What To Tell AI About Your Business
Open the last serious question you asked AI about your business. Now tell me what it knew.
·
Ninety minutes once, then quarterly
THE SYMPTOM
Open the last serious question you asked AI about your business. Now tell me what it knew about your business when it answered.
Not what you meant it to know. What was actually in front of it.
The other tells. You re-explain the business at the start of every conversation, slightly differently each time, and you have never checked whether the differences matter. The answers are always sensible and never specific: they would fit a firm twice your size, or half, or in another country. It has never once told you an idea was bad for a reason particular to you. And somewhere along the way you quietly stopped asking it the questions that mattered most, and kept using it for drafting, because the important answers felt thin and you could not say why.
The clean test: ask it the same significant question in a fresh conversation tomorrow. You will get a different answer, and nothing in either one will help you decide which was better.
THE CAUSE
The model works with the information available to it. Where facts about your business are missing, it falls back on general patterns and assumptions. Advice that fits businesses like yours may not fit yours. It can end up answering a more generic question than the one you thought you asked. That is not the same as the model making things up, and a better model would not supply what is missing.
You correct for this by supplying context in conversation, often slightly differently each time. You include what feels relevant. But the facts that would change the answer are often the ones you did not think to mention.
A good adviser stops you and says: before I answer that, what is your cost to serve? The model can construct a coherent answer without knowing. Sometimes it asks. Sometimes it carries on without checking whether the missing information would change its advice.
That makes the failure easy to miss. An answer built on too little can sound just as convincing as one grounded in your actual numbers. Fluency alone won't tell you which is which.
A better question only takes you so far. Your margin, your cash, your capacity and what you are trying to do this year can each change which advice is right. Rewording the question won't supply those facts. You have to provide them.
THE MOVE
1. Write a context pack. One document, one page, kept current. Not a folder, not a wiki, not a knowledge base. One thing short enough that pasting it costs you nothing, because a context pack you do not paste is a filing exercise.
2. Start with shape, in two paragraphs. What you sell, to whom, how you make money, revenue, headcount, how long you have been going. Plain sentences.
3. Then the numbers that change answers. This is the part that does the work. Gross margin. Cost to serve, if you have done Fix 03. Cash position and runway. Debtor days. What proportion of revenue comes from your largest three customers. Utilisation, if you sell time. Six or seven figures, each with a date attached.
If you cannot fill this section, that is not a failure of the pack. It is a list of which other Fixes you need.
4. Then the constraints, which are the things you cannot do. The cash ceiling above which nothing is possible. The person everything routes through. Contractual restrictions, regulatory limits, geography, the customer you cannot afford to annoy. Advice that ignores your constraints is not advice, it is a wish, and the model has no way of knowing your constraints exist.
5. Then the decisions already taken, with reasons. We do not do fixed price work, because we tried it in 2024 and here is what happened. Without this you will be sold the same idea repeatedly, plausibly, forever, and each time it will sound fresh.
6. Then what you are actually bad at. Honestly, in writing. This is the section everyone skips and it changes more answers than any other. A model told you are strong at delivery and weak at sales will give you a materially different plan from one that assumes even competence, which is what it assumes by default.
7. Then what good looks like. Your real objective, which is frequently not growth. More profit on the same revenue. Less of your own time. A business someone would buy in four years. Say it plainly, because the default assumption is growth and the default advice follows.
8. Leave four things out. Personal data about named individuals, including salaries and anything about their health or performance. Anything under an NDA. Anything you would mind seeing outside your business, because you should treat what you paste as though it might be. And your own marketing copy: feeding a positioning statement into a model gets you advice built on your own spin, reflected back with a straight face.
9. Paste it at the start of anything that matters, and update it quarterly or whenever a number moves materially. Put the date on it. An out-of-date context pack is worse than none, because it is confidently wrong rather than obviously absent.
10. Then check every answer with three questions. Which number in what I gave you drove this? What would change it? And what did you assume that I did not tell you? The third is the one that earns its place: it makes the model state its own gaps, which is the closest you will get to it asking you a question.
THE ARTEFACT
The context pack template itself. One page, seven sections, with prompts under each so you are answering questions rather than facing a blank document.
It is a document rather than a spreadsheet, and deliberately so. Some exposures close with a model. This one closes with a page you keep current and actually use.
Two things in the template are worth the attention. Every number has a date field next to it, because a margin from eighteen months ago is not a fact about your business, it is a fact about a business you used to run. And there is a short section at the end, headed What I would rather you did not assume, which exists because the assumptions a model makes about a firm your size are predictable: that you have a sales function, that finance is somebody’s job, that growth is the goal, that you can hire when you decide to.
Fill it in once properly and it takes about ninety minutes. Most of that is spent finding numbers, which tells you something in itself.
Download the document
THE CHECK
By the end of week one: the pack exists, it is one page, it has a date on it, and it contains at least five numbers. Adjectives do not count. If your pack says we are premium and service-led, you have written marketing, and the model will reason from it exactly as if it were true.
By the end of month one: you have used it on three real questions, and on at least one of them the answer was materially different from what you would have got without it. If the answers never change, the pack is too vague, and the fastest test is to delete every sentence with no number or constraint in it and see what survives.
By the end of the quarter: it has told you an idea was a bad one, for a reason specific to your business, that you had not already thought of. That is the marker. A model with no context agrees with you, because it has nothing to disagree with you from.
THE AI REVIEW
Every Fix ends with an AI Review, and it is worth running one here first, on the pack itself. What is specific here. The decisive point is whether the pack survives a cold reading: open a fresh session, hand it over, and find what the model still cannot answer about your business. The assumption to stress is that what you left out did not matter. And it will fill your gaps with plausible averages rather than telling you it does not know, which is the whole reason the pack exists. The method is on The AI Review page.
PUBLISHED
LAST REVISED
The AI Review is now taught in full on its own page rather than described here. A reference to Cost to Serve now gives its correct number. Nothing in the move, the artefact or the Check has changed.