Your AI has never met your company

A customer writes in, annoyed that a delivery is three days late. Someone on the team pastes the email into ChatGPT and asks for a reply. What comes back is polite, well organised and confident. It offers a full refund, promises dispatch within 48 hours and closes with a line about valuing the relationship.

The company does not offer full refunds on late deliveries. It credits the next order. The warehouse cannot promise 48 hours this month. And the customer is one of the accounts the owner handles personally, which everyone on staff knows and the model had no way to know.

Nothing in that draft was stupid. A smart new hire on day one, with no handbook and no one to ask, would write something similar. That is the position most AI tools are in every time someone opens a chat window at work.

What the model actually knows

A general-purpose model has read a large part of the public internet. It has never read your price list, your refund terms, the three customers who get special payment terms, the supplier you stopped using in March, or the way your team signs off emails. When it lacks a fact, it fills the gap with the most likely version of that fact for a company like yours. Most of the time the likely version is close. Close is the problem, because close reads as correct.

Larger companies feel this too. Atlassian's 2025 developer experience report found that developers saved real hours with AI and then lost a similar amount to friction elsewhere, with finding information named as the top time-waster. In June 2025 Bloomberg reported that staff at some companies paying for Microsoft Copilot, which is wired into company files, still preferred ChatGPT. Having the data connected did not settle it. People went with the tool they liked using, and the company context did not follow them there.

Two things follow from that. The AI needs your company's context, and the context has to travel with whichever tool your people actually open.

A written company memory

The fix we use is plain. Write down how the company works in a small set of pages, keep them current, and have the AI read them before it does any work.

Those pages hold a few kinds of things:

• Facts with a source and a date. Prices, terms, service areas, lead times, who owns which account.
• Rules. What we never promise in writing. Who approves refunds. Which customers go to the owner.
• Voice. How we greet people, how we sign off, words we avoid.
• Decisions, with the reason. "We stopped offering same-day delivery in May because the courier contract changed."

None of this is new. It is the handbook most firms mean to write and never do. The difference now is that a model will read every page, every time, without getting bored, which no new hire has ever done.

How we run it ourselves

Provecta Group runs its own work this way. Every AI working session starts by reading the same company context file and the same written standards before it touches a task. It ends by writing down what was decided and what changed, so the next session starts from there. When a session gets something wrong because a fact was missing, the fix is to add the fact to the shared pages, not to correct one chat and move on.

The part people skip is the second half. A memory that only gets read goes stale within weeks. It has to get written to as well.

What a company memory will not do

It will not make the model know anything you have not written down. If the refund policy changed last week and nobody updated the page, the AI will quote last week's policy with the same confidence it quoted the right one.

It will not replace judgement on the angry customer. Whether to call them, credit them, or let the owner handle it is still a person's call. The memory can tell the AI that this account goes to the owner. It cannot decide what the owner should say.

And it is work. Somebody has to own each page and check it. We would rather say that plainly than pretend a connector does it for you.

One thing to try this week

Pull up the last ten AI drafts your team had to fix before sending. For each one, write a single line naming the fact the model did not have. "Did not know we credit instead of refund." "Did not know the Kisumu route runs Tuesdays only." "Used a price from last year."

Those ten lines are the first page of your company memory. Put them somewhere shared, paste them at the top of the next prompt, and watch how many of the same corrections stop showing up.

Find out where your firm stands at pgco.world/tools/ai-readiness.