An AI agent that prepares entries, matches transactions or releases payments is part of your internal control over financial reporting. The control framework does not change; what changes is who (or what) performs the steps, and how you evidence them.
What auditors and regulators will look for
- Scope: which processes and key controls involve AI.
- Design: is the control still effective with an AI performing or supporting it?
- Operating evidence: logs of what the agent did, on what inputs, and the human review.
- IT general controls: access, change management (model and prompt changes), operations.
- Segregation of duties: an agent's permissions must not combine incompatible duties.
Design patterns that work
- AI prepares, human approves for anything that posts to the ledger or moves cash.
- Thresholds: agents act alone only under an amount or risk limit; above it, human approval.
- Change control: treat prompt, model and configuration changes like code changes, with testing and approval.
- Monitoring: sample agent outputs; track override and error rates; alert on unusual volume.
- Kill switch: a tested way to stop the agent and fall back to the manual process.
A pre-AI lesson about automation at speed
Automated systems without adequate pre-trade controls can do enormous damage in minutes. The SEC found that on August 1, 2012, Knight Capital's automated router sent more than 4 million orders in 45 minutes while trying to fill 212 customer orders, losing more than $460 million; the SEC cited inadequate controls, deficient code deployment and testing, and weak incident response. Agents are not trading systems, but the control lessons — limits, testing, monitoring, a kill switch — transfer directly.