Buyer's guidePublic Sector
Buying AI in government: from problem statement to contract management
Buying AI is still buying a service, but three things make it harder: performance depends on your own data, the product keeps changing after award as models are updated, and costs often scale with usage. This guide covers the problem statement and market engagement, the legal frame in the UK, EU and US, evaluation on your own data, the contract terms that keep you in control and the choice of route to market.
On this page
- Writing the problem statement before meeting AI suppliers
- The legal frame for buying AI in the UK, EU and US
- A hypothetical scorecard for a document-processing service
- Clauses an AI contract with a public body should contain
- Choosing a route to market for an AI purchase
- Where government AI purchases go wrong
- Questions and answers
- Sources
Writing the problem statement before meeting AI suppliers
Many failed AI purchases trace back to a specification that named a technology instead of an outcome. A request for an AI chatbot is not a requirement. Answering routine council tax questions accurately from published guidance, in English and Welsh, with a route to a person, is. Describe the current process and where it fails, the decision the AI would touch, the data that exists and who owns it, and how you will know the service has improved. Without that baseline no bid can be scored on outcomes.
Then talk to the market before writing the tender. The Procurement Act 2023 provides for preliminary market engagement, as long as the buyer publishes a notice and does not give any supplier an unfair advantage1. US federal guidance encourages demonstrations and tests of candidate AI systems during market research6. Use these conversations to try approaches on a sample of your own data, learn how suppliers price, and find out what each will and will not disclose about its models. The AI vendor evaluation guide covers the technical due diligence behind those conversations.
The legal frame for buying AI in the UK, EU and US
Procurement law decides how you may run the competition; AI guidance shapes what you ask for. Both apply.
| Question | United Kingdom | EU member states | US federal agencies |
|---|---|---|---|
| Main rules | Procurement Act 2023, in force since 24 February 202512 | Directive 2014/24/EU as transposed into each country's law3 | Federal Acquisition Regulation, with OMB Memorandum M-25-22 for AI6 |
| Procedures that suit novel AI | Competitive flexible procedure, which lets buyers design stages such as demonstrations1 | Competitive dialogue or innovation partnership3 | Performance-based acquisition, with proposed solutions tested during evaluation where practicable6 |
| AI-specific guidance | Guidelines for AI procurement from the Office for AI4 | Model contractual clauses for AI (MCC-AI), in high-risk and light versions5 | M-25-22 requirements on data, IP rights, lock-in and monitoring6 |
| Buying from existing agreements | Frameworks and dynamic markets, such as Crown Commercial Service agreements | Framework agreements and dynamic purchasing systems | Existing contract vehicles; GSA was tasked with planning public guides to AI acquisition approaches6 |
| Use of the buyer's data | Set by contract terms and UK GDPR processor duties | Set by contract terms and GDPR processor duties | Contracts must bar training on non-public agency data without explicit consent6 |
A summary for orientation, not legal advice. Thresholds, exemptions and national transposition details change what applies to a given purchase.
A hypothetical scorecard for a document-processing service
Suppose a benefits team is buying a service that reads supporting evidence and pre-fills case records for staff to confirm. The weightings are illustrative; fix your own before you see any bid.
| Criterion | Weight | A strong bid shows | A weak bid shows |
|---|---|---|---|
| Accuracy on your documents | High | Field-level accuracy on your held-back sample, broken down by document type and scan quality | Accuracy figures from the supplier's own demonstration set |
| Handling uncertainty | High | A confidence signal per field, with low-confidence items routed to staff and nothing saved unconfirmed | Every field filled automatically, with no confidence signal |
| Fairness | High | Error rates compared across handwriting, languages and document types that correlate with protected groups | One overall accuracy figure |
| Explainability for staff | Medium | Each extracted value linked to the place in the document it came from | Values with no link back to their source |
| Security and data location | Medium | Processing locations named, access controls evidenced, accreditation route agreed | Location to be confirmed, or dependent on unnamed subprocessors |
| Whole-life cost and exit | Medium | A price for your stated volumes over the full term, including upgrades, and an export plan in open formats | A unit price only, proprietary formats and no exit assistance |
Illustrative only. Your published award criteria must reflect your own service and the rules of the procedure you use.
Clauses an AI contract with a public body should contain
Model clauses such as the MCC-AI are a useful starting point, but they need tailoring to each purchase5.
Choosing a route to market for an AI purchase
- If
A framework or dynamic market already lists suppliers offering the kind of service you need.
ThenCall off from it, but run your own evaluation on your data and add the AI clauses above where the call-off terms allow.
Framework terms were written for general IT services and rarely cover model changes, training on your data or independent testing.
- If
The problem is new and you cannot specify the solution in advance.
- If
You only need to learn whether an approach works at all.
ThenBuy a short, separately priced proof of concept with your data and success criteria defined, and decide in advance how a full contract would be competed.
Pilots bought without a lawful route to scale end as working prototypes that nobody can deploy.
- If
The AI arrives as a feature inside software you already license, such as a writing assistant in an office suite.
ThenAssess it through change control on the existing contract: data terms, opt-outs and whether it creates a new transparency duty.
Embedded features arrive by update, so the decision is made by default unless someone asks.
Where government AI purchases go wrong
The specification names a technology, not an outcome
Early signalThe tender asks for an AI solution with no baseline, volumes or success measures.
MitigationSpecify outcomes and the decision boundary, and publish how bids will be tested.
Bids are scored on the supplier's data
Early signalScores rest on demonstrations and case studies rather than your own records.
MitigationHold back a representative sample of your data and test shortlisted bids on it, as US federal guidance now expects where practicable6.
Usage-based costs are not modelled
Early signalPrice schedules quote per-request rates without volumes or growth.
MitigationGive every bidder the same standard workload to price and evaluate total cost over the contract term.
There is no exit plan
Early signalData export, ownership of fine-tuned models and transition support are missing from the contract.
MitigationAgree exit deliverables at award and rehearse an export before go-live. Keep transparency records current as the contract runs; see algorithmic transparency.
Questions and answers
Can we buy AI services through frameworks such as G-Cloud?
Often, yes. Many AI products and cloud services are listed on Crown Commercial Service agreements, and calling off from one can be quicker than running a full procedure. Check that the supplier's listing covers what you need, run your own evaluation within the call-off rules, and add AI-specific terms where the framework allows. For novel development work, a framework built for commodity cloud services may not fit.
How do we compare usage-based pricing between AI suppliers?
Give every bidder the same standard workload: expected monthly requests, typical input and output length, peak periods and likely growth. Ask for a total price over the contract term that includes storage, support and model upgrades, and for caps or alerts on spending. Then test how each price behaves if volumes turn out higher than planned, because usage-based charges diverge quickly.
Who owns a model fine-tuned on government data?
Whoever the contract says, which is why it must say so. Default supplier terms often keep the weights with the supplier. Public buyers usually seek ownership of their data and outputs, plus either ownership of fine-tuned weights or a perpetual licence with the right to export them, and a bar on reuse for other customers. US federal guidance requires agencies to settle data ownership and IP rights in AI contracts, with particular care where agency data is used to train or fine-tune a system.
What should happen when a supplier changes the underlying model?
The contract should require advance notice, a chance to test the new version against your evaluation set, and the option to stay on the previous version or roll back if results fall short. Treat a model change as a change to the service: re-run acceptance tests, update impact assessments and transparency records if the behaviour shifts, and record the decision.
Sources
- Procurement Act 2023 — legislation.gov.uk · checked 10 October 2026
- The Procurement Act 2023: a short guide for senior leaders — GOV.UK · checked 10 October 2026
- Directive 2014/24/EU on public procurement — EUR-Lex · checked 10 October 2026
- Guidelines for AI procurement — Office for Artificial Intelligence, GOV.UK · checked 10 October 2026
- Updated EU AI model contractual clauses (MCC-AI) — Public Buyers Community, European Commission · checked 10 October 2026
- M-25-22: Driving Efficient Acquisition of Artificial Intelligence in Government — Office of Management and Budget · checked 10 October 2026