Skip to content

Company data in public AI tools: where the line is

Your team already uses ChatGPT — the question isn't whether, it's with what data. A three-category rule anyone can remember, without blocking the work.

Dan Cristian Alexandrescu4 min read

Your team is almost certainly already using ChatGPT or its equivalent, whatever the company policy says. The useful question isn’t “who uses it”, because the answer is “everyone, quietly”. The useful question is with what data.

Most companies react in one of two ways, both bad. Either they ban everything — and people use AI from their personal phone, where you see nothing. Or they say nothing — and a signed contract ends up in a public chat because someone wanted a summary.

The rule below has three categories, because four don’t get remembered.

The short version, for management
04
  • A blanket ban doesn’t work: it moves AI use onto personal devices, where you have no visibility at all.
  • Three data categories — green, amber, red — are enough. Anyone in the company should be able to classify a document in five seconds.
  • What actually makes the difference is which account people use, not which model. A company account with a contract changes the whole conversation.
  • The rule fits on one page and gets tested on three real documents from your company, not on textbook examples.
— The rule

Three categories, not a list of rules

03
  • Green

    Goes anywhere

    Information that is already public, or could be public without bothering you: site copy, product descriptions, press releases, marketing material, general professional questions. You don't want friction here — you want people using AI as much as possible.

  • Amber

    Company account only

    Internal data that isn't secret but isn't public either: processes, template quotes, training material, aggregated sales figures, ordinary correspondence. It can go into an AI, but only through the organisation's account with its contract — not from a personal address.

  • Red

    Doesn't go up anywhere

    Identifiable personal data of customers or employees, signed contracts, credentials, anything under NDA, unpublished financials, anything tied to a live case. If someone genuinely needs AI on that data, it gets solved with a hosted setup, not a public chat.

— What actually matters

The account, not the model

How the team uses AIWhat control you haveVerdict
Free personal account, on a phoneNone. You don't know what went up and you can't prove anything.Green only
Paid personal accountIndividual settings you don't administer.Green only
Organisation account, with a contractCentral administration, contractual terms, control over data retention.Green + amber
Model hosted in your own infrastructureComplete — the data never leaves your perimeter.Including red, with controlled access

Vendors' default settings on retention and training change often. That's why the rule is tied to account type and contract, not to a particular toggle in a menu.

The link to 2 August. The data rule and the AI Act’s transparency obligations are two different things, but they get written into the same document and explained in the same meeting. If you’re gathering the team anyway to settle what data goes where, use the occasion for the chatbot disclosure and content marking too.
— To do

Two decisions for this week

  • Decide whether you pay for company accounts

    It's the cheapest security measure available to you this quarter. The per-person cost is below an hour of a lawyer's time, and the effect is that work moves off personal accounts, where you see nothing, onto accounts you administer.

  • Test the rule on three real documents

    Take three documents circulating in the company right now — a quote, a contract, a report — and ask three different people to classify them green / amber / red. If they don't agree, the rule isn't clear enough and needs rewriting before you publish it.

— FAQ

Questions we got on this edition

03
  • Wouldn't a blanket ban be simpler?

    Simpler to write, yes. Safer, no. A blanket ban doesn't stop AI use, it relocates it to personal devices, where you have no visibility, no contract and no way to prove anything. The companies that tried found this out about six months in, usually through an incident.

  • How do we know whether a tool trains on our data?

    It's in the terms of use and the business contract, and it differs between the free plan and the organisation plan. Because those terms change, our recommendation is to tie the internal rule to the account type — which you control — rather than to a setting the vendor can change between quarters.

  • Where does customer personal data fit?

    Red, by default. Processing it through an external vendor raises GDPR questions — legal basis, processing agreement, transfers outside the EU — that get resolved before, not after. If you have a genuine business case for it, it's worth building properly, with a hosted setup and your DPO in the conversation from the start.

— AI Briefing

The one-page rule, written for your company.

We write it with your team, test it on real documents and fold it into the internal AI policy — alongside the register of systems and the transparency obligations.