The Chief Compliance Officer's Guide to AI Agents

Compliance officers meet agentic AI in an awkward position: asked to approve systems they did not design, held responsible if those systems breach a rule, and simultaneously offered agents that could transform compliance monitoring. This guide covers both directions: the compliance program for agents, and agents in the compliance program.

How should the compliance program extend to agents?

By extending, not duplicating. The existing program has policy, training, monitoring, third-party diligence, investigation, and reporting. Each needs an agent-specific addition:

Program element	Agent-specific addition
Policy	Permitted and prohibited uses; data rules; human-decision requirements; approval route; incident duties
Training	Specific to deployed agents: what to check, what not to input, how to report
Monitoring	Agent behavior monitored like employee conduct: what they did, exceptions, incidents
Third-party diligence	AI vendors and AI embedded in existing vendor products
Investigation	Traces as evidence; agent incidents in the case management process
Reporting	Agent inventory, incidents, and control status in compliance reporting

The executive guide to AI agent governance describes the governance structure compliance plugs into.

What makes an AI policy enforceable?

Concreteness. A policy saying "use AI responsibly" enforces nothing. A policy that states which data classes may never be sent to external models, which actions always require a human decision, which uses are prohibited outright, and what the approval route is can be enforced through permissions, gateway rules, and approval gates. Write the policy so that an engineer can implement it and an auditor can test it. The how to set AI risk appetite guide covers writing rules in enforceable terms.

Why does generic AI training fail?

Because it is disconnected from what people do. An employee who completes a module on AI ethics and returns to a job where an agent now drafts their customer emails has learned nothing useful. Training that works is specific: here is the agent you will use, here is what it does and does not do, here is what you must check before sending, here is what you must never paste into it, here is how to report it getting something wrong, and here are the decisions that remain yours. Tie training to deployments and measure behavior, not completion.

What third-party risk is missed?

AI embedded in products the company already buys. A CRM adds an AI assistant; a helpdesk tool adds automated responses; a payroll system adds anomaly detection. None goes through AI procurement because none is an AI purchase. Compliance should add AI questions to routine vendor reviews and renewals: what AI features are present, what data do they process, are they on by default, can they be disabled, what are the training and retention terms, and how are model changes notified. The how to evaluate AI vendors guide covers the assessment.

How do agents change compliance monitoring?

They change what is feasible. Sampling exists because reviewing everything was impossible; agents make population testing practical. Transaction monitoring against policy, communications review, expense and gift checks, conflict of interest screening, and control evidence gathering can run continuously rather than periodically, with compliance officers reviewing the exceptions the agent surfaces.

Dispositions, escalations, regulatory judgments, and anything with consequence for an individual remain with compliance officers who are accountable and can explain the decision. The digital FTE for compliance monitoring guide covers the build.

How does compliance stay independent?

The same way internal audit does: by not owning the agents it monitors, specifying its own tools rather than inheriting business-built ones, documenting its agent-assisted procedures so they can be reviewed, and having the reliability of its tools assessed by someone other than their daily user. Where compliance uses an agent to test a control the business also automated, the independence question should be explicit.

What records evidence the program?

The inventory with risk tiers and owners; policy versions and acknowledgements; training records tied to deployments; monitoring outputs and exception dispositions; third-party assessments including embedded AI; incident records and remediation; and reporting to the board or committee. Regulators assessing a compliance program will look for the same structure they expect elsewhere, applied to agents. Obligations vary by sector and jurisdiction; this is general guidance, not legal advice.

What should chief compliance officers ask?

• Is our AI policy concrete enough that permissions could enforce it?
• Does our training describe the agents people actually use?
• Which vendor products have added AI features since we last reviewed them?
• What agent incidents occurred this period, and how were they handled?
• Where compliance uses agents, who assessed their reliability?

How can FISTA Solutions help compliance functions?

FISTA Solutions builds AI agents whose permissions and gates enforce written policy, with traces that serve investigations and evidence that supports program reporting, and works with compliance leaders through its AI enablement practice to extend policy, training, and third-party diligence to agents. Since 2017, FISTA has delivered 150+ projects for 50+ companies across 12+ countries.

To make your AI policy enforceable in software, talk to FISTA on WhatsApp, or read the AI policy template.