The S3 Brief · Issue 15 · AI Governance

Your Agentic AI Audit Evidence Is Missing. Here's What Regulators Expect.

Regulators now expect documented agent identity, intervention points, and immutable audit trails. Most lean organizations have none.

Agents are autonomous. Auditors are skeptical. Your audit evidence isn't ready.

In October 2026, banking examiners began demanding decision logs, approval gates, and named accountability owners for every deployed AI agent. Teams that could not produce this evidence on demand faced supervisory findings. The same month, the UK Information Commissioner's Office opened a formal call for evidence on agentic AI governance, signaling that regulators are turning their attention from models to the systems that act.

Your existing AI governance program was built for chatbots and inference pipelines. Agents are different. They take actions without human approval between deployments. They have memory. They call tools. They fail differently. Your audit trail and your oversight structure both need to change.

Start with agent identity and access.

Every agent that takes an action needs a unique, auditable identity. Not a generic service account. Not a shared token. A documented identity with an approval date, named owner, and a record of who authorized it to do what. If you cannot say "Agent X was approved on [date] by [person] to [action] using [tools]," you have no audit trail.

When an examiner asks you to produce evidence that an agent had the right to execute a specific action, screenshots and policy memos do not satisfy the requirement. You need an immutable log entry. That log must capture the agent's identity, the action it took, the timestamp, the tool it used, any data it accessed, and any human intervention point. The EU AI Act requires high-risk AI system logs to be retained for at least six months. Your logs need hash-chain integrity (SHA-256 minimum) to prove no one altered them after the fact.

Second: document intervention points.

Agents are evaluated on their autonomy. Regulators evaluate them on your ability to stop them. Before an agent goes into production, map where a human can pause it, review its decision, override it, or escalate it. Write this down. Assign specific job functions (not titles, not teams) authority to intervene. Define what triggers escalation. Define the bounds on iteration: how many steps can an agent take before requiring human review?

Federal Reserve guidance published in February 2026 emphasized that even in financial services, where model risk management has decades of institutional weight, examiners found that autonomy thresholds drifted upward over time without anyone making an explicit decision to raise them. Authorization creep is real. Document your starting position now.

Third: separate tool authorization from resource authorization.

Cloud Security Alliance guidance published in 2026 calls this the two-gate model. An agent needs permission to call a tool (tool authorization). It also needs permission to access a specific resource through that tool (resource authorization). If you grant an agent blanket API key access without per-resource gates, you have one gate. Auditors will flag it. Implement separate controls for each gate and document both.

Fourth: set spending limits for agents with financial authority.

If your agent can execute a payment, initiate a transfer, or modify a record that has business impact, regulators expect documented spending limits, explicit pre-deployment behavioral testing, and a model card stating which foundation model powers the agent. Behavioral testing means you run the agent through scenarios that exercise its decision-making, log its outputs, and verify it does not exceed its authority. Write the results down. Keep them.

Fifth: audit trail architecture.

Audit-ready logging is not optional. Screenshots are not logs. Declarations of intent are not evidence. You need continuous, real-time log streaming to a centralized platform (a SIEM or equivalent) that supports immutable, tamper-evident retention. Every agent action entry must include the agent identity, the action, the timestamp, the input data, the output, any human decision, and any system policy that governed the decision.

The Financial Services AI Risk Management Framework released in February 2026 includes 230 control objectives mapped across the AI lifecycle. Three of them focus on third-party and supply chain AI risk. When your audit evidence flows through a third-party logging or monitoring service, you carry responsibility for that service's integrity. Document the SLA. Verify the hash chains.

Sixth: align to frameworks auditors already know.

If you operate under SOC 2, auditors will look for agent controls under CC6 (logical and physical access controls), CC7 (system monitoring), and CC8 (change management). Map your agent governance to these categories. If you report under ISO 27001, ISO 42001 is built to layer on top of it. If you follow NIST AI RMF, your GOVERN and MANAGE functions apply directly to agents: continuous monitoring and escalation protocols are not optional.

Start this week.

Pick one deployed agent. Collect its identity documentation. Find its logs. Verify the log retention window and integrity. Map the intervention points in production. Check if tool authorization and resource authorization are separate. If you use a third party for agent deployment or logging, review that vendor contract for audit liability.

Then repeat for every other agent.

Regulators are asking these questions now. Auditors will ask them during your next examination or third-party assessment. The difference between a governance pass and a finding is evidence.

Want a template for agent governance controls and audit evidence requirements mapped to SOC 2, ISO 27001, NIST AI RMF, and the EU AI Act? Start with the AI Security Checklist for a baseline inventory of agent deployments and vendor relationships: https://s3advisorygroup.com/resources/ai-security-checklist

Sources

More past issues

All issues