Leaders who won't give AI "too much", and whether an enterprise licence already protects you

Both sides are half right. The enterprise licence stops the provider learning from your data. It does nothing about what a tool can reach inside your company, and that's usually what the cautious leader is really worried about, even if they can't quite say so.

By Matthew Steele4 min read28 October 2026

A boardroom with two directors in conversation across the corner of a table, one with arms folded, a closed laptop between them
Illustrative scene, not a customer's premises.

What the licence does and doesn't cover

The enterprise licence covers It doesn't cover
Your prompts and files aren't used to train the provider's models. What the tool can reach inside your business.
Company-owned accounts and workspaces, so a leaver's account goes with them. Whether the shared drive is tidy enough to be safe to read.
A contract with retention, deletion and sub-processor terms. What an agent may write, send or change.
Administration: who has a seat, usage, some controls. Whether anyone is checking the output.

The left column is real and worth paying for. The right column is where the actual risk sits, and it's entirely in your hands – the rings in what data AI should see are how we'd carve it up.

What the cautious leader is actually worried about

In our experience it's one of three things, and it's worth finding out which:

  • It'll see things it shouldn't. The answer is the rings: it sees the internal ring and not the restricted one. The leader's right that the restricted ring should stay closed, and they should hear that from you.
  • It'll put them somewhere. The answer is the licence and the settings: training off, company workspace, retention set. This is the one worry the licence does solve.
  • It'll decide something. The answer is the three tiers: anything that acts on its own gets a review, an owner and a log. An assistant that only drafts isn't the concern.

A leader who says "too much" is rarely wrong about the restricted ring. They're usually wrong about the internal ring, which is where the useful work is and where the risk is low. Separating the two is the whole conversation, really.

IT's worry about the folders

IT is right, and for a reason that predates AI. An assistant connected to the shared drive can see whatever the person using it can see, and in most businesses that's far more than it should be, because drive permissions have been drifting for a decade. The assistant doesn't create the exposure. It makes it obvious – a person who'd never have opened the finance folder now gets an answer drawn from it.

What the licence protects and what it leaves to youTwo columns. The licence covers: no training on your data, company-owned accounts, a contract with retention and deletion terms, administration of seats. You still decide: what the tool can reach inside your business, whether the shared drive is tidy enough to be safe, what an agent may write or send, and whether anyone checks the output. The folder problem sits in the right-hand column.The enterprise licence coversYour prompts and files are not used to trainthe provider's models.Company-owned accounts; a leaver's goes with them.A contract with retention, deletion andsub-processor terms.Administration: seats, usage, some controls.You still decideWhat the tool can reach inside your business.Whether the shared drive is tidy enough to besafe to read. (IT is right about this one.)What an agent may write, send or change.Whether anyone is checking the output.The left column is real and worth paying for. The risk the cautious leader senses lives in the right column, and it is yours.
What the licence protects and what it leaves to you. The folder problem sits on the right-hand side.

The fix is the one IT has wanted for years anyway: tidy the permissions so the finance folder is visible to finance and the HR folder to HR. Do that before you connect any assistant to the drive, and connect it to a few named folders first rather than the lot. IT's worry becomes the project plan.

The conversation that works

Not "how much should AI see?". That one has no answer and ends in a stalemate. Instead:

  1. Draw the four rings together, in an hour.
  2. Agree that the restricted ring stays closed to tools. The leader hears their caution confirmed.
  3. Pick one job in the internal ring, put a number on it, and run it for a quarter.
  4. Review the number, then widen one ring or don't.

The date is what matters. Caution with a review date on it is a plan you can work to. Without one it's a veto, and nothing moves.