Buying government AI means buying an exit route and evidence

US federal memoranda issued in April 2025 link adoption with governance and procurement. Contract questions should follow the public-service task.

A government buyer can procure a model that appears impressive and still acquire a service that is difficult to audit, expensive to change or unsuitable for citizens. The challenge is to turn an AI demonstration into a contractual description of work. That description needs an acceptance method, a clear responsibility for errors and a credible way to continue the service if the supplier changes direction.

In April 2025 the US Office of Management and Budget issued [M-25-21](https://www.whitehouse.gov/wp-content/uploads/2025/02/M-25-21-Accelerating-Federal-Use-of-AI-through-Innovation-Governance-and-Public-Trust.pdf) on federal AI use and [M-25-22](https://www.whitehouse.gov/wp-content/uploads/2025/02/M-25-22-Driving-Efficient-Acquisition-of-Artificial-Intelligence-in-Government.pdf) on acquisition. M-25-21 replaced M-24-10. The [official announcement](https://www.whitehouse.gov/releases/2025/04/white-house-releases-new-policies-on-federal-agency-ai-use-and-procurement/) presented the changes as an effort to accelerate adoption while retaining public protections. This is analysis of that dated policy change, not a certification of a procurement or a claim that no subsequent guidance exists.

The practical distinction is between purchasing access to a technology and accepting a public-service workflow. An original illustrative agency wants assistance answering routine grant-program questions. The contract can specify an approved document collection, the date on which those documents were valid and a requirement to show the source behind an answer. It can also specify that an answer about eligibility remains informational until the responsible officer makes a determination.

Acceptance should include unanswered questions. A test set can contain an outdated deadline, an applicant whose circumstances are missing and a request involving a different agency’s program. The correct result may be a refusal to guess and a clear route to the relevant office. If the evaluation rewards only completed answers, the supplier has an incentive to make the system sound certain even when its evidence is weak.

A useful contract also distinguishes defects. An unavailable model, an incorrect document and a misunderstood question need different remedies. In the illustrative help service, a temporary outage can show an ordinary support option. A potentially incorrect answer should trigger review of the cited material and notice to affected users where appropriate. The buyer should retain enough redacted operational evidence to investigate without storing unnecessary applicant information.

Dependency enters through several layers. The application supplier may call a separate model provider and use another service for search. A promise to export conversations does not necessarily provide the source documents, prompts, evaluation cases or configuration needed to reproduce the workflow elsewhere. Buyers can request a concrete handover package and test it before renewal. This is an editorial procurement example, not an assertion that the memoranda prescribe that exact package.

Pricing needs the same specificity. A low unit price can hide repeated calls, large document inputs, storage, integrations and staff review. Run a representative workload and calculate the cost of accepted answers, including escalations. Ask how model or price changes are notified and what happens at a quota limit. A public-service queue should not silently stop because a demonstration account reached its allowance.

Security and rights review must follow the data. Grant questions can reveal financial or personal circumstances even when they arrive as free text. Decide which fields the service needs, which may be removed and who can access retained records. Test the supplier’s claimed controls with documentation and the contract rather than assuming that an enterprise product label settles those questions.

There is a workforce issue too. Officers need to recognise unsupported answers, correct the knowledge collection and suspend a problematic feature. Someone must own the service after the procurement team finishes. In a pilot, include staff review time in the business case and reserve capacity for incident response. Otherwise savings on paper can become unpaid work in another department.

The April memoranda are a documented policy baseline for understanding federal buying choices. Their operational value depends on the procurement that follows. A buyer who starts with a bounded public task can ask better questions about evidence, responsibility and portability. The result worth paying for is a service that stays useful and accountable through ordinary change, rather than an impressive demonstration that only its original supplier can maintain.