How to train AI on your company data, in the order that works
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.
Angus McDonald · 25 Aug 2026 · 10 min read
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 | 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, 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 | 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". 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. 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. 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". 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. 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, 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 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 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
tags: company-brain ● context ● knowledge-base ● sops ● adoption
Angus McDonald
Founder, Azgard
Builds and operates production AI systems for organisations that need results, not slide decks.