Lighter rules, heavier contracts: OMB’s April 2025 AI memos handed governance to the buyer

The week

Government rules about technology rarely make headlines, but they decide how billions of dollars are spent and how millions of people are treated by public services. In the first week of April 2025, the White House budget office rewrote the rules for how federal agencies use and buy artificial intelligence.

On April 3, 2025, the Office of Management and Budget, the part of the White House that oversees how agencies manage money and operations, issued two memoranda on artificial intelligence. OMB memos are instructions to federal agencies. They do not pass through Congress, but agencies are expected to follow them.

The timing had been set in January. Executive Order 14179, signed on January 23, 2025, revoked the previous administration’s 2023 AI order and directed OMB to revise the two 2024 memos that implemented it. The April memos were those revisions. They kept the statutory scaffolding that Congress had built, including the AI in Government Act of 2020 and the Advancing American AI Act of 2022. They changed the tone, the categories and the deadlines.

The first, M-25-21, covered how agencies use AI. It replaced the 2024 memo M-24-10. It kept two features of the earlier approach: chief AI officers, senior officials in each agency responsible for AI, and public use case inventories, published lists of where each agency uses AI. It merged the earlier categories of “rights-impacting” and “safety-impacting” AI into a single category of high-impact AI, meaning uses that could significantly affect people’s rights, safety or access to services. Agencies got one year to document minimum risk practices for those uses.

The memo defined high-impact AI as AI whose output serves as the primary basis for decisions with legal, material, binding or similarly significant effects. Its examples ranged from benefits adjudication and federal hiring to biometric identification and the control of critical infrastructure. The minimum practices for those uses were concrete. They included pre-deployment testing and a documented AI impact assessment reviewed by someone independent of the project. They also required ongoing monitoring, operator training, and human oversight with a fail-safe where practicable. Affected people had to have a route to timely human review and appeal.

M-25-21 also set a sequence of deadlines. Each agency had 60 days to retain or designate a chief AI officer. The 24 large agencies covered by the Chief Financial Officers Act had 90 days to convene an AI governance board and 180 days to publish an AI strategy. Every agency had 180 days to publish a compliance plan, repeated every two years until 2036, and was expected within 270 days to adopt a policy on generative AI. A chief AI officer could waive a minimum practice for a specific use, in writing, but had to report each waiver to OMB within 30 days and publish a summary. A high-impact use that could not meet the practices had to be stopped.

The second, M-25-22, covered how agencies buy AI. It replaced M-24-18. It told agencies to stop vendors from training commercial models on non-public agency data without consent. Training is the process by which an AI model learns from examples, so this provision was about making sure government information does not become raw material for a company’s commercial products. The memo also called for agencies to secure rights in data and outputs, and for interoperability and portability terms to avoid lock-in, the situation in which switching to another vendor becomes so costly that an agency is effectively stuck.

M-25-22 rested on a statute. The Advancing American AI Act had directed OMB to make sure AI contracts align with the government’s AI guidance. The memo applied to agencies generally, though not to elements of the intelligence community, and it did not apply to AI bought as a component of a national security system. It also excluded common commercial products with AI embedded in them, such as a word processor that suggests text. To decide whether that exception applied, agencies were told to weigh two things: whether the product was widely available to the public, and whether it had substantial purposes other than AI.

The acquisition memo set its own clock. GSA had 100 days to plan public guides for the acquisition workforce. It had 200 days to build a web-based repository, open only to executive branch agencies, for sharing contract language, negotiated prices and other AI buying resources. Agencies had 270 days to update their internal acquisition procedures. Those updates had to bring privacy officials into planning early, set terms for intellectual property, and convene cross-functional teams that included acquisition, IT, security, privacy, civil rights, legal and budget staff.

M-25-22 applied to contracts awarded from solicitations issued 180 days or more after the memo, which meant from the end of September 2025. A solicitation is the formal request an agency publishes inviting vendors to bid. The memo also reached existing contracts: it applied to any option to renew or extend performance exercised after the same date. An agency could not avoid the new terms by extending an old agreement.

The headlines framed the April memos as a lighter touch. Read closely, they also moved a great deal of responsibility onto the people who write and manage government contracts.

The January 2025 rescission changed the vocabulary of federal AI work more than its substance, and the inventories created under the 2024 memo remained the durable plumbing. The April memos confirmed both points. Their more consequential move was elsewhere.

M-25-22 asked contracts to do much of the governance work that the earlier framework had placed with internal review. A review board can decide whether a tool should be used. Only the contract can compel a vendor to disclose, test, notify, cooperate and leave cleanly. That shifts real accountability to contracting officers and program managers at four stages of an acquisition: defining the requirement, setting terms at award, managing performance, and planning the exit.

Although the memos bind only federal agencies, the logic applies to state and local governments and, frankly, to any large organization buying AI.

The distinction matters because the two kinds of governance fail in different ways. Internal governance fails loudly: a board that never meets or an inventory that is never published is visible to OMB, inspectors general and the press. Contract governance fails without anyone noticing. A missing notification clause or a vague data-use term produces no finding until something goes wrong, and by then the agency has spent its bargaining power. The sections below follow the acquisition in order, because the agency’s bargaining power is highest at the start and falls with each stage.

What an agency is actually buying

When an agency “buys AI,” it usually is not buying a robot or a single program. It is typically buying access to a service, often a large language model of the kind behind popular chat assistants, that runs on the vendor’s computers. A large language model is software trained on enormous amounts of text to predict and generate language. It can draft letters, summarize documents, answer questions and classify cases.

A few related terms recur.

Hosted model. The model runs on the vendor’s servers in the cloud. The agency sends it questions (prompts) and receives answers (outputs). The agency never holds the model itself.

Fine-tuning. Adjusting a general model with an organization’s own examples so it performs better on that organization’s tasks, such as an agency’s forms and terminology.

Embeddings and retrieval indexes. To let a model answer questions about an agency’s own documents, those documents are converted into a kind of numerical map, called embeddings, and stored in an index the model can search.

Evaluation sets. Collections of test questions with known correct answers, used to check whether a model is performing well enough for a particular job.

Over time, an AI deployment accumulates assets that are specific to the agency: its prompts, its fine-tuned models, its indexes, its tests. Who owns those assets, where they live and whether they can be taken elsewhere are contract questions, not technical ones.

Three ways AI enters an agency

AI reaches a public agency through three doors, and each one puts the contract in a different position.

Direct model access. The agency buys access to a model through a cloud provider’s marketplace or an application programming interface, and its own staff or an integrator build the application. Here the agency controls most of the design and can write detailed terms on data handling, logging and evaluation. It also carries the most engineering responsibility.

AI-first products. The agency buys a product whose main purpose is AI, such as a document summarization service or a contact center assistant. The vendor makes most design choices. The contract is the agency’s only window into how the model is chosen, tested and changed.

AI embedded in something else. The agency already owns a case management, HR or productivity platform, and the vendor adds AI features to it. M-25-22’s exception for common commercial products often applies here, but not always. A feature that becomes the main reason staff use a product, or one built specifically for an agency, can fall back inside the memo. The two-part test, wide public availability and substantial non-AI purpose, has to be applied feature by feature.

A fourth route lies outside the agency’s purchasing entirely: contractors using AI to do the work they were hired for. A firm processing applications or drafting reports may run them through a model the agency never evaluated. M-25-22 told agencies to decide when solicitations should require vendors to disclose that kind of use. For services contracts that touch personal data or eligibility decisions, the disclosure clause matters as much as any term on a product the agency buys directly.

Why the contract became the control

Under the 2024 approach, much of the governance work happened inside the agency: review boards, impact assessments and approvals. Those still matter. But an internal board has a structural limit. It can tell its own staff what to do. It cannot tell a vendor what to do.

A homeowner can decide, at the kitchen table, that a renovation must use non-toxic materials, finish on time and leave the site clean. Unless those requirements are written into the contract with the builder, they are wishes. The contract is the only document the builder has agreed to follow, and the only one a court will enforce.

AI is the same, with one twist. Much of what matters about an AI system, including what data it keeps, whether it changes, and how it performs, sits on the vendor’s side of the line. The agency can see it only if the contract requires the vendor to show it.

That is why the memo’s shift matters in practice. It moves real responsibility onto contracting officers and program managers, people who are often stretched thin and rarely trained in AI. Giving them a clear, reusable set of terms, and someone technical to call on, is one of the most useful things an agency leader can do. The four stages below are where that responsibility shows up.

What the memo told contracts to carry

M-25-22 was specific about the terms. Contracts for AI systems had to address intellectual property rights and the use of government data, privacy, protection against lock-in, compliance with M-25-21’s minimum practices for high-impact uses, and ongoing testing and monitoring. The testing provisions were the most demanding.

• Agencies were to evaluate systems using their own data, such as validation and test sets the vendor could not see, kept as close as possible to the data the system would meet in service.
• Vendors had to give agencies the access and time needed for that independent evaluation. Where vendors did the testing instead, results had to be detailed enough to verify or reproduce where practicable.
• Contracts had to describe the vendor’s own testing procedures and could not prohibit the agency from disclosing internally how the vendor tests or what it found.
• Monitoring was expected on a defined cycle, for example quarterly or twice a year, based on the program’s needs.

The memo also encouraged agencies to require vendors to meet performance standards before deploying a new model version, and to roll back if the new version fell short. Each of these is a promise only a vendor can keep. None of them can be achieved by an internal policy alone.

The case against putting so much in the contract

A serious argument holds that this approach asks too much of procurement. The federal acquisition workforce was already stretched, and contracting officers are trained in competition rules, not model evaluation. Large AI vendors sell on standard terms to millions of customers and often decline to negotiate. A small agency buying a few hundred licenses has little power to change them. Detailed clauses also raise the cost of bidding, which can drive away the newer, smaller firms the memo itself wanted to attract.

Those objections are real, and they point to how the approach should be run rather than to abandoning it. Most of the burden falls on drafting terms once, not on each purchase. Government-wide vehicles let one negotiation cover many agencies, which is where GSA’s role became important. Terms can also be scaled to risk: the full set for a high-impact use, a short standard set for a drafting assistant. The alternative, relying on internal review alone, leaves the agency with rules it has no means to enforce against the party that controls the system.

Requirement: deciding what is being bought

Governance starts before a solicitation is written. The memo encouraged agencies to describe the outcomes they need and to plan how they will test whether a system delivers them. Many AI purchases begin instead as product demonstrations, with requirements written around a tool someone has already seen, which reverses the proper order.

The practical work is to decide early whether the use is likely to be high-impact under M-25-21. That classification drives everything downstream. A tool that informs benefit eligibility, fraud flags or enforcement priorities needs testing data, evaluation criteria and human review built into the statement of work, the part of the contract that describes what the vendor will deliver. A drafting assistant for internal memos does not.

Program offices should also ask whether AI is embedded in products they are buying for other reasons, such as case management or contact center software. Embedded features often arrive through a routine upgrade and escape review entirely. An agency may discover that its call center software now suggests answers to callers using a model nobody evaluated.

Market research belongs in this stage too. Before a solicitation, agencies can ask vendors how their models are evaluated, what data they retain, and how they handle government content. The answers reveal which offerings can meet the agency’s eventual terms, and they spare a contracting officer from discovering incompatibilities after award, when bargaining power is lowest.

Building the team before the solicitation

M-25-22 required agencies to convene a cross-functional team for AI acquisitions, with the effort scaled to the complexity and risk of the purchase. The memo’s list of relevant officials included acquisition staff and competition advocates, IT, cybersecurity, privacy, civil rights and civil liberties, budget, data, legal and program evaluation. That team was asked to produce an initial list of risks and an early view of whether the system was likely to host high-impact uses.

The team’s first practical output should be a decision on the acquisition technique. The memo strongly encouraged performance-based methods. A statement of objectives or a performance work statement describes the outcome the agency needs and lets vendors propose how to meet it. A quality assurance surveillance plan, or QASP, sets out how the government will measure performance after award. For AI, the QASP is where accuracy thresholds, review samples and monitoring cycles are written down. Agencies that skipped it had nothing concrete to hold a vendor to later.

The memo also required agencies, when practicable, to tell vendors in the solicitation whether the planned use was high-impact or likely to become so. That disclosure has a practical effect. It warns vendors that they will need to supply the documentation an agency needs for its AI impact assessment, such as a description of training data and known limitations. Vendors that cannot provide it self-select out, which is better than discovering the gap during deployment.

How the stage differs by agency

The 24 CFO Act agencies had governance boards, chief AI officers with staff and, in many cases, existing AI programs. They could run these teams as standing bodies. Smaller agencies and commissions usually had a chief AI officer who also held another job, and might buy AI only occasionally. For them, the realistic approach was to borrow: use a government-wide vehicle, adopt terms another agency had already negotiated, and spend their own limited expertise on the requirement and the test data.

Demonstrations still have a place. The memo encouraged agencies to test candidate systems in conditions close to real use, including on the agency’s own networks where possible. A demonstration run against the agency’s own sample cases is evidence. A demonstration run against the vendor’s prepared examples is marketing.

Award: data and rights terms

The award is where the memo’s data protections become enforceable. Agencies were told to restrict vendors from using non-public agency data to train publicly or commercially available models without explicit consent. They were also told to secure the rights they need in government data, outputs and any derived models, and to address intellectual property in a tailored way rather than accepting commercial defaults.

These terms sound standard, but most commercial AI agreements lack them and typically reserve broad rights for the vendor to improve its services. Contracting officers need specific language on four points.

• Where agency data is stored and processed.
• Whether prompts and outputs are retained, and for how long.
• Whether any agency data feeds model improvement.
• Who owns fine-tuned models or retrieval indexes built from agency content.

Each point needs a verification mechanism, such as an attestation (a signed statement from the vendor) or an audit right (the agency’s right to check). A technical control can also serve, such as tenant isolation, which walls one customer’s data off from every other customer’s on the vendor’s systems. Without one, the agency has no way to confirm the vendor kept its promise.

Privacy, minimization and the authority to operate

M-25-22 tied its data terms to existing privacy law and OMB Circular A-130. Where an AI system would create, collect, use or store personally identifiable information, agencies had to involve their senior agency official for privacy from pre-solicitation planning onward. The memo also set a minimization rule: vendors should collect and retain agency information only when reasonably necessary for the contract’s purposes. For hosted models, that principle maps directly to retention settings for prompts, outputs and logs.

The prohibition on training was stronger than most commercial terms. The memo told agencies to ensure contracts permanently prohibit using non-public agency inputs and outputs to train publicly or commercially available models, absent explicit agency consent. “Permanently” matters. Many commercial agreements allow the vendor to change data-use terms on notice, and a prohibition that lapses at renewal protects nothing.

Award is also where security authorization meets AI governance. The memo reminded agencies that any AI system operated as an information system for an agency needs an authorization to operate before deployment. For cloud services, that usually means a FedRAMP authorization at the right impact level. The two reviews ask different questions. The security review asks whether the system is protected. The AI review asks whether its outputs are fit for the decision they inform. A system can pass one and fail the other.

Award terms outside the federal government

State and local buyers work under their own procurement codes, not the Federal Acquisition Regulation, and often buy through cooperative contracts negotiated by a lead state or a purchasing consortium such as NASPO ValuePoint. Those master agreements set baseline terms that each participating entity can supplement. A city that relies on a cooperative contract should read its AI and data provisions closely and add its own participating addendum where they fall short. State privacy statutes and public records laws add obligations a federal buyer does not face, including rules on how long conversation logs must be retained and who can request them.

Performance: watching a system that changes

Traditional software changes when the agency installs an update. Hosted AI models can change underneath an agency without notice. A model version swap can alter accuracy, tone or refusal behavior (when a model declines to answer) in ways that matter for a high-impact use. M-25-21’s minimum practices include ongoing monitoring, and monitoring is impossible if the vendor will not say what changed.

Suppose a state benefits agency uses a hosted model to summarize applications for caseworkers. One month, the summaries start leaving out a field that caseworkers rely on. Nobody at the agency changed anything. The vendor swapped in a newer model version. Without a requirement to notify the agency of changes and share test results, the agency learns about the problem from a backlog of complaints rather than a notice.

Contracts therefore need model change notification, access to the evaluation results the vendor runs, and logs sufficient for the agency to investigate a complaint or an error. The contracting officer’s representative, a role often staffed by program personnel with little technical background, carries day-to-day oversight. Agencies that paired those representatives with a technical evaluator from the chief AI officer’s team had a realistic chance of noticing drift, the gradual change in a system’s behavior over time. Agencies that did not were relying on the vendor to report its own problems.

Incident terms belong here as well. A wrong benefit determination produced by a model is not a security breach, but it may need the same urgency. Contracts should define what counts as an AI incident, how fast the vendor must report one, and what cooperation it owes the agency’s investigation.

Contract administration in detail

The memo’s contract administration section gave the performance stage its structure. Agencies had to oversee systems against the contract, including monitoring for emerging risks to privacy, civil rights and civil liberties. They were encouraged to evaluate periodically whether the system still justified its cost, considering effectiveness, risk and operating expense. They were also encouraged to set sunset criteria in advance, such as a change in cost, mission need, vendor terms or model performance that would trigger a decision on continued use.

Sunset criteria are the least used of these tools and among the most valuable. Without them, an AI service tends to continue because stopping it requires a decision nobody is assigned to make. With them, the agency has agreed in advance on what would end the arrangement. A drop in accuracy below a set threshold, a price rise above a set percentage, or a vendor’s refusal to provide test results each becomes a defined event with a defined owner.

The evidence an auditor would ask for at this stage is predictable. It includes the QASP and its results, records of model change notices and the agency’s response, any re-run evaluation results, complaints and their resolution, and the chief AI officer’s determinations for the use. An agency that cannot produce those records has, in effect, no performance management, whatever its policy says.

Remedies depend on vendor cooperation

M-25-21 requires agencies to give people affected by high-impact AI decisions access to timely human review and an appeal. Agencies can extend an existing appeals process to cover AI-assisted decisions. A human reviewer can only reconsider a decision meaningfully if the agency can reconstruct what the system did. That requires inputs, outputs, model version and configuration to be logged and retained in a form the agency can read. The appeal right is created by policy, but whether it means anything depends on the contract.

Exit: leaving without losing the work

Lock-in was the memo’s most distinctive concern. AI systems accumulate agency-specific assets over time: prompt libraries, fine-tuned models, embeddings, evaluation sets and workflow integrations. If those assets exist only in the vendor’s proprietary formats, recompeting the contract means starting over, which makes recompetition unlikely. Years of agency work built on the vendor’s platform would stay behind when the agency left.

Exit terms should require export of agency data and agency-specific artifacts in documented formats, knowledge transfer to a successor, and transition assistance for a defined period. Pricing transparency matters too. Introductory pricing that rises sharply at renewal is a form of lock-in when switching costs are high.

What can actually be carried out

Portability for AI is narrower than it sounds. Fine-tuned versions of proprietary models usually cannot be exported in a form another vendor can run, because the underlying model weights belong to the vendor. An agency that writes “transfer the fine-tuned model” into an exit clause may receive nothing usable. The assets that can move are the ones the agency created: training and evaluation data, prompt libraries, labeled examples, retrieval documents, configuration records and integration specifications.

The realistic goal is therefore re-creatability. A successor should be able to rebuild equivalent capability from the agency’s own materials in weeks rather than years. That shapes the terms. The agency should require those materials to be maintained in documented, open formats throughout the contract, not assembled in a rush at closeout. M-25-22 pointed the same way, encouraging well-defined interfaces, documentation that supports moving tools between vendors, and rights to code and models produced in performance of the contract.

Closeout begins at the decision not to renew

The memo told agencies to start closeout as soon as they decide not to extend an AI contract. That means agreeing with the vendor on the format and usability of data and derived products, any circumstances under which access would expire, and a plan for transferring assets. Waiting until the final month gives the outgoing vendor every incentive to deliver the minimum.

The memo also named pricing practices that work against competition. It suggested agencies favor vendors who avoid tying arrangements, minimum spend requirements, exclusive discounts and similar terms that consolidate spending with one supplier. Those practices matter at exit because they raise the cost of leaving. An agency whose discount depends on buying everything from one vendor has given up some ability to recompete any part of it.

State and local agencies are not bound by the memo, but they buy from the same vendors, and clauses written for federal buyers tend to appear in their agreements. A county or state that understands the federal clauses can ask for the same protections.

The trade-off across all four stages is speed. Each of these terms takes negotiation, and the memos also told agencies to move faster. Agencies that drafted a standard AI clause set once, with procurement and the chief AI officer together, could buy quickly without giving up protections. Agencies that negotiated each deal from scratch tended to trade governance away for speed.

Who bears the cost of a thin contract

Thin contract terms show up later as wrong benefit decisions, rising renewal prices and questions from auditors that the agency cannot answer.

The public. When AI informs who receives benefits, which cases get flagged for fraud or where enforcement attention goes, errors fall on real people. A contract that lets the agency see test results, learn about model changes and investigate complaints is, in practice, a protection for the people the agency serves.

Money. AI services are often priced attractively at first. The real cost appears over years, in renewals, in the staff time spent building prompts and indexes, and in the price of switching. Exit and pricing terms decide whether that investment stays with the agency or stays with the vendor.

Oversight. Inspectors general, auditors, legislators and courts will ask agencies to explain how an AI-assisted decision was made. Agencies can answer only if the contract gave them logs, evaluation results and vendor cooperation.

Reputation and trust. A single public failure, such as a chatbot giving residents wrong information about deadlines, can undermine confidence in an agency’s whole digital program. The contract is where an agency decides in advance how it will detect and respond to such failures.

Who pays differs by type of agency

For a federal benefits agency, the costs of a thin contract fall mainly on applicants, and they show up as appeals, congressional inquiries and inspector general reviews. For a state agency administering federally funded programs, they can also show up as findings from the federal funding agency. For a county or city, the exposure is usually more immediate. It arrives as a local news story, a council hearing, or a public records request for chatbot transcripts that the county cannot produce because the vendor holds them. Smaller governments also feel price increases more sharply, since a renewal increase can consume a line item with no room to absorb it.

Regulated public entities face one more layer. A public hospital, a state-chartered utility or a criminal justice agency must meet HIPAA, sector rules or the FBI’s CJIS Security Policy on top of procurement law. AI terms have to fit inside those obligations. A vendor willing to sign a business associate agreement for a health system, or to meet CJIS requirements for a sheriff’s office, may still refuse terms on model changes or exit. Each of those has to be negotiated separately.

A county buys a resident assistant

Suppose a county government wants an AI assistant on its website to answer residents’ questions about permits, property taxes and trash pickup. It is not bound by the federal memos, but its procurement director decides to borrow their logic.

At the requirement stage, the county’s team asks whether the assistant could significantly affect residents’ rights or access to services. Answering general questions is lower risk, but if the assistant will tell residents whether they qualify for a tax exemption, the stakes rise. The team decides the assistant may explain rules and link to forms, but may not make eligibility determinations, and writes that boundary into the statement of work along with a set of test questions.

At award, the county insists that residents’ questions are not used to train the vendor’s commercial models, that conversation logs are kept for a set period in the county’s control, and that the index built from county documents belongs to the county.

During performance, the contract requires the vendor to give notice before changing the underlying model and to rerun the county’s test questions afterward. A staff member in the IT office reviews the results each quarter alongside the program manager.

For the exit, the contract requires the vendor to export the index, the test questions and the conversation logs in documented formats, and to support a transition for a defined period. When the contract comes up for renewal, the county can credibly invite other bids. That credibility is what keeps the renewal price honest.

The approach has limits, and the county should know them. A small county has little negotiating power with a large vendor, and some terms will be refused. The procurement director then has three options. The county can buy through a cooperative contract that already includes stronger terms. It can narrow the use until the remaining risk is acceptable under standard terms. Or it can choose a smaller vendor willing to negotiate. Each choice trades capability, price and protection differently, and the decision should be written down.

Habits of agencies that bought well

Agencies that took the contract-centered approach seriously tend to show a few habits.

• They classify each proposed AI use against the high-impact definition before writing the requirement.
• They hold a standard AI clause set, agreed by procurement, counsel and the chief AI officer, covering data, rights, monitoring, incidents and exit.
• They pair every contracting officer’s representative on an AI contract with someone who can read test results.
• They keep an inventory of AI features inside products bought for other reasons.
• They plan the exit at the start, including what will be exported and in what format.
• They write a quality assurance surveillance plan with numeric thresholds and keep the test data away from the vendor.
• They agree sunset criteria before deployment, so stopping a weak system is a scheduled decision rather than a fight.

The errors are easy to spot in hindsight.

• Buying the demo. Writing requirements around a product someone liked, instead of the outcome the agency needs.
• Accepting commercial terms by default. Standard vendor agreements often allow broad use of customer data and leave ownership of derived assets vague.
• Assuming the system stays the same. Hosted models change, and contracts that do not require notice leave the agency blind.
• Ignoring introductory pricing. Discounts that end at renewal, combined with high switching costs, can turn a bargain into a trap.
• Treating governance and speed as opposites. A reusable clause set lets agencies have both.

Where the bargaining power sits at renewal

1. Which of our current or planned AI uses would count as high-impact, and who made that call?
2. Do our AI contracts say clearly where our data is stored, how long prompts and outputs are kept, and whether any of it trains the vendor’s models?
3. Who owns the fine-tuned models, indexes and test sets built from our information, and could we take them with us?
4. Will the vendor tell us before it changes the model we rely on, and will it share the results of its own testing?
5. What happens to our pricing at renewal, and what would it cost us, in time and money, to switch vendors?

With hindsight

The contract-centered approach was tested quickly. In July 2025 the administration released America’s AI Action Plan and Executive Order 14319, Preventing Woke AI in the Federal Government, which added an ideological-neutrality expectation for large language models the government buys. Weeks later, through GSA’s OneGov program, agencies gained access to major commercial AI models at steep introductory discounts. In August 2025 several of the largest model developers offered enterprise versions of their assistants to agencies for nominal prices, in some cases a dollar per agency for a year. GSA, the General Services Administration, is the federal government’s central buyer, and OneGov is its effort to negotiate government-wide deals with large technology vendors. Cheap, fast access is attractive, and it made data, logging and exit terms urgent: the easier it is to start using a tool, the more important it is to know what happens to the data and how to leave.

M-25-22’s 180-day mark arrived at the end of September 2025. From then on, new solicitations and option exercises had to carry the memo’s terms. Because the memo reached options on existing contracts, the effect was spread over many months as option years came due, rather than landing on one date.

On December 11, 2025, OMB memo M-26-04 implemented the July order. It required solicitations and orders for large language models issued after that date to include contract terms on the administration’s two Unbiased AI Principles, truth-seeking and ideological neutrality. Agencies were told to modify existing LLM contracts where practicable, at the latest before exercising an option. Their procurement policies had to be updated by March 11, 2026, including a way for users to report outputs that violated the principles. The memo described itself as complementing M-25-22 and set itself to expire after two years unless OMB extended it.

Whatever one thinks of its substance, M-26-04 confirmed the mechanism. The administration’s chosen tool for changing how vendors’ models behave was a contract clause, applied at the solicitation and at the option exercise. That was the same lever M-25-22 had built, and agencies that already had a clause set and a process for amending it could absorb the new requirement more easily.

By September 2026, as the first OneGov AI offers expired, GSA was renewing some on new commercial terms. On September 10, 2026, it announced a 27-month OneGov agreement with OpenAI, expected to take effect on October 1. The offer replaced flat nominal pricing with a discount on metered, token-based usage, with no platform fee, minimum order or spend commitment. It was open to state, local and tribal governments as well as federal ones, and GSA said it incorporated the agency’s own AI terms and conditions. GSA presented the agreement as supporting the AI Action Plan, M-25-21 and M-25-22. Renewal is exactly the moment the memo’s exit and portability terms are meant to matter. It is when an agency learns whether it can realistically walk away, and therefore whether it can negotiate.

The shift from nominal fees to consumption pricing changed the risk agencies carried. Under a flat fee, the cost of heavy use fell on the vendor. Under metered pricing it falls on the agency’s budget, and it grows with every workflow staff build on the tool. That makes usage monitoring, budget alerts and the exit terms in the original order more important, not less.

The memo’s contract-first logic held up. Whether agencies used that bargaining position at renewal is the open question. Public reporting through late 2026 said little about how many agencies wrote exit, portability or evaluation terms into their own orders, as opposed to relying on terms negotiated centrally by GSA.

Inside and outside government, the conclusion is simple. The protections that matter most with AI vendors are the ones written down before the agency becomes dependent. Introductory discounts end. Models change. Staff who negotiated the original deal move on. What remains is the contract, and an agency that wrote its exit terms on day one arrives at renewal with options rather than obligations.

That applies just as much to a city or a state buying the same products as it does to a federal department, and leaders at every level can ask to see those terms before they sign.