What an AI Policy Analyst Delivers, and When to Make the Hire

What Does an AI Policy Analyst Hand You in the First Ninety Days?

The deputy counsel puts it plainly, halfway through a Tuesday budget meeting: which of the tools staff already use send resident data to a model, and who approved them. The room goes quiet. Somebody offers that the call center bought a transcription add-on last spring. Nobody knows what it does with the recordings. The first deliverable of an AI policy analyst is the end of that silence, written down.

It takes the form of an inventory naming each system, its owner, the data it touches, the decision it influences, and whether a law already covers it. Three more deliverables follow within the quarter. A short approval rule that a caseworker can read once and apply: these tools for these data classes, this route for anything else, disclosure required here. An impact assessment for each system that factors into a consequential decision about a resident, including the ones where the honest conclusion is "do not deploy." And a tracked list of the statutes, executive orders and directives that bind the agency, with the dates they take effect.

None of this is invented from scratch. Legislatures have already written the shape of the job. In 2025 every state introduced AI legislation and thirty-eight adopted around a hundred measures, including New York requiring agencies to publish an inventory of their automated decision-making tools, Arkansas requiring public entities to adopt an authorized-use policy, and Alaska requiring a prioritized plan of AI project uses and costs ¹. Colorado's rules for deployers of high-risk systems, effective February 1, 2026, ask for a risk management policy, impact assessments, annual review for discriminatory outcomes, notice to the person affected, and an appeal to human review ². Law here is jurisdiction-specific and moving; check with counsel before treating any of it as advice.

So the trigger for the hire is not enthusiasm about AI. It is one of three facts: a law names your agency, procurement is asking for contract language nobody has written, or your use-case count has passed what a volunteer can hold alongside their real job.

Test the Candidate on the Transcription Add-On

Hand a candidate the call center's transcription add-on and watch what they ask first. The one you want asks about the data: retention period, who else can read the recordings, whether the vendor trains on them. Then who is developer and who is deployer in your setup, because the duty follows that answer ². A candidate who opens on model architecture is interested in the wrong half of the job.

The second thing to listen for is a failure mode named before a framework. Ask what goes wrong with an eligibility screening tool and wait for something concrete, a proxy variable or a fallback that quietly denies, rather than "bias and transparency concerns." The ones who have done this work reach for a specific case, because they have lived through one.

Ask, too, about a time they recommended against a system and what happened next. The whole value of an assessment rests on it sometimes concluding no, and an analyst who has never written that sentence has not yet been tested. Then hand them a real vendor page and ask for the two-paragraph rule you would send to three hundred staff. Legal fluency that cannot be compressed to a paragraph will not change anybody's behavior, and the strong ones write for the caseworker who has to comply rather than for the lawyer who reviews the memo.

The boundary with the compliance officer is worth stating out loud in the interview. Compliance asks whether a control was followed. This role decides what the control should be for a technology with no settled practice, which is a drafting job as much as an audit one. If your agency already has a senior owner for that, you may be hiring one level up instead: see what an AI governance lead owns.

Why a Records Officer Outperforms a Certificate

The person who eventually answered the deputy counsel's question had been the agency's records officer for eleven years. The obvious background is a policy degree plus a privacy certification; the unexpected ones often work better. A procurement analyst who reads contracts for a living, a legislative staffer who has tracked bills through committee, an accessibility coordinator. Each already does the core motion, which is turning a legal duty into a rule a busy person can follow.

What separates the strong ones is practice with the technology rather than opinions about it. Ask what they have built with an AI assistant and what it got wrong. The answers that mean something are specific: drafting a bill summary with a model and then reading the enrolled text to find the two clauses it flattened; trying to make an agency chatbot answer a benefits question incorrectly and documenting the prompt that did it; keeping a folder of model cards and system cards so a vendor claim can be checked against the developer's own disclosure.

That habit matters because most of the job is judging confident text. A model will produce a fluent paragraph asserting that a system falls outside a statute, and the analyst's contribution is knowing which sentence in that paragraph needs a source and going to get it. Candidates who use AI heavily and check it constantly do this well. Candidates who have never used it, and candidates who trust it, both fail in the same place.

They also need a working relationship with the people who run the systems. The best analysts spend their first month in program offices rather than in the policy library, and they arrive with enough technical vocabulary to talk to the agency IT staff who actually administer the tools without an interpreter.

Look Inside the Building Before Running a Search

That records officer was already on the payroll, which is the usual case. Privacy program managers, procurement analysts and legislative liaisons hold the relationships and the institutional memory that take an outside hire a year to build, and many of them have been doing a version of this work uncompensated since the first chatbot pilot. Promoting one and backfilling the vacated seat is often faster and cheaper than an external search.

It also pairs well with a broader effort to build AI capability across existing staff. If you do go outside, post where people already read statutes for a living. Privacy professional associations and their AI governance certification cohorts, state CIO and CISO working groups, public-sector privacy convenings, and graduate policy and information-school programs all concentrate the right background. A generic job board returns people who have read about AI governance; these venues return people who have written a policy someone had to follow.

Look also at the public-sector practices of large consultancies, at civic technology organizations whose alumni have delivered inside government, and at legislative staff who worked on an AI bill and are ready to implement one. Ask every candidate for the policy document they are proudest of and read it before the interview. This is a writing job, and writing samples are abundant and honest.

One sourcing warning: certification counts have grown quickly and a credential alone now separates almost nobody. Screen on artifacts, not on letters after a name.

How Do You Close an AI Policy Analyst, and Where Does the Work Sit?

Close on authority first and pay second, because the authority is what they cannot get elsewhere. As of mid-2026 no published wage series covers this title in government. The number you are bidding against is the private market: one 2026 report triangulated mid-career US AI governance pay at roughly 140,000 to 218,000 dollars from job boards, posting samples and recruiters, with data collected between December 2025 and May 2026 ³. Public salary bands rarely reach that.

What you can offer instead is scope that a private compliance seat does not carry: the inventory is published, the assessments are read by an oversight body, and the rule the analyst writes governs a whole agency rather than one product line. Name the executive sponsor in the offer conversation. Give the role a standing seat in procurement review and the explicit authority to say no, subject to appeal. Fund a bill-tracking subscription and conference travel, both small numbers that read as seriousness.

What kills the offer is predictable. The role buried three levels down with no route to a decision-maker. A job description that turns out to mean approving software licenses. And a hiring process that takes four months, in a market where demand for AI governance skills grew about 150 percent year over year by LinkedIn's own 2026 accounting ³.

On location: the drafting, tracking and assessment work travels fine, and most postings for this title allow remote or hybrid arrangements. The parts that do not travel are the ones that make the inventory real, which are the interviews with program staff, the vendor demonstrations, and occasional testimony at the capitol. Many state roles also require in-state residency, and systems handling criminal justice or federal tax data may require review on-site under their own access rules. A workable norm is remote with monthly on-site weeks, agreed in writing before the start date.

Read the evidence >