Governments are buying AI with contracts written for stationery.

Public procurement is designed around a specification, a competitive process, an acceptance test and a warranty. Each of those steps assumes the delivered thing stays the same.

A system built on a hosted model does not. The vendor updates it, behaviour shifts, and the acceptance evidence becomes historical.

What is going wrong in practice

Contracts that specify accuracy against a fixed test set, with no obligation to maintain it. No requirement to notify the buyer of a model change. No right to rerun acceptance tests after an update. And no defined remedy when performance drifts within contract term.

The clauses that fix most of it

Notification of material model changes with a defined notice period. A buyer right to rerun an agreed evaluation suite at any point, with the suite held by the buyer. Performance floors expressed as ongoing obligations rather than delivery conditions. And exit terms that include data and configuration export in a usable form.

Why it is not happening faster

Because procurement teams are not model engineers and the vendors drafting responses have no incentive to propose these clauses. Several national digital agencies now publish model contract terms, which is the cheapest available intervention and the one least used.

The advice one procurement lead gave

Buy the evaluation, not the model. If you control the test and the right to run it, the rest of the contract becomes negotiable.