AI agents and the compliance audit trail
As AI agents act on mandates, every action needs provable authority. What a trust layer for the agent economy looks like.
- AI agents acting on wealth data need scope-bound, time-limited authority.
- Every agent action must be auditable against the authority it ran under.
- Moving funds should be outside an agent's scope by design.
The new risk
As wealth management becomes agent-driven, an AI agent that reads a portfolio or drafts a report is acting on sensitive data. Without provable authority, there is no audit trail and no accountability.
A trust layer
The answer is explicit, scope-bound, time-limited mandates for agents — read portfolio, draft report, never move funds — with every action recorded against the authority it ran under. This is auditability built for the agent economy.
What an audit trail for AI agents must show
When software agents prepare documents, propose transactions or extract data from statements, the audit trail has to answer the same questions as for a human employee: who or what acted, on whose authority, with which inputs, and what was the result. In practice this means logging each agent action with its mandate, the data it used, the output it produced and the human who approved or rejected it.
Agents should work with the least privilege necessary and never approve their own proposals where a four-eyes rule applies.
Keeping humans accountable
AI can accelerate routine work — classifying documents, drafting summaries, flagging inconsistencies — but responsibility stays with people. Define which steps require human confirmation, make the agent's reasoning or sources visible to the reviewer, and record the decision. A clear separation between suggestion and decision is what allows a family office to use automation without weakening its control framework.
Practical checklist
Inventory where AI agents act in your processes and what each may do: read data, draft documents, propose transactions, contact third parties. Give each agent the minimum permissions it needs and a clear mandate. Log every action with inputs, outputs and the approving person.
Define the steps that always require human confirmation, especially payments, client communication and changes to beneficial-ownership data. Review samples of agent output regularly for errors and bias, and document how the agents' results are used in decisions. Treat the agent configuration itself as a controlled document with versioning.
Common mistakes
A frequent mistake is giving agents broad access 'to make them useful' and planning to restrict it later. In practice the restriction rarely happens. Start with the minimum permissions and expand deliberately. Another mistake is logging only the final result: without inputs and intermediate steps, nobody can reconstruct why an agent proposed a particular action.
Teams also tend to let reviewers approve agent output without seeing the sources. Approval then becomes a formality. Make the evidence visible and record what the reviewer actually checked. Finally, agent configurations change — prompts, models, tools — and each change can alter behaviour. Treat configuration changes like code changes: versioned, reviewed and documented.
Example from practice
A family office lets an AI agent extract positions from monthly bank statements. The agent proposes the bookings, flags two unknown instruments and logs its sources. A team member reviews the flagged items, confirms the rest and approves. When an auditor later asks how a figure was produced, the log shows the original statement, the agent's proposal and the person who approved it.
FAQ
- RFC 9396 rich authorization requests
- NIST AI RMF
Related guides
Compliance, built into the platform
FLIORE is a family office platform with KYC and compliance built into the core.
Explore FLIORE