Notes on working with AI

What gets written here

Posts come out of questions that arrive more than once, usually from someone already partway through a rollout and finding that the hard part was not the technology. They are written for the person who has to make the decision and then live with it, rather than for a feed. There is deliberately no news coverage here. A model released this week will have been superseded by the time it could matter to a policy you are still drafting, and there are faster places to read about it. What changes slowly is what is worth writing down: how a decision ages, where a rollout loses momentum, and which questions to ask a vendor before the documents are already in their system.

The recurring threads

Governance that survives contact with staff

What to decide before the first licence is issued, what can safely wait until later, and which of those decisions turn out to be expensive to reverse once people are already working in the tool every day.

Security and where documents end up

What an enterprise AI vendor has to prove before it touches your files, how retrieval over your own documents actually behaves once you point it at a real archive, and what unsanctioned use is already telling you.

Capability, not licences

Why usage plateaus in week three, what separates training that sticks from a one-off workshop, and how to read the difference when several models answer the same prompt in different ways.