AI doesn't know your business until you write it down

AI tools produce generic output for one reason: they know nothing about your business beyond what you give them. A language model arrives knowing the averaged text of the public internet — not your buyer, your offer, your pricing, or why clients choose you over the firm down the road. Until that context exists in writing, every AI tool you buy will return plausible filler.

Why does everything AI writes sound the same?

Because the model is doing exactly what it was built to do. A large language model predicts the most likely next words given what it has already seen. When your prompt says "write a cold email for my marketing agency", the most likely output is the average cold email for the average marketing agency — which is precisely the email your prospect has already deleted forty times this quarter.

This is not a flaw you can escape with cleverer phrasing. It is a context problem. I make the broader case in AI Automation for B2B: what actually works: the firms getting real value from AI are not the ones with the best tools, they are the ones with the best-documented businesses. The tool is a commodity. The context is not.

What counts as "writing it down"?

Not a brand book. A handful of short, dense documents that answer the questions a competent new hire would ask in week one:

• Who you sell to. Sector, size band, job titles, and the trigger events that make them buy. One page.
• What you sell. The offer, the price, the delivery mechanism, and what is explicitly out of scope.
• Why you win. The three reasons clients pick you, in their words rather than yours.
• How you sound. Sentence length, words you never use, how formal a reply should be.
• How work gets done. The actual steps of your core processes — the same documentation that decides when a process is ready to be promoted into an automation.

Most 5–50 person firms have none of this written anywhere. It lives in the founder's head, which is exactly where no AI tool can reach.

How does a written document change what the model does?

The mechanism is straightforward. When the context you supply contains your ICP definition, then the model's research and drafts anchor to that definition — it compares each prospect against your stated criteria instead of guessing. When the context contains your tone rules, then the output starts sounding like you rather than like LinkedIn. When the context is empty, then the model fills every gap with the statistical average of its training data — and the average is generic by definition.

Garbage in, garbage out has a quieter sibling: nothing in, average out.

Where does this show up first in practice?

Outbound, in my experience. We run campaigns at 25–40 cold emails a day per inbox, with a roughly 4% positive-reply rate as the working expectation — below 3% means fix the campaign. The gap between those two numbers is rarely the AI model used for research or drafting. It is whether the campaign was fed a precise, written ICP and offer, or a vague verbal brief.

The parallel with data hygiene is exact. You would not send to an unchecked list — catch-alls and bounces are why you verify before you send. Unwritten context is the copy-side version of an unverified list: an input nobody has checked, feeding a system that cannot check it for you.

What should you actually do this week?

Block two hours and write the ICP page and the offer page. Rough is fine; a wrong sentence you can correct beats a right idea nobody can access. Then use them — paste them into every AI-assisted task for a fortnight and note where the documents proved wrong or thin. The documents improve, and the output improves with them.

There is a second payoff. A written business is easier to automate at all: the documentation that makes AI output specific is the same documentation that tells you where no-code tools will carry a process and where a real build begins. It is easier to hand to a human, too. The exercise was never really about AI. AI just made the cost of an undocumented business impossible to ignore.

---

Next step: the Growth System Audit — £450, seven days, credited against any build — maps where your growth system leaks and what to build first.