Procurement files, social programme beneficiary lists and SPEI payment batches are where Mexican public sector auditors need to audit AI agents. Training should cover handling held actions, spotting manipulated supplier documents and testing approval authority from the record. Audit conclusions stay with people. Transparency duties, public-sector data protection law and SPEI controls frame oversight.
What do internal auditors in Mexican government focus on?
Each federal agency and many state bodies have an internal control unit that audits operations, investigates irregularities and reports on compliance. The heaviest areas are procurement, from tender to contract to payment; social programmes that pay transfers to large beneficiary lists; payroll; and public works. The federal supreme audit institution reviews public accounts from outside, and internal auditors prepare the ground.
Payments to suppliers and beneficiaries move through SPEI, which settles quickly. Auditors test who authorised each batch, whether the beneficiary was eligible and alive, whether a supplier's bank account changed shortly before payment, and whether the contract was awarded by the right procedure.
Where do agents enter the work?
Public bodies are adopting agents where documents pile up. Each runs at a depth the body has proven on its own cases.
- Tender file review (depth 1). The agent checks bids against the published terms and lists missing documents or inconsistencies.
- Beneficiary list validation (depth 1). It compares a programme's payment list with the body's register and flags duplicates or ineligible records.
- SPEI batch preparation (depth 2). It assembles the payment batch. A named official reviews and releases it.
- Transparency request drafting (depth 2). It finds responsive documents and drafts the reply. An official decides and sends.
Audit can use agents too, at depth 1, to test full populations, such as every payment batch in a quarter against the approval record.
What must auditors learn, and what stays with them?
The first lesson is the held action. When an agent prepares a payment batch and stops, the verdict names the person it is waiting for and the bound that held it. Auditors need to know how to read that and what to test. Our guide on how internal auditors handle an action an AI agent has held explains the trail: who received the held item, how long it waited, who released it and on what basis.
The second lesson is manipulation. Bids and supplier letters arrive from outside, and a document can contain hidden text written to steer an agent, such as an instruction to mark the bid compliant. Agents governed by provenance treat that text as external, and an external claim cannot authorise anything. Auditors should still learn to recognise attempts in the record. Our guide on training internal auditors to spot prompt injection and pressure gives exercises.
The third is authority. An approval is only valid if the person had delegated authority for that amount and type. An approver authority matrix for the internal audit team shows the format, and auditors can test the business's own matrix against the record.
Red lines round this out. Our list of what internal auditors should never let an AI agent do includes awarding a contract and releasing a payment to a changed bank account without a held verdict.
What stays with the auditor: the plan, findings, conclusions and any referral of irregularities. These cannot be delegated.
Which Mexican rules shape oversight and literacy?
Mexico has no single AI law. Public bodies are bound by transparency and access to information duties, and by a separate personal data protection law for the public sector that governs how beneficiary and employee data may be used. The supervisory arrangements for both have been restructured, so auditors should confirm which authority now oversees their body. Public bodies also carry fairness duties in how they treat applicants and beneficiaries.
SPEI is operated by Banxico. Audit's interest is the body's control over what it sends: release authority, bank detail changes and error handling. The CNBV and CONDUSEF supervise financial institutions and their customers and do not govern a ministry's internal audit.
Cross-border references help structure an AI audit: the NIST AI Risk Management Framework, ISO/IEC 42001, and the OWASP Top 10 for LLM Applications, which includes prompt injection and excessive agency.
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 of governed agents: depth, the five bounds, the four verdicts and the record. Build an AI agent competency matrix for the internal audit team to show who needs which level.
Days 31 to 60. Pilot an audit of SPEI batch preparation. Test every held batch, who released it, bank detail changes in the prior 30 days, and certificate currency. Run a prompt injection exercise using sample bids.
Days 61 to 90. Write the standing audit programme for agents, add the injection test to procurement audits, and brief management on findings.
Where the Academy fits
Pryme Intelligence Academy's Foundations track gives the whole unit a common vocabulary in about four hours, free and stamped with the product version. Auditors who lead agent audits can take the Operator track, whose module on when something goes wrong covers held actions and incidents. Certificates are listed on a public register with their expiry.
Questions people ask
Can an agent release a SPEI batch to beneficiaries?
That would be depth 4, which must be earned and stays bounded by the body's own limits. Most public bodies keep release with a named official, and audit should confirm it.
How would audit spot a manipulated bid?
By sampling refused and held verdicts in the record where the agent flagged instructions inside an external document, then checking how officials responded.
What should a first finding look like?
Specific and evidenced from the record: which action, which bound, who approved, and why it matters. The guide to AI agent skills by role shows how audit fits with other roles.