Buying AI in the public sector: getting value under the Procurement Act

Public-sector AI spending has grown quickly over the past year, driven by pressure to do more with less and a clear government push to modernise services. The risk is familiar: buying technology faster than the organisation can absorb it, on terms that make it hard to measure value or change direction later.

The Procurement Act 2023, in force since February 2025, gives public bodies more flexibility than the old regime. Used deliberately, it is a strong tool for buying AI well. Used by default, it can repeat the mistakes of the last generation of IT contracts.

What the Act changes for AI buying

Several features of the Act matter directly for AI:

• Most advantageous tender. Contracts are awarded on the most advantageous tender rather than the old "most economically advantageous" test, giving more room to weigh quality, social value and risk alongside price.
• The competitive flexible procedure. Buyers can design a procedure that fits the market — including stages for demonstrations, testing and negotiation. For AI, where capability claims need to be tested on real data, this is valuable.
• Transparency across the lifecycle. Notices are published on a central digital platform from planning through to contract change and termination.
• Published performance. For larger contracts, key performance indicators must be set and performance against them published.
• Supplier exclusion and debarment. Clearer grounds and a central debarment list strengthen the position of buyers when suppliers fail.

Buy outcomes, not tools

The most common weakness in AI procurement is specifying a product rather than a result. A tender that asks for "an AI tool for case handling" invites suppliers to sell features. A tender that asks for "a 20% reduction in average case handling time, with no fall in quality scores, measured against a published baseline" invites them to commit to value. The target in that example is illustrative; the principle is not.

Before going to market, agree:

• the service outcome and the baseline it will be measured against;
• the data the solution will need and who owns it;
• the human oversight required and who provides it;
• how performance will be reported — which, for larger contracts, the Act now requires you to publish.

Test before you commit

AI capability is easy to claim and hard to verify on paper. Use the flexibility of the competitive flexible procedure to build in proof:

• a structured demonstration on representative, anonymised data;
• a time-boxed proof of value with clear pass and fail criteria;
• independent review of the supplier's approach to accuracy, bias and explainability.

This takes slightly longer up front. It saves far more time than recovering a contract that does not deliver.

Design out lock-in from day one

AI contracts can create lock-in in new ways: proprietary models trained on your data, workflows built inside a supplier's platform, and pricing that rises with usage. Protect your position by:

• securing rights to your data, outputs and any fine-tuned models;
• requiring documented exit and migration support;
• separating the platform from the services where practical;
• setting usage-based pricing caps and review points.

Govern AI as a service, not a project

Once the contract is signed, the work shifts to assurance. Treat each AI deployment as a live service with an owner, published performance measures, a risk register and a regular review against the outcomes in the contract. The Act's transparency requirements make performance visible to the public; your governance needs to make it visible to your board first.

A practical checklist

Before any AI procurement goes to market, a senior responsible owner should be able to answer yes to five questions:

• Is the outcome defined and baselined?
• Do we know what data is needed and do we have the right to use it?
• Have we designed a way to test claims before award?
• Are exit, data rights and pricing protections in the terms?
• Is there a named owner for performance after go-live?

Public bodies have a real opportunity to improve services with AI. The Procurement Act gives them the tools to do it on terms that protect value. The discipline is in using those tools on purpose.