The full argument for why enforcement is the half that counts sits on the AI governance page. This is the shorter version, and what the work looks like.
How to tell whether you have this
Ask around the firm what AI tools people use for work. Not whether they should. What they actually do.
You will not get a complete answer, and the incompleteness is the finding. Most AI tooling in a small firm arrives through individuals, not through procurement, so there is no purchase record, no contract and nothing to audit against.
Why it happens
Someone summarised a client document in a chatbot, probably in the last fortnight, almost certainly for a sensible reason and with no intent to do anything wrong. That is the ordinary case. It is also why this is difficult to address by telling people off.
The tools also arrived faster than any policy cycle could move.
How I would approach it
- The register first. What is in use, by whom, against what kind of data. Everything downstream depends on it, it is the quickest of the three, and it is routinely the one that surprises people.
- The boundary second. Classification of the data that matters, and a written decision about which categories may reach which tools. Specific enough to follow. A template that says be careful is not.
- The enforcement third, and this is the part that changes your answer. The same position implemented as policy in Microsoft Purview, so it acts on the data itself and does not depend on everyone remembering what the document said.
Only the second of those is a writing exercise. A policy in a document holds if everybody remembers it. A policy at the data layer holds whether they do or not, and produces a record either way.