Ten people, ten ChatGPTs, and no shared rules

A client email arrives that sounds like your firm wrote it. Another arrives the same day, from the same firm, that does not. Different greeting, different terms, a price that was retired in March. Both were drafted with AI. Two staff members, two personal accounts, two sets of instructions that nobody else has seen.

This comes up in most of our first conversations with owners, usually phrased as a question. Here are the ones we hear most.

"Is my team even using AI?"

Almost certainly, whether or not you bought it. The University of Melbourne and KPMG surveyed more than 48,000 people across 47 countries for their 2025 study on trust in AI. Among employees who use it, 57% said they hide their use and present AI work as their own. Only about 40% said their workplace has clear guidance on generative AI.

So the honest starting assumption is that AI is already in your workflows and you cannot see most of it.

"Why would they hide it?"

People worry it looks lazy, or like cheating, or that it will be used to justify cutting their role. Those worries are about culture and trust, and no tool fixes them. What owners can control is whether using AI properly is easy and expected, or something people do quietly on their phones.

"What is the actual harm if everyone uses it their own way?"

Three things, in our experience.

The firm stops sounding like one firm. Each person's assistant has its own idea of your tone, your terms and your prices, built from whatever they typed into it.

Facts drift. One person's chat history says the service fee changed last quarter. Another's does not. Clients notice before you do.

Corrections do not travel. When someone catches the AI making a mistake and fixes it in their own chat, the fix stays in that chat. Everyone else's assistant keeps making the same mistake.

None of this is visible from the owner's desk. It tends to surface as a confused client or an awkward phone call, long after the drafts went out.

Writer's 2025 survey of 1,600 executives and employees found that 42% of C-suite respondents said adopting generative AI was tearing their company apart. That is a vendor survey of large firms, so treat it with some care. But the fragmentation it describes shows up in ten-person firms too.

"Should we ban personal AI tools?"

You can, and in regulated work you may have to. A ban on its own tends to push use further out of sight. The firms that handle this well give people an approved way to use AI and make it better than the private one.

"What makes the approved way better?"

The company's own context. A shared memory that every person's AI session reads first: the firm's rules, current prices and terms, approved wording, the voice you write in, and the corrections people have already made. When one person fixes a mistake, the fix is written into that memory with the reason, and every later session picks it up.

That is how we run AI inside Provecta Group. Our rules and corrections live in one written memory the company owns, and every session loads it before doing anything. Ten people using it get one firm's answers, not ten.

"Does that solve shadow AI?"

It solves part of it. Shared memory standardises what goes into the AI. It does not make people comfortable admitting they use it, and it does not train anyone. Memory is the artefact. Getting people to use it, and to say when it gets something wrong, is management work, and it takes longer than setting up the files.

It also does nothing about data people paste into personal accounts. That is a security question with its own answers, and we will come back to it.

"Where would we start?"

Ask your team, without blame, which AI tools they use and for what. Collect the three or four tasks that come up most. For each one, write down the rules a new hire would need: the facts, the wording, the things never to promise. That page is the first entry in your company memory.

The AI readiness check gives you a quick read on where your team stands. Book 20 minutes at pgco.world/book to go through the results.