Revoking the AI executive order could not revoke the AI inventory, because Congress had written it

The week

On January 20, 2025, the day of his inauguration, President Trump signed an executive order rescinding a long list of prior executive actions. Signed between ceremonies, it was one of the day’s less noticed acts and one of the most consequential for anyone running artificial intelligence programs in the federal government. An executive order is a written directive from the President to the executive branch; it carries the force of law for federal agencies, but a later President can cancel it with another signature. Among the orders cancelled was Executive Order 14110 of October 2023 on the safe, secure and trustworthy development and use of AI, the previous administration’s main framework for how government should handle the technology.

Three days later, on January 23, he signed an executive order on removing barriers to American leadership in AI. It directed the development of an AI action plan within 180 days and ordered a review of actions taken under the revoked order.

It also told OMB to revise, within 60 days, its 2024 memoranda on agency use and acquisition of AI. OMB, the Office of Management and Budget, is the White House office that turns presidential policy into specific instructions for agencies, usually through documents called memoranda. Those memoranda had required Chief AI Officers, AI use case inventories and minimum practices for rights-impacting and safety-impacting AI, and agencies had spent 2024 implementing them.

A use case inventory is a list of every place an agency uses AI: a chatbot answering questions, a model that flags suspicious tax returns, a tool that helps sort benefit applications. “Rights-impacting” and “safety-impacting” uses were the ones that could affect people’s rights or physical safety, and they had to meet a minimum set of protections such as testing and human review.

The same week, a large private AI infrastructure venture was announced at the White House, underlining the new emphasis on building and adopting the technology quickly.

For the officials who had spent a year standing up AI governance, the week posed a practical problem. With the order gone and the memoranda under review, they had to work out which parts of their work still had to be done. The answer, it turned out, depended less on the subject of each task than on what kind of legal instrument stood behind it.

For agency AI leaders, the week opened a gap. The executive order behind a year of governance work was gone. The OMB memoranda that turned it into agency requirements were still on the books but marked for revision. Nobody knew what the replacements would keep.

An obligation’s legal instrument tends to predict its survival better than its subject does, and rescission week put that to the test.

Agency AI work in late January 2025 rested on one of three footings. Some activities rested only on the revoked order. Some rested on OMB memoranda that remained in force pending revision. And some rested on statute, which no executive order could remove. Every activity in an agency AI program falls under one of those three, and each called for a different decision during the interval.

Statutes, orders and memoranda

The three kinds of instrument behave very differently when administrations change.

Statutes are laws passed by Congress and signed by the President. They are the bedrock. Only Congress can change or repeal them, and a new President cannot simply sign them away. If a statute says an agency must do something, the agency must keep doing it until Congress says otherwise.

Executive orders are directives from the President to the executive branch. They are binding on agencies, but they last only as long as the President who signed them, or a successor, wants them to. One signature creates them; another removes them.

OMB memoranda are the detailed instructions that tell agencies how to carry out policy. They often implement an executive order, a statute, or both. They can be revised or replaced by OMB itself, which makes them more durable than a revoked order’s specific deliverables but less durable than a law.

The arrangement resembles a building. Statutes are the foundations and load-bearing walls. Executive orders are the interior design scheme a new owner can rip out. OMB memoranda are the building management rules: they can be rewritten, but the building keeps running under the old ones until the new ones are posted.

Why this matters to AI programs specifically

Federal AI governance in 2024 had been assembled from all three layers at once. An agency official doing the work day to day often could not say which layer a given task came from, because the guidance documents cited each other. That was fine while all three pointed the same way. When one layer was removed overnight, the distinction suddenly mattered a great deal.

What AI governance looks like inside an agency

For leaders who have not seen it up close, “AI governance” can sound abstract. In practice it is a handful of concrete routines, most of which would be familiar to anyone who has run a risk or compliance function.

Knowing what you have. The use case inventory is the foundation. Someone has to ask each program office what AI tools it uses or is building, record what each one does and keep the list current as projects start and stop. Without it, nothing else is possible, because an agency cannot review or improve tools it does not know exist.

Deciding what needs a closer look. Most AI uses are low stakes, such as summarizing documents for internal use. A few affect people directly: who gets a benefit, which case gets investigated, which traveler is pulled aside. Governance means sorting the inventory so that the high-consequence uses get more attention.

Checking that high-consequence tools work. For those uses, the minimum practices included steps such as testing a tool before relying on it, monitoring it afterward, making sure a human could review or override its outputs and giving people a way to raise concerns.

Buying carefully. Much government AI is bought rather than built. Acquisition expectations shaped what agencies asked vendors to provide: documentation, testing results, and terms about data and performance.

Someone in charge. The Chief AI Officer and the governance board tie it together, deciding what proceeds, what waits and what needs more work.

Seen this way, most of these routines are not tied to any one administration’s priorities. They are what any careful organization does with a powerful tool. That is part of why sorting them by legal footing, rather than by the language used to describe them, was so useful.

Footing one: work resting only on the revoked order, pause and document

EO 14110 had assigned many specific deliverables: reports with deadlines, interagency working groups, and tasks framed in the order’s language on equity and civil rights. Where an activity existed only because the order required it, the requirement ended on January 20.

The sensible step was to pause such work and record what had been done, not to delete it. Assessments, test results and documentation already produced had value for any future framework, for audits and for litigation. An agency that discarded them risked needing them within months.

Deletion also raised a separate legal issue. Documents created in the course of agency business are federal records, and the Federal Records Act governs their disposal whatever happens to the policy that prompted them. A change of policy does not give anyone permission to shred the files. Records officers should have been part of any decision to wind down order-specific work.

The risk of overcorrection was real. An activity described in the old order’s language might still serve a purpose the new policy would value. Measuring whether a benefits model makes more errors for some groups of applicants, for example, is also a measure of accuracy. A model that is wrong more often for one group is, by definition, less accurate, and an administration focused on effective government has every reason to want to know that.

Footing two: work resting on OMB memoranda in revision, keep running lightly

The 2024 memoranda had created Chief AI Officers, AI governance boards and minimum practices for rights- and safety-impacting uses, and set acquisition expectations. A Chief AI Officer is a senior official in each agency responsible for coordinating its use of AI; a governance board is the group of senior leaders that reviews major AI decisions. The January 23 order directed the memoranda’s revision, not their immediate withdrawal. For roughly ten weeks they sat in limbo.

The practical approach was to keep these functions running at reduced intensity. Chief AI Officers stayed in place. Reviews of high-consequence uses continued, but agencies could defer new procedural steps that might not survive.

The new administration’s emphasis on accelerating adoption gave these functions a second justification. An official who knows where AI is used, which projects deliver value and which are stuck is as useful for removing barriers as for imposing controls. Agencies that described their governance this way, as the machinery that lets AI scale safely, were better placed in budget discussions than those that described it as compliance with a revoked order.

The procurement question

Procurement deserved particular care. A solicitation is the formal request an agency issues when it wants to buy something, setting out what it needs and the terms vendors must accept. Solicitations in progress that cited the 2024 acquisition memorandum needed a decision: proceed under existing terms, or amend.

Contracting officers, the officials with legal authority to sign contracts on the government’s behalf, needed guidance quickly, because vendors were asking. A short written instruction from the senior procurement executive was enough to keep awards moving. Without one, each contracting officer would have had to decide alone, and the predictable result would have been delay.

Footing three: work resting on statute, unaffected

The third footing was the most important and the least discussed that week. Congress had legislated parts of the federal AI governance framework.

The Advancing American AI Act, enacted in the fiscal 2023 defense authorization, required agencies to prepare and maintain inventories of their AI use cases, share them across government and publish them where lawful. The AI in Government Act of 2020 required OMB guidance on agency AI use. Neither depended on an executive order.

The inventory provision had its own limits worth knowing. It was written as a five-year requirement, and it did not apply to the Department of Defense or the intelligence community. An agency program could therefore plan on the inventory continuing at least into 2027, with its scope set by the statute rather than by whichever memorandum was current.

Other statutes kept applying as before. The Privacy Act governs systems of records, including those that feed models. The E-Government Act requires privacy impact assessments, which are structured reviews of how a system collects and uses personal information. FISMA, the Federal Information Security Modernization Act, governs the security of agency systems, AI included. Civil rights laws apply to decisions about the public whatever tools inform them. And the Administrative Procedure Act still lets people challenge arbitrary agency action, which reaches automated decisions that cannot be explained.

For this footing, the right decision was simply to continue. An agency that paused its use case inventory because the executive order was gone would have been out of compliance with a statute.

One agency sorts its work plan

Say a federal benefits agency uses AI in three places: a chatbot on its website, a model that prioritizes applications for review, and a tool that drafts routine letters.

On January 21, its Chief AI Officer convenes a short meeting and asks, of every item on the governance work plan, what its legal footing is.

A report on equity impacts, due under the revoked order, rests on the order alone. The team stops work on new sections, files the analysis completed so far as a federal record, and notes that its accuracy findings on the prioritization model are worth keeping regardless.

The governance board and the review of the prioritization model rest on the OMB memoranda. They continue, but the board agrees to postpone a new documentation template until the revised memoranda arrive.

The use case inventory rests on the Advancing American AI Act. It continues unchanged. So do the privacy impact assessment for the chatbot and the security controls on all three systems, which rest on the E-Government Act and FISMA.

Meanwhile, a solicitation for a new letter-drafting tool cites the 2024 acquisition memorandum. The senior procurement executive issues a one-paragraph instruction to proceed under existing terms and amend later if needed. The award goes ahead.

Notice what the Chief AI Officer did not do. She did not wait for the new memoranda before acting, and she did not try to predict what they would say. She simply asked what each piece of work was built on and let the answer drive the decision. That approach required no forecasting of politics, only a clear reading of the rulebook, and it produced decisions that staff, auditors and vendors could all understand.

By the end of the week, the agency has a one-page table showing every activity, its footing and its status. When the revised memoranda arrive, it knows exactly what to check.

The same sort in banks and hospitals

The same sorting applies to any organization that lives under shifting rules. A bank operating under supervisory guidance, a hospital working to accreditation standards or a contractor bound by agency policies all face moments when part of the rulebook changes and part does not.

Take a regional bank that built a model-risk program partly around a supervisor’s guidance letter and partly around requirements written into law. If the supervisor withdraws the letter, the bank faces the same three-way sort: what existed only because of the letter, what rests on standing internal policy, and what the law still demands. A bank that has tagged its controls this way can respond in days. One that has not may spend months untangling which controls it can safely retire and which it cannot.

Cost. Organizations that cannot tell which requirements are durable tend to overspend on those about to disappear and underspend on those that will last.

Legal exposure. Stopping work a statute requires creates a compliance gap; destroying records can create a legal problem of its own.

Credibility. Leaders who can explain what still applies, and why, give staff and vendors confidence during a confusing period. Those who cannot invite paralysis or improvisation.

Continuity. Governance described in terms of its purpose, such as knowing where AI is used and whether it works, survives changes in political emphasis far better than governance described in terms of a particular document.

Pause, keep running, continue

Agencies that tagged each governance activity by its legal footing that week avoided both mistakes: spending on requirements about to disappear, and stopping work the law still required. Their approach looked like this:

• A single list of governance activities, each tagged as order-only, memorandum or statute.
• Order-only work paused, with completed work archived as federal records.
• Memorandum-based functions kept running, with new procedural steps deferred.
• Statutory work, including the use case inventory, continuing without interruption.
• A prompt written instruction to contracting officers on in-flight solicitations.

Others went wrong in predictable ways:

• Stopping everything. Treating the revocation as the end of AI governance put statutory duties at risk.
• Changing nothing. Spending heavily on deliverables that no longer had a legal basis.
• Deleting work. Discarding assessments that would be needed for audits, litigation or the next framework.
• Silence on procurement. Leaving contracting officers to guess, and vendors to wait.

Sorting your own obligations

1. For each of our AI governance activities, what is the legal basis: an executive order, a policy memorandum, or a statute?
2. Which activities would stop tomorrow if a particular policy document were withdrawn, and which would continue regardless?
3. What have we already produced that we should keep, and have our records officers been consulted?
4. Is our AI use case inventory current, and do we treat it as a legal obligation rather than a policy preference?
5. Are any of our current solicitations or contracts tied to guidance under revision, and has anyone told the contracting officers what to do?

With hindsight

The footing proved predictive. On April 3, 2025, OMB replaced the 2024 guidance with two memoranda, on federal use of AI and on AI acquisition. They framed policy around accelerating adoption but kept Chief AI Officers, use case inventories and minimum risk practices for what they called “high-impact” AI, a category close to the earlier rights- and safety-impacting uses.

Much of the machinery that agencies had kept running “lightly” during the ten-week gap survived the transition, renamed and reframed rather than removed. Agencies that had dismantled it would have had to rebuild it within months.

Inventories continued as a statutory duty: the new use memorandum listed the Advancing American AI Act among the laws its requirements implemented and told agencies to keep updating their annual use case inventories. The work that rested on a statute in January was still required in April, exactly as its legal footing suggested it would be.

In July 2025 the administration released its AI Action Plan and further executive orders, including one setting requirements for large language models procured by agencies. Large language models are the kind of AI behind modern chatbots and writing assistants, trained on vast amounts of text to generate human-like responses. That order showed that the new administration, too, would use executive instruments to shape how agencies buy and use AI, which means those requirements carry the same vulnerability to future revision as the order revoked in January.

Budget and staff reductions in 2025 made sustaining governance harder at many agencies, whatever its legal footing. A duty that survives on paper still needs people to carry it out.

Inside government and out, the practical point holds. When rules change, the first thing to establish is what each obligation was actually built on, before guessing at the new policy. That can be answered in an afternoon, and the answer tells an organization what to pause, what to keep and what it cannot stop.