Who Operates the AI Agents Your Agency Already Put in Front of Constituents?

Your Benefits Agent Told Nine Hundred People the Wrong Filing Deadline

It surfaced on a Tuesday, in a caseworker's note, three weeks after the agent started saying it. The deadline had moved in a policy bulletin the agent never ingested. Throughput was up, deflection was up, the vendor dashboard was green, and nobody in the building could say who was supposed to have caught it. That gap is the job you are hiring for.

The uncomfortable part is how ordinary the failure is. A constituent-facing agent does not usually break loudly. It answers, in a steady voice, using last quarter's policy, and it does that at a volume no phone bank could reach. Route Fifty's coverage of agents in service delivery frames the arrival of agentic systems as a change to how government work gets done rather than a new channel bolted onto the old one ¹. A change to how work gets done needs someone whose job is the work, not the tool.

So the first thing to write into the posting is authority, before duties. This person can suspend a production agent without convening anyone. They can change a guardrail between shifts. They can send a class of intents back to human handling and defend that in front of the program director. If your draft posting says "monitors," "reports on," or "partners with the vendor to," you are describing a seat next to the decision rather than the decision.

Second, name the surface. Which agents, which programs, which populations, and what happens to a case the agent hands back. An operations manager with three named agents and a documented escalation path will do more in a month than one given a portfolio described as "AI initiatives."

Which Backgrounds Produce an Agent Operations Manager Who Can Read a Queue?

The strongest candidates have run a queue where being wrong had consequences for a person. Call center and eligibility supervisors, records managers, unemployment insurance adjudication leads, 911 and dispatch supervisors, benefits appeals staff. They already think in escalation paths, sampling, and what a bad hour looks like before the metrics move. What they need to add is agent-specific: reading a trace, writing a guardrail, and knowing when a model's confidence means nothing.

The unexpected sources are worth naming, because agencies overlook them. Library reference desk managers spend their careers deciding when a question exceeds what a source can answer. Clinical charge nurses run a floor of skilled people and pull someone off an assignment without drama. Air traffic and utility control room supervisors are trained to distrust a normal-looking board. Freedom of information officers know exactly what a log has to contain to survive a request, which is knowledge most engineering candidates simply do not have.

The background that misleads most often is pure vendor implementation. Someone who configured the platform knows the console, and console fluency reads as expertise in an interview. Ask them to describe a decision they made against the vendor's recommendation and what it cost. If the answer is a feature list, keep looking.

Separate this hire from adjacent ones before you post. An AI system auditor reviews systems from outside the program and should not report to the person running them. A guardian agent engineer builds the controls. Your operations manager works the live floor and is accountable for what constituents were told today.

Ask How They Used AI on Their Own Casework Before They Ever Supervised It

The people who are good at this got good by using these systems on work they personally owned, and getting burned in a way they can still describe. That story is the interview. Ask for one specific case where a model gave them a confident answer they acted on, how they found out it was wrong, and what they changed afterward. Real answers have a date, a document, and a small amount of embarrassment in them.

The tells separate quickly. A candidate with practice will tell you what they stopped delegating. They drafted correspondence with an assistant but kept the eligibility determination, because the determination is the part a constituent can appeal. They will describe checking a claim against something outside the conversation, a statute, a bulletin, a colleague, rather than asking the model whether it was sure. Someone performing fluency describes prompts and tools; someone with judgment describes a boundary they drew and why.

Then make them work in front of you. Give a real transcript from one of your agents, forty minutes, and ask for three things: what went wrong, what class of failure it belongs to, and what guardrail would have prevented it without breaking the good cases. Watching a candidate look for the counterexample to their own first theory tells you more than any answer they can prepare for.

One screening rule that saves time later. Do not ask whether a writing sample was AI-assisted; you cannot tell, and the question teaches candidates to hide the thing you actually want to see. Ask them to work with an assistant in front of you instead, and grade the moments where they checked, overrode, or declined.

Find This Person Inside the Program Office, Not on the Vendor Bench

Look internally first, because the scarce ingredient is program knowledge and the agent knowledge is teachable in a quarter. The best hire is often the supervisor already spending afternoons cleaning up after the pilot. Post the role at a level that makes an internal move a promotion rather than a lateral favor, and let it carry the authority described above, or your candidate will decline and stay where the ladder is clear.

Outside the building, the venues that actually work are the ones where practitioners talk about production. The Code for America Summit and its brigade network, NASCIO and NASWA annual conferences for state officials, the Beeck Center's digital service work, GovAI Coalition materials out of San Jose, state digital service teams and the alumni of federal and state digital service corps. Those alumni networks matter because the people in them have already shipped inside procurement and privacy constraints and will not be surprised by either.

Closing is where agencies lose these candidates, and the reasons are consistent. They care about scope they can name, a reporting line into the program rather than into a central innovation office, and a written statement that they can take an agent offline. They ask what happens the first time they do it. If your honest answer is that a deputy director would reverse them, they will hear it in the pause.

What kills the offer: a nine-week hiring timeline against a private offer that closes in ten days, a title that reads as coordinator, and a scope that turns out to be writing the AI policy memo. Analysts who want to write policy exist. This is not that job. Related but distinct is an AI output verification counsel, who reviews what the agency can defend having said.

What Does This Role Cost, and Should It Sit On-Site or Remote?

No published salary survey covers this title in the public sector yet, so any specific number you see quoted for it is someone's guess. Price it by the classification you already have. Most agencies land it where a program supervisor or a senior IT manager sits, one band above the queue supervisors it draws from, because the authority to take a service offline is supervisory authority.

The honest planning note is that private demand for the same skills is rising while agency bands move slowly: Gartner projects task-specific agents in 40 percent of enterprise applications by the end of 2026, up from under 5 percent in 2025 ². Expect to compete on scope, mission and stability rather than on cash, and expect turnover if the role is graded as coordination work.

On location, the work is hybrid by nature. Reading transcripts, tuning guardrails, and auditing logs are remote-friendly. The parts that are not: sitting with caseworkers during the first weeks of a deployment, being in the room when an agent is pulled, and the periodic sampling that goes faster when the person who knows the policy is at the next desk. Two to three days on-site during any active rollout, mostly remote between them, is the pattern that holds.

One structural decision belongs in the posting. Whoever holds this role needs credentialed access to agent logs, identity records, and the systems the agents act inside, which is its own governance question and increasingly its own hire, the non-human identity security manager. Guidance aimed at public sector governance frames agentic deployment as a set of new oversight roles and training obligations rather than a tooling purchase ³. Budget the training time. Six months of program experience plus a quarter of agent-specific coaching produces a better operator than any external hire you can close this fiscal year.

See the benchmarks >