How Internal Audit Can Re-Perform AI Agent KYC and AML Decisions Under the 2026 Model Risk Guidance
Internal audit can re-perform an AI agent's KYC or AML decision only if the agent's inputs, the policy version applied, the policy outcome, every tool call and the final output were captured at the time of the decision and can be retrieved by case or alert identifier. With that evidence, auditors reconcile a population, draw a risk-based and stratified sample, trace each item from the case record to the interaction record, re-perform it and classify exceptions. Evidence produced by policy enforcement in the execution path supports this. Logs reconstructed after the fact often cannot.
The re-performance procedure
- 1
1. Select and retrieve
Take the case or alert record and retrieve the linked interaction, policy decision and tool-call records.
- 2
2. Confirm the policy in force
Verify the policy version applied was the approved version for that date and decision type.
- 3
3. Re-perform
Apply the policy to the recorded inputs and compare the expected outcome with the agent's output and tool calls.
- 4
4. Check escalation and approval
Where the policy required human approval, confirm a named approver and timestamp separate from the agent's actions.
- 5
5. Classify exceptions
Record missing evidence, policy mismatch, unapproved escalation or unexplained output, with the evidence reference.
- 6
6. Document and report
Preserve workpapers, report findings with management responses and track corrective action.
The evidence needed to re-perform one decision
What changed in the model risk baseline
Earlier versions of this procedure were anchored in OCC Bulletin 2011-12. That bulletin is now listed in the OCC's rescinded bulletins. On April 17, 2026, the OCC, the Federal Reserve Board and the FDIC issued updated interagency model risk management guidance (OCC Bulletin 2026-13). As of November 2026, it sets out a risk-based approach and rescinds prior model risk management guidance and other issuances.
The guidance covers model development and use, including testing, validation and monitoring, and governance and controls, including clear policies and roles. It describes generative and agentic AI as novel and rapidly evolving, and the agencies plan a request for information on banks' use of AI. It also states that practices appropriate for one banking organization may be inappropriate for another with a different risk profile or model purpose. For audit, that means scope, depth and sample size should follow the agent's purpose and risk, and audit programs should cite the 2026 guidance as the current reference.
A Federal Register document published April 10, 2026 on AML/CFT programs adds a related expectation. It describes independent program testing that focuses on whether the program is effective and identifies issues for remediation. Re-performance of agent decisions supports that conclusion for one component. It does not replace testing of the wider program.
Defining the population and drawing a defensible sample
Define the population from case or alert records for KYC onboarding and AML alert handling over a stated period. Reconcile the count to the source system's totals before sampling, and document any difference. A population drawn from the agent's own records alone can miss cases that never produced evidence, which is itself a finding.
Examiner practice offers a model for sample design. OCC examination procedures base samples on risk assessment, prior examination reports and audit findings, and include new accounts for customer identification program compliance and customer due diligence information for higher-risk customers. Internal audit can follow the same logic: weight selection toward higher-risk customers and prior findings, and stratify by decision type and by escalation outcome, so that auto-cleared alerts, escalated alerts and blocked actions are each tested. Examiners also review audit reports and workpapers, so record the population, the stratification rationale and the selection method in a form a third party can follow.
Sample size should reflect the agent's purpose and risk profile rather than a fixed number. Whether to use statistical selection, judgmental selection or a combination is an audit methodology decision, and it should be stated and approved before fieldwork.
Judging whether the evidence layer is sufficient
The central distinction is when and how the evidence was created. Evidence produced by policy enforcement in the execution path records what was allowed, redacted, blocked or escalated at the moment it happened. Evidence reconstructed afterward from application logs or reviews shows what the system says occurred, but often cannot show what the agent was permitted to do. Auditors should weigh reliability accordingly and should not rely on evidence they cannot trace to an enforced policy.
- Capture point: Execution-path evidence is created at decision time. Reconstructed evidence depends on what was logged and how it is stitched together later.
- Policy attribution: Execution-path evidence carries the policy version and outcome. Reconstructed evidence usually infers them.
- Retrieval: Test whether records come back by case, alert or decision identifier and can be exported for examiner inquiry within a defined time.
- Durability: Verify tamper-resistance, retention period and data residency directly, not from a vendor statement.
- Scope limits: Confirm what is explicitly out of scope, such as workflow-level tracing or model lineage, so the audit conclusion is not overstated.
Where Trussed fits, and where it does not
Trussed enforces policy in the execution path across models, agents, tools and MCP servers, with outcomes of allow, redact, block, rate limit or escalate for human approval. That enforcement produces continuous audit-ready evidence at the level of individual interactions and tool calls, which maps to the inputs, policy decision and tool-call record described above. It can be self-managed in the customer's cloud, with configurable retention and data residency, which are the properties auditors need to verify.
The scope is deliberately narrow. Trussed does not trace agentic workflows end to end, govern at the transaction level or provide model lineage, so audit conclusions should be limited to the interaction and tool-call level. Trussed is deepest in insurance, a regulated financial sector, and has production experience of 18+ months in a regulated enterprise. This guide does not claim banking or other financial-services deployments. Use the criteria in this guide to evaluate any evidence layer, including this one.
Re-performance trace for one sampled decision
Case or alert record
The sampled item from the source system, used as the join key.
Interaction and policy decision
Inputs received, policy version evaluated, outcome of allow, redact, block or escalate.
Tool-call record
Each call with parameters and returned results.
Output and approvals
The agent's output, plus any escalation and approval by a named person.
Test your evidence layer against a real sample
See how policy enforcement in the execution path produces audit-ready evidence for individual interactions and tool calls, and how it can be retrieved for re-performance.
See How Evidence Is ProducedHow does your AI governance program compare?
See where your program has gaps in less than 2 minutes.