US public sector internal auditors need training to audit AI agents in benefit eligibility checks, improper payment reviews and grant monitoring. They must test verdicts, approvals, certificates and provenance from the record, and check agency AI inventories and risk practices. Audit conclusions stay with them. Federal AI policy, the NIST framework, HIPAA where relevant and state laws frame oversight.
What do public sector internal auditors in the US work on?
The picture varies by level. Federal agencies have inspectors general and internal control teams; states have auditors and internal audit units in large agencies; cities and counties have their own auditors, often reporting to an elected official or an audit committee. Most follow government auditing standards alongside professional internal audit standards.
Common themes run across levels. Improper payments in benefit programmes such as unemployment insurance and health coverage draw sustained attention, as do grant oversight, procurement, payroll and IT security. Audit reports are usually public, so findings must be precise and evidence must hold up to outside reading.
Where do agents enter the work?
Agencies are introducing agents where caseloads are heavy and rules are written down. Each capability carries a depth earned on the agency's own past cases.
- Eligibility document review (depth 1). The agent checks submitted documents against programme rules and lists what is missing or inconsistent. A caseworker decides.
- Improper payment flagging (depth 1). It compares payments with wage, death and residency records and flags likely errors for review.
- Grant reporting review (depth 2). It reads grantee reports, compares spending with the award, and drafts follow-up questions. A grants officer sends.
- Notice drafting (depth 2). It drafts notices to claimants with the reasons for a decision taken by a person.
Audit can use agents too, at depth 1, to test full populations from the append-only record, such as every overpayment flag and what happened to it.
What must auditors learn, and what stays with them?
Begin with the fundamentals in our guide to what internal auditors should learn before working with AI agents: depth, the five bounds of the minimum chain, the four verdicts and the record.
The government-specific test is whether a person decided where a person must. Eligibility determinations and benefit reductions affect people's rights and are subject to due process. An auditor should show from the record that the agent stayed at depth 1 or 2, that each determination was made by a caseworker, and that the notice gave the real reasons. A notice that recites the agent's flag without the caseworker's reasoning is a finding.
Provenance is the second skill. A wage record from the state's system is sealed. A claimant's statement is declared. A tip received by email is external, and an external claim cannot by itself reduce a benefit. Auditors should test that overpayment actions rested on sealed or signed data.
Certification is third. Agent certificates lapse after 90 days, and a run that clears no depth drops the agent to depth 1 the same day. Auditors test certificate dates against action dates. Those who want to understand review from the inside can take operator-level training and record practice in a practicum log template for internal auditors.
What stays with auditors: scope, conclusions, recommendations and reporting. Our piece on whether AI agents will replace internal auditors explains why independence and judgement remain human.
Which US rules shape oversight and literacy?
There is no single federal AI law. Federal policy requires agencies to inventory their AI use cases, name accountable officials, and apply minimum risk management practices to uses that significantly affect people's rights or safety, including testing, monitoring and human review. Auditors should test that each agent appears in the inventory, is classified correctly, and has the required practices in place.
The NIST AI Risk Management Framework, voluntary, with its Govern, Map, Measure and Manage functions, is a natural spine for an audit programme at any level of government. HIPAA applies where an agency handles protected health information as a covered entity, as some health programmes do. Federal improper payment requirements shape audit attention on payment integrity. At state level, Colorado has passed an AI act aimed at high-risk systems that make consequential decisions, and auditors should confirm which state laws apply to their body.
Public bodies also carry transparency and fairness duties, and freedom of information laws may reach the agent's record.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
A 90-day training plan
Days 1 to 30. All auditors learn the vocabulary. The audit director obtains the agency's AI inventory and maps each agent to its capabilities, depth and certificate dates.
Days 31 to 60. Pilot an audit of improper payment flagging. Test every flag in a period: provenance of the data, who acted, what the notice said, and certificate currency.
Days 61 to 90. Write the standing audit programme mapped to the NIST functions and federal minimum practices, and set up banking corrections for audit's agent.
Where the Academy fits
Pryme Intelligence Academy's Foundations track is free and takes about four hours, with lessons available as audio. Auditors leading agent audits can take the Operator track; its certificates are issued through VerifyDoc.ai and anyone can check them at source, which helps when an oversight committee asks about reviewer competence.
Questions people ask
Does due process apply when an agent flags an overpayment?
Due process attaches to the decision to reduce or recover benefits, which a person should make with notice and an opportunity to respond. The agent's flag is input, not the decision.
Are inspectors general expected to audit AI use?
Where agencies use AI in programmes, oversight bodies can be expected to examine it. Inventories and records make that practical.
How often should audit check the agency's AI inventory?
Whenever the inventory is refreshed and whenever an agent's depth or use changes. An inventory that lags behind real use is a finding in itself. Our piece on why internal auditors are the real control and the guide to AI agent skills by role cover the wider picture.