs-an-owner-and-an-expiry-date),
which is an argument about ownership and staleness and has nothing to do with silicon.
Meta's VRAM accounting arrives at roughly the same size from the other end. When a
discipline problem and a hardware constraint agree on a number, I stop treating that
number as a matter of taste.

#### Curated context beats complete context, even when you can afford both

The instinct when you hit a context limit is to fight the limit: scale the rope, buy a
bigger card, wait for the local model with the million-token window. Worth knowing that
the load-everything approach loses even when you can pay for it.

Oracle published
[research on agent memory](https://blogs.oracle.com/developers/agent-memory-is-a-database-problem-oracle-research-makes-the-case)
the same day Glimmer landed. Their curated memory system scored 93.8% on LongMemEval while
holding about 1,300 input tokens per request at turn 80. The comparison was flat
conversation history: the full transcript, nothing dropped, about 13,900 tokens per
request. Curation won 48, lost 13, tied 19. Their line for it is "bounded, relevant context
beat complete, unfocused context".

Ten times fewer tokens and better answers, against a baseline that could see everything.
That is the case for a company brain in one sentence, and it holds whether your window is
128K or a million.

#### Running AI agents locally for a small business is a filing problem

We describe AI capability to clients in four parts: what the system can **know**, **do**,
**reach** and **remember**.

Glimmer, plus [NVIDIA's local-agent push](https://blogs.nvidia.com/blog/local-ai-open-source-models-agents-nemotron/)
the following day, hands a small business do and reach for the price of a graphics card.
Tool calling, multi-step execution, picking itself up after a failed step, all of it
private and offline. That used to be the expensive half.

It hands you nothing on know and remember. Look at NVIDIA's own list of what this is for.
"Private data processing: Read, summarize and act on local files, documents, emails and
messages." "Long-running workflows: Break larger projects into steps, track progress and
resume interrupted sessions with context intact."

That is a company-brain spec with no brain supplied.

Meta says it more plainly than anyone. An agent that "manages your schedule, drafts your
messages, organizes your files, and learns how you work needs deep access to personal
context." Agreed. And then the post ships the model, because the model is the part Meta
can ship.

We talk with clients about five levels of AI maturity. What happened this week is that the
middle levels got cheap and the top one, an agent working with your whole organisation's
context behind it, didn't move at all.

Which is the same reason a model you downloaded is still
[a tool you bought before mapping the work](https://www.azgard.tech/learn/stop-buying-ai-tools-before-you-map-the-work).
Free and local doesn't excuse you from knowing which job it is for.

#### What an agent-navigable business looks like

The signage manufacturer we work with is the cleanest example I have. Apex's
decision-bearing context is small: how a job gets priced, what suppliers charge this month,
the rules an estimator applies when a drawing is ambiguous, which enquiries are worth
quoting at all. Written down properly, that is a short, structured set of documents an
agent can navigate.

It did not start that way. It started as years of files, a spreadsheet with a tab per
product type, and pricing rules that lived in one person's head. Most of our work with them
has been converting the second thing into the first. None of it was about models. All of it
is what makes a model useful, and it would fit inside a 128K window with room left over.

That is my answer to a question I get asked in different words every month: what do you
need before an AI agent can use your company files? Not a platform. A short, current, owned
set of documents that say how decisions get made, and somebody whose job it is to keep them
true.

#### Where I think this gets oversold

Two things cut against the local story, and I would rather say them than have you find out
in month three.

Hardware first. The Register
[measured](https://www.theregister.com/ai-and-ml/2026/08/10/zuck-rekindles-open-weights-llama-drama-with-muse-glimmer/5285666)
about 12.2 tok/s on a DGX Spark, and reckons that without a dedicated GPU with enough
memory "the best you can expect is around 6 to 14 tok/s". That is a model you can watch
typing. Most Australian SMBs do not have an RTX 5090 under a desk and I am not going to
tell a plumbing business to buy one.

Second, the cloud frontier models are still better at hard reasoning, and for most owners
the privacy case for local weights is weaker than it feels. The enterprise agreement you
already signed probably covers what you are actually worried about. Open weight model
versus cloud AI is a real decision for a clinic sitting on health records; for most trade
businesses it is a preference, not a requirement.

So this isn't a "go local" piece. The point is that the work you would do to make a local
agent useful is the same work that makes a cloud agent useful, and it is the only part of
the stack nobody can hand you for free. The Register reads Glimmer as positioned for
"small-to-medium sized enterprises or enthusiasts". Fair enough. The ones it will work for
are the ones who did the filing.

#### Where I would start this week

- Pick the workflow where the decisions are worth the most money, not the one that annoys
  you most.
- Write down the rules that govern it. Real rules, current numbers, one place.
- Give it an owner and a review date. A brain that has gone stale is worse than no brain,
  because people trust it.
- Then point whatever model you like at it, cloud or local, and watch where it gets
  confused. That is your next page to write.

Do that and the next open-weights release is a genuine upgrade for you. Skip it and you
will have downloaded a very fast machine for reading a mess, decided local models are not
there yet, and been wrong about why.

If you want a second opinion on which forty pages actually matter, that is what our
[discovery call](https://www.azgard.tech/contact) is for.

**FAQ**

Q: Is a local open-weight model good enough to run my business?
A: For narrow, repeatable jobs against well-organised context, yes. For hard reasoning the cloud frontier models are still ahead. The bigger practical limit is speed: without a dedicated GPU with enough memory you are looking at roughly 6 to 14 tokens per second, which is fine for an overnight batch and painful for anything interactive.

Q: How much company context fits in a 128K context window?
A: About 96,000 words, or roughly a 300-page book. That sounds small until you try to write down the rules that actually govern your business, at which point most owners find they have closer to forty pages of genuinely decision-bearing knowledge scattered across thousands of files.

Q: Do I need a knowledge base before using AI agents?
A: You need the decision-bearing part of one. Not every document you own, but the current rules, numbers and exceptions that determine how work gets priced, scheduled and approved. Curated context consistently outperforms dumping everything in, so the work is choosing what matters rather than uploading more.

---

### Your company brain needs an owner and an expiry date

URL: https://www.azgard.tech/learn/your-company-brain-needs-an-owner-and-an-expiry-date
Published: 2026-08-11
Topic: perspectives · Tags: company-brain, governance, agents, context · ~8 min read

> Agents cite stale internal docs because retrieval treats every source as equal. Split your company brain into a wide evidence layer and an owned canon.

**TL;DR**

- A company brain needs two tiers that never blur: a wide evidence layer that can hold anything, and a narrow canon layer where every document has a named owner and a review date.
- Agents produce confident wrong answers because retrieval ranks sources by similarity to the question, not by whether anyone still stands behind them.
- Crawling Slack is worth doing, because informal chatter is the only place you find out that the written policy and the practised policy have drifted apart.
- The cheapest governance fix for a small business is writing down what each canonical document does not cover, so an agent stops applying it to cases nobody approved.
- Nobody has solved freshness, so make staleness visible instead: keep canon small enough that one person can reread all of it in an afternoon.

Two arguments about company knowledge landed on the same day last week. They flatly
contradict each other, and I think both sides are half right.

On 10 August, Flo Crivello launched Lindy Teammate and
[made the case against curation](https://www.cognitiverevolution.ai/lindy-teammate-flo-crivello-on-multiplayer-agents-memory-why-he-d-ban-the-chinese-models-he-uses/)
on The Cognitive Revolution. Wikis fail, he argues, because "the moment written
documentation is written, it's out of date." So Teammate skips the curation step. Its
onboarding routine, which he calls hydration, crawls all of Slack, mess and all, because
agents don't mind the mess. Three to five million tokens for a twenty-person company with
a few years of history behind it.

Same day, No Jitter published
[a piece on making institutional knowledge AI-ready](https://www.nojitter.com/data-management/make-your-institutional-knowledge-ai-ready)
that opens with exactly the failure a crawl produces. An agent applies the documented
Policy A. The sales team has quietly been running Policy B for months. Kuber Sharma of
UiPath puts it about as well as it can be put: "The agent confidently produces a wrong
answer based on a wrong source. That's worse than producing no answer. It's an authority
failure with an AI signature on it."

Both camps are arguing about volume. How much do you feed the thing? I don't think that
is the interesting variable. What decides whether your company brain helps or hurts is
provenance: for any sentence the agent reads, does anybody still stand behind it?

Mess makes perfectly good evidence. It makes lethal law. Most businesses are about to
build one undifferentiated pile in which those two are the same thing.

#### Why AI agents give confident wrong answers from internal docs

Retrieval ranks by similarity to the question. It does not rank by standing. A Slack
message from March 2024 saying "yeah just waive the fee for anyone over $5k" and an
approved rate card from last month look like equally good candidates to an embedding
model. The Slack message often wins, because people ask questions in the register of
Slack, not in the register of policy documents.

Then the agent does the job it was built for. It answers in a confident voice and it
attaches a citation.

That citation is the part that does the damage. A colleague who gets it wrong out loud
gets challenged, because everyone can hear the guess in their voice. An agent that
produces the wrong answer with a link attached gets believed, and then gets pasted into
an email to a customer. The failure isn't that the model lied. It did its job faithfully
on a source that had no business being treated as authoritative.

#### Crawl the Slack anyway

Here is where I part company with the tidy camp, because Crivello is right about the
thing they keep glossing over. Wikis really do rot. The curation-first pitch has been
made to small businesses for fifteen years and I have never once walked into a
twenty-person company with a maintained wiki. If the advice has a nil success rate across
that many attempts, the problem is the advice.

The stronger point in favour of crawling is one I don't think either side has made
loudly enough. The mess is the only place the Policy B problem is visible at all. A tidy,
curated-only brain will answer the Policy A question flawlessly and will never know that
sales stopped doing it that way in March. The informal record is a detection instrument.
It tells you what your business actually does, which is frequently not what your
documents say it does.

So crawling earns its place for a better reason than cost. It buys you something curation
cannot buy at any price: a live picture of what your business is actually doing. Issuing
rulings is a separate job.

#### What should go into a knowledge base for AI agents

Two tiers, and they never blur into each other.

The evidence layer is wide and cheap and makes no promises. Slack, email threads, meeting
transcripts, superseded drafts. The agent may quote any of it with attribution attached
and a date on the front, and may never convert it into a rule.

The canon layer is narrow. Every document in it carries four things: a named owner who is
a person rather than a department, a last-reviewed date, an explicit statement of what
the document does not cover, and a retirement path.

That third one gets skipped by nearly everybody and it is the cheapest insurance in the
system. A refund policy written for online orders will be applied cheerfully to a trade
account, because nothing in the document ever said it wasn't about trade accounts. Humans
infer scope from context they absorbed in the room where the policy was argued out. The
agent was not in the room. So write the boundary into the document: this covers standard
online orders, it does not cover rush jobs or trade accounts, and if you are being asked
about those, no approved policy exists yet.

Which gives you a rule simple enough to actually enforce. The agent may only state a rule
that comes from canon. Where canon is silent, the correct answer is "there's no approved
policy on this, though here's what people have said about it in #sales". Nobody ever gets
fired over that sentence.

#### Company brain vs enterprise search for a small business

Enterprise search answers the question "what can be found". A brain has to answer "which
version is valid". Those are different products, and the market is starting to say so out
loud.

[Stack Internal](https://stackoverflow.co/internal/), formerly Stack Overflow for Teams,
now sells itself as "the trusted knowledge layer" that captures, curates, validates and
delivers knowledge with "your own version of our iconic green check mark", under the
banner "where human and agentic curation meet". Meanwhile Tencent Cloud
[open-sourced a team memory hub](https://www.marktechpost.com/2026/08/07/tencent-cloud-open-sources-tencentdb-agent-memory-v2-0/)
whose write-up concludes that governance, not retrieval, is the actual differentiator
over standard RAG. New memories there default to private, so sharing is a deliberate act
rather than a side effect.

An American developer community and a Chinese cloud provider have landed in the same
place from opposite directions. Both treat the standing of a document as a first-class
field, on the same footing as its text.

#### Who owns the knowledge an AI agent relies on

Sharma's checklist in the No Jitter piece is five questions long: who owns it, who updates
it, who approves changes, who retires obsolete documents, and who audits the agents when
official docs change. For a twenty-person business, answering those is a
forty-minute exercise, not a project.

List everything your agents can currently read. Put a human name beside each row. Most
rows will come back blank. That blankness is the result, and you didn't have to buy
anything to get it.

Which is my objection to the reflex I keep seeing, where a governance problem gets
answered with a purchase order. A memory product is not a knowledge policy. It's the same
trap as [buying tooling before you have mapped the work](https://www.azgard.tech/learn/stop-buying-ai-tools-before-you-map-the-work):
the purchase feels like the decision, so the decision never gets made. Context engineering
is [hardening into a named hiring category](https://www.techradar.com/pro/why-context-engineering-is-ais-next-hiring-challenge)
now, with budgets and job titles attached, which is broadly good news. It also guarantees
somebody will try to sell you the job title before you have answered the five questions.

#### Keeping an AI knowledge base current without a full-time librarian

Freshness is the one axis nobody has solved yet, including the organisations Redis rates
most mature and the vendors selling governance as a product.

Redis published its
[State of Context Engineering survey](https://redis.io/resources/state-of-context-engineering-2026/)
on 3 August. It found that 97% of respondents believe context matters and 4% have built
for it. Another 39% rate themselves ineffective at scoping memory for agents, meaning they
can't say what should persist, for how long, or for whom. And 21% admit they frequently or
usually run critical decisions on stale or batch-synchronised data, a figure that climbs to
69% among the organisations Redis classes as most mature. Maturity has not yet bought
anybody freshness.

The finding I keep chewing on is a pair: 60% rate themselves ineffective at governing
access to AI context sources, while 58% are confident their context systems are governed
effectively. Redis calls it confidence detaching from capability,
which is polite. The same people are grading themselves as failing and passing on the
same question.

Crivello concedes the point himself, incidentally, while describing his own memory agent's
cache: "every cache at some point got a TTL. It grows stale." That's the crawl-everything
camp acknowledging the exact worry the curation camp is built around.

So stop trying to eliminate staleness and make it visible instead. Two habits do most of
the work. Keep canon small enough that one person can reread the entire thing in an
afternoon each quarter, which in practice means around 15 documents covering the handful
of things your business genuinely makes rulings on. If your approved set runs to 400
documents you don't have canon, you have a wiki wearing a better name.

And make expiry demote rather than delete. When a document blows its review date, it drops
out of canon into the evidence layer, where the agent can still quote it with a date
attached and can no longer state it as policy. Demotion keeps the audit trail intact while
stripping the document of its authority to give orders.

#### What I would write down first

Before you buy a memory product, before you argue about crawl depth, make one page with
three columns: the thing your agents can read, the person who owns it, the date it was
last checked. Add a fourth column if you're feeling brave, for what would kill it.

We run our own brain at Azgard and I've done this exercise on it. It was not a
comfortable afternoon. But it cost nothing, it took less time than a vendor demo, and it
told me which questions our agents were not qualified to answer, which is the only thing
on that page that would have shown up in a customer email otherwise.

**FAQ**

Q: How do I stop an AI agent from using outdated company documents?
A: Stop treating all your documents as equally authoritative. Put a small number of approved documents in a canon layer, each with an owner and a review date, and tell the agent it may only state a rule from that layer. Everything else it can quote with attribution, never as policy. When a canonical document blows its review date, demote it to quotable evidence rather than deleting it.

Q: What should go into a knowledge base for AI agents?
A: Two separate things. A wide, cheap evidence layer holding Slack, email threads, transcripts and old drafts, which the agent may quote but never rule from. And a narrow canon layer, ideally under 20 documents, covering the handful of things your business actually makes rulings on: pricing, refunds, escalation, safety obligations.

Q: Do I need a full-time person to keep an AI knowledge base current?
A: No, but you do need canon small enough that maintaining it is an afternoon each quarter rather than a job. The moment your approved set runs to hundreds of documents, nobody rereads it, review dates rot, and you are back to an undifferentiated pile. Small canon is what makes part-time ownership realistic.

---

### Stop buying AI tools before you map the work.

URL: https://www.azgard.tech/learn/stop-buying-ai-tools-before-you-map-the-work
Published: 2026-06-11
Topic: strategy · Tags: adoption, tooling, process · ~5 min read

> Most AI spend fails before procurement because nobody mapped the work it was meant to change. Here is the mapping sequence we run before any tool decision.

**TL;DR**

- Most failed AI projects were lost before procurement, when a tool was chosen before the work it had to change was mapped.
- Mapping the work means writing down each step of one real workflow, who does it, what enters and exits, and where the time actually goes.
- One mapped workflow with a measured baseline beats a platform licence; the tool decision becomes obvious once the work is visible.

Every month, somewhere in Sydney, a leadership team signs for an AI platform licence
before anyone has written down what the platform is supposed to change. Six months later
the licence renews, adoption sits under 10%, and the conclusion drawn is "AI isn't ready."

The model was never the problem. The sequence was.

#### The buying reflex

AI procurement usually starts with a demo. Demos are built to show off a model, so the
conversation anchors on capability (what the tool *could* do) rather than on your
operation (what must *change*). Capability is abundant. Fit is scarce.

The money involved is not small. A mid-tier AI platform for a 30-person business runs
$40 to $90 a seat each month. Call it $20,000 to $35,000 a year once onboarding and
the inevitable consulting top-up land. Against that, the diagnostic work that would
have told you whether the tool fits costs a half-day of three people's time.

When the tool arrives, it lands on a workflow nobody has described precisely. Staff are
asked to "use AI" in the gaps of their day. There is no baseline, so nobody can say
whether anything improved. The renewal decision gets made on anecdotes.

There is a tell for this pattern. The licence was bought to answer a board question
("are we doing something about AI?") rather than an operations question ("which ten
hours a week are we buying back, and from whom?"). Tools bought to answer board questions
get measured by board metrics, which is to say: not at all.

#### What mapping the work means

Pick one workflow that hurts: quoting, intake, reporting, scheduling. Then write down,
with the people who do it:

1. **Each step**, in order, including the informal ones that live in someone's head.
2. **Who touches it**, and where it waits between people.
3. **What enters and exits** each step: emails, spreadsheets, phone calls, PDFs.
4. **Where the hours go**, measured roughly but honestly.

The map fits on one page. The first time you draw it, two or three steps will surprise
the owner of the process. Those surprises are usually where the value is.

##### A worked example: quoting at a trade business

Here is the shape of it at a signage manufacturer we work with. The workflow that hurt
was quoting. Before mapping, the owner described it in one sentence: "the estimator
reads the drawings and prices the job." The map said otherwise:

- An enquiry lands by email or phone and waits, on average, a day and a half before
  anyone logs it. Nobody owned the inbox.
- The estimator re-types customer details that already exist in the CRM.
- Measurements come off PDF drawings by hand, into a spreadsheet with one tab per
  product type.
- Half the pricing rules live in that spreadsheet. The other half live in the
  estimator's head.
- Won quotes are re-keyed, line by line, into the job-management system.

Rough but honest numbers: about 3.5 hours per quote, around 12 quotes a week. Call it
40 hours: a full role spent producing documents rather than winning work. None of
that was visible in the one-sentence version of the process.

The surprises follow a pattern, too. Across the maps we draw, they are almost always
one of three things: work sitting in a queue nobody owns, the same data entered twice
in two systems, or a decision rule that exists only in one person's head. None of
those shows up in a vendor demo. All of them show up on a one-page map.

#### The tool decision makes itself

With the work mapped, the question stops being "which AI platform?" and becomes "which
step do we change first?" Often the answer needs no new licence at all: a model you
already have access to, wrapped around one step, with its output checked by the person
who used to do that step manually.

"Wrapped around one step" is smaller than people expect. At the signage business it
was a single skill that reads the measurement table out of a PDF drawing and drops the
line items into the estimating spreadsheet. The estimator checks every line before
pricing: the model proposes, the human approves. No platform, no new login, nothing
changes for the customer. Per-quote time came down from about 3.5 hours to a little
over two, and because checking is faster than transcribing, errors fell rather than
rose.

Run that for a month against the baseline you measured. If the hours come back, scale
it to the next step. If they don't, you have spent weeks, not a year, learning that.

The baseline is what makes the result defensible. "Quotes took 3.5 hours and now take
just over two" survives a board meeting. "The team feels faster" does not. It is also
what stops scope creep: every proposed second step has to clear the same bar the first
one did.

#### Where to start this week

Choose the workflow whose owner complains the loudest. Book a half-day with them.
Draw the map, measure the baseline, and only then open the conversation about tools.

A few rules keep the half-day honest:

- Map what actually happens, not what the procedure document says happens.
- Time the waits as well as the work: waiting is usually the bigger number.
- Write down every re-key and every export-import. Each one is a candidate.
- Keep tool names out of the room. The map first, the shopping list later.

If you want a second pair of eyes on the map, that is exactly what our
[discovery call](https://www.azgard.tech/contact) is for.

**FAQ**

Q: How long does mapping a workflow actually take?
A: For a single workflow like quoting or onboarding, a half-day workshop with the people who do the work gets you 80% of the map. The remaining detail surfaces in the first week of the pilot. It is hours, not months.

Q: We already bought a tool. Is it wasted?
A: Usually not. Map the work now and you will either find the workflow where the tool earns its licence, or learn quickly that it never will. Both outcomes beat a renewal decision made on hope.

Q: Who should own the work map?
A: The person who runs the process day to day, not IT and not the vendor. They know where the real exceptions live, and they are the ones whose hours you are trying to win back.

## Contact

- Email: angus@azgard.tech
- Contact form: https://www.azgard.tech/contact — Angus reads every message and replies within one business day
- Enquiries about builds (Audit, Pilot, Custom System, Embed) start with a free 30-minute call, booked from https://www.azgard.tech/work

## Policies

- Privacy policy: https://www.azgard.tech/privacy
- Terms of service: https://www.azgard.tech/terms