Invest1 publisher2 min readPublished
OMB redefines agency Chief AI Officers as adoption advocates and risk advisers
The White House's two revised memos fold AI approvals into the IT accountability agencies already run and tell them to buy American AI, without saying in the fact sheet what makes AI American or high-impact.
The Investor · Invest desk

What happened
- OMB, working with the Assistant to the President for Science and Technology, issued two revised AI policies for the Executive Branch, M-25-21 on use and M-25-22 on acquisition.
- Accountability for federal AI is to mirror the process agencies already run for government IT, instead of creating new layers of approvals for AI specifically.
- The policies create a single high-impact AI category covering use cases that need heightened due diligence because of potential effects on Americans' rights or safety.
- The acquisition memo removes agency reporting requirements it calls burdensome and says the acquisition process is optimised while privacy and lawful use of government data continue to be protected.
- Agencies are told to support a competitive American AI marketplace and to maximise use of American AI systems and services when they seek new AI solutions.
Compiled by The InvestorSomething wrong?How this is made
Why it matters
- decision A vendor's first meeting is with an officer told both to champion adoption and to classify risk, so the deal turns on which side of the lower-risk line the use case is put.
- constraint Requirements written to avoid vendor lock-in are requirements written to be recompeted, so the policy that makes entry cheaper also shortens how long a winner holds a programme.
- exposure Suppliers with foreign ownership, foreign-trained models or foreign hosting face a preference with no published test, and the burden of proving Americanness falls on the bidder.
- precedent Routing AI through existing IT governance makes an agency's software accountability chain the default venue for arguments about models, staffed by people who never signed up to adjudicate them.
One officer at each agency now holds three jobs at once: advocate for adoption, classifier of risk, adviser on AI spending [3][4][14]. For a vendor that turns a compliance problem into a sales problem. Land on the lower-risk side and the product inherits the accountability chain the agency already runs for its IT [6]; land in the high-impact category and it inherits heightened due diligence whose steps the fact sheet does not set out [5][17].
How unsettled the sort is shows in the White House's own illustrative list: four uses across three agencies, two of them at Veterans Affairs, where AI standardises patient care and flags pulmonary nodules during lung cancer screening, one at Justice analysing the global drug market for trafficking investigations, and one at NASA navigating the Mars 2020 rover [13][15]. The document places none of the four in either category [17], and screening support in oncology sits close enough to safety that a clinical vendor should treat the classification as a live question rather than a formality.
The phrase "American AI" appears five times in the fact sheet, which never says what makes AI American [16]. Corporate domicile of the seller, ownership of the weights, where the training compute ran, whose paper the reseller holds: each reading admits a different set of bidders, and the pairing of a competitive marketplace with maximised American use does not choose between them [10][11].
Read as acceleration, the package lowers an agency's cost of trying a supplier, which tends to produce more pilots at smaller sizes rather than fewer large awards [8][9]. Read as risk triage, or rather the version that matters to the biggest programmes, heightened due diligence for high-impact systems may run slower than the approvals it replaced, and rights and safety are exactly where the large agency caseloads sit [5]. Agencies are also to produce an AI adoption maturity assessment, and the fact sheet attaches no date to it [7][17]. What would settle the question is the memo text rather than the summary of it: if M-25-21's high-impact procedures take longer than what they replace, or if the existing IT accountability chain proves to be the slower gate, the change is one of venue and not of pace [6][17]. Until then, the qualifying bar for selling into the Executive Branch is one officer's classification, and the White House has told that officer to say yes to the lower-risk half [4][12].
What to watch
- The published text of M-25-21 and M-25-22, which is where the high-impact due diligence steps either add time or do not.
- Whether agency AI adoption maturity assessments are published, and what they say about unmet needs.
- Any agency guidance that puts an eligibility test behind the phrase American AI, such as domicile, ownership or training location.