OPM has told agencies that where an official independently reviews the record and adopts the result, an AI output is generally not a principal basis for the hiring decision, and therefore generally not high-impact under the federal AI memorandum

The move the memo makes

OPM issued a memorandum on 27 August 2026 on the use of artificial intelligence in the federal hiring process. Its substance is an application of the high-impact definition in OMB Memorandum M-25-21 to hiring.

The hook is a footnote in the OMB memorandum. OPM quotes it directly, identifying it by number: footnote 27 states that AI may be integrated in decision or activity pipelines in high-impact categories without meeting the definition of high-impact because the AI's output does not actually serve as a principal basis for the relevant type of agency action or decision.

OPM then supplies its own application. In OPM's view, where an appropriate agency official independently reviews supporting data, evaluates the decisional logic, and adopts the AI-assisted work product, the output generally does not serve as a principal basis for the resulting decision or action.

The memo immediately guards that sentence against the reading it would otherwise invite. It says that is not because human oversight, by itself, exempts a use from the high-impact definition, but because independent review, evaluation, and adoption supply the decision's own justification.

That distinction is the whole memo. Oversight is not the test. Whose reasoning carries the decision is the test.

The test cuts both ways, and the memo says so

It would be easy to read this as a permission slip. The memo forecloses that reading in its own words.

The inquiry is functional, not formal: identify what in fact supplies the justification for the decision or action in the documented workflow. An AI output may serve as a principal basis even where a human formally approves the result. Conversely, an output that merely informs a decision that is independently justified by the record does not.

And more directly: a high-impact determination is possible whether there is or is not human oversight for the decision or action. Asserted human validation or review does not, by itself, remove a use from the definition. What matters is whether the documented record shows that the official's independent judgment, rather than the AI output, supplied the decision's justification.

The operative word there is documented. The memo locates the answer in the record of the workflow, not in the org chart. A rubber stamp leaves a record that looks like a rubber stamp.

The four classes, and the conditions attached to each

OPM names four uses it considers generally not high-impact under M-25-21, after following the assessment process in its Appendix A.

First, content development for job opportunity announcements, including position descriptions, evaluation statements, announcement text, job analyses, qualification requirements, crediting plans and interview questions, when the process of drafting content itself does not affect an individual's access to employment. The memo adds that before operational use the material must be reviewed, validated and adopted under classification, job analysis and subject matter expert requirements, which it notes apply independently of the AI memorandum.

Second, decision-support for applicant-level evaluations, including resume screening, minimum qualifications and eligibility review, and candidate summarisation, when the AI output is not a principal basis for a decision. The memo's own example requires the AI to reference back to the underlying documentation and the deciding official to review that underlying record and decide on that basis.

Third, pre-offer quality control, when the AI is limited to administrative completeness and quality-control support and is not a principal basis for approving, delaying or rejecting the selection or final offer.

Fourth, aggregate hiring program evaluation, covering completed hiring actions, applicant flow, time-to-hire and assessment performance, where no individual employment decision turns on the AI output.

Every one of the four carries a when or where clause. None is a categorical clearance, and the phrase throughout is generally not high-impact rather than not high-impact. A determination still has to be made and documented for the particular workflow.

What survives a not-high-impact determination

The memo is explicit that clearing the high-impact hurdle changes less than it appears to. A non-high-impact determination changes nothing else, and AI-assisted steps remain subject to job analysis and validation requirements.

It then lists five things agencies must verify, before deployment and periodically: that AI-assisted rating, ranking and referral correctly apply veterans' preference; that AI-enabled steps meet accessibility and reasonable accommodation requirements; that applicant data is processed only in agency-authorised tools with applicable privacy protections; that each hiring action remains documented and auditable; and that a reconsideration avenue remains available for adverse eligibility determinations.

Note the modality shift. The four use classes are phrased as generally not high-impact. These five are phrased as things agencies must verify. That is the difference between a characterisation and a duty, and it is the part a compliance reader should take away.

The memo also disclaims novelty on the last item. A footnote states that it creates no new process or right, that it restates avenues existing independently of the memorandum, and that it requires only that AI-assisted steps not impede access to them.

The same footnote is already in use at another agency

OPM is not the only federal body relying on this carve-out. A second instance is worth recording, with the limits of what it shows stated alongside it.

The Internal Revenue Manual at 10.24.1, the IRS policy for AI governance, reproduces the same sentence in its glossary exhibit. AI may be integrated in decision or activity pipelines in high-impact categories without meeting the definition of high-impact when the AI's output does not actually serve as a principal basis for the relevant type of agency action or decision.

Two differences are worth recording precisely. OPM attributes the sentence, naming OMB footnote 27. The IRM does not attribute it at all; it sits unsourced in a glossary. And where OMB and OPM write because, the IRM writes when. Whether the IRS intended any change of meaning by that is stated nowhere, and this report does not assume one.

What the two documents do not show is any relationship between them. Neither cites the other. Neither mentions the other agency. There is no shared template named or attached in either. The only documented common ancestor is OMB M-25-21.

The dates matter too, and they cut against the tidier version of this story. The IRM's AI policy text carries a 10 February 2026 stamp; the 10 August 2026 action on that manual section is recorded as superseding the February version, with the declared material change confined to a recordkeeping subsection. So the carve-out was already in the IRM roughly six and a half months before OPM's memo. What the two documents establish between them is that chronology and nothing further. They do not record how either came to use the language, and this report does not fill that gap in either direction.

What to do about it

If you run federal hiring, the actionable object is the determination, not the conclusion. The memo supplies an assessment process and a model determination form and describes that form as offered for convenience rather than as a requirement. Producing one for each workflow is what makes the generally in generally not high-impact defensible.

Write the record so it answers the memo's own question. If the file shows an official who reviewed the underlying documentation and reached a conclusion on it, the test is met. If it shows an official who approved an AI summary, the memo says that can still be a principal basis.

Do not retire the five verification items on the strength of a non-high-impact determination. The memo says in terms that they survive it.

For private employers, this binds nothing. Its value outside government is the test itself, which is a more usable formulation than most of what is available: ask what supplies the justification in the documented workflow, and check whether your record would survive somebody asking.

Related briefings

• AI hiring: four state rulebooks, one workflow
• AI hiring laws
• Canada: federal AI policy and the ADM directive
• North America AI regulation hub
• AI Regulation News hub