Rule 206(4)-7 Annual Compliance Review: What Evidence to Collect for AI Advisor Assistants
For AI advisor assistants, an annual review under Rule 206(4)-7 should draw on six evidence categories: written policies, documented control points, enforcement records, exceptions and human approvals, retention and residency settings, and control testing results. The test is whether a reviewer can reconstruct what the assistant was allowed to do and what it actually did. Records created by runtime enforcement support that more reliably than records assembled after the fact.
Evidence categories mapped to control points
What the annual review asks of an adviser using AI assistants
As of November 2026, Rule 206(4)-7 requires a registered adviser to adopt and implement written policies and procedures reasonably designed to prevent violations of the Advisers Act and its rules by the adviser and its supervised persons. It also requires a review, no less frequently than annually, of the adequacy of those policies and the effectiveness of their implementation, and a designated chief compliance officer to administer them.
The rule does not list the elements a policy must contain. The adopting release reasoned that advisers are too varied in their operations for one universal set. It also says nothing specific about AI. The evidence categories in this guide are therefore a practitioner mapping onto the rule's two prongs, adequacy and effectiveness of implementation. They are not a regulator-specified list, and this page does not predict what any examiner will request.
The SEC has said the annual review should consider compliance matters that arose during the prior year, changes in the adviser's business activities, and changes in applicable law. Adding an AI assistant to research, client communication or advisory workflows is a change in business activity. The review should show that the policies were updated to reflect it and that the updated policies operated as intended.
Policy on paper versus policy enforced at runtime
The two prongs of the rule call for different kinds of evidence. A written policy, its approval history and its stated scope speak to adequacy. They say little about whether the policy operated on a given Tuesday when an assistant drafted a client summary or called a tool that touched account data.
Evidence assembled after the fact has a structural limit. Prompt histories, vendor dashboards and manual attestations show what was recorded or asserted. They do not necessarily show which policy version applied to a specific interaction, or what the assistant was permitted to do at that moment. Reviewers end up inferring controls from outputs.
When a control is enforced in the execution path, the decision is made and recorded at the same moment the interaction happens. The record is a by-product of the control rather than a reconstruction. That makes it direct evidence of implementation, which is the second prong. It also keeps the record of what was permitted separate from the record of what the assistant produced.
Evaluating evidence quality
Having records is not the same as having usable evidence. Five criteria help separate the two. These are editorial criteria for internal evaluation, not published standards.
Completeness: do records cover every in-scope interaction and tool call, or only a sample the vendor chose to retain? Linkage: can each record be tied to the policy version in force at that time? Without that link, a decision record cannot be tested against a policy. Tamper resistance: can the people being reviewed alter or delete the records? Time coverage: do records span the full review period, including periods after policy changes? Independence: is the evidence generated outside the system under review, or does the assistant's own output serve as proof of its own behavior?
Apply the same criteria to vendor attestations. The SEC has proposed rules on due diligence and monitoring of third-party service providers and models, though the final status of those proposals is not established here. Whatever happens with them, an attestation describes a vendor's position, and the adviser's review concerns the adviser's own implementation. Attestations can supplement internal records but should not replace them.
What runtime enforcement produces, and what it does not close
Trussed enforces policy in the execution path at the level of individual interactions and tool calls: allow, redact, block, rate limit, or escalate for human approval. Each decision generates audit-ready evidence as it is made. Trussed can be deployed self-managed in the customer's cloud, with configurable retention and data residency, so the adviser can set those values to match its own retention analysis. That analysis, including any Rule 204-2 obligations, remains a determination for the adviser and its counsel.
Enforcement records address several checklist items directly: decisions, redaction and block events, escalations and approvals, and the policy applied. They do not address everything. Whether the policy is adequately designed is a judgment the review itself must make. Model behavior outside the enforcement points, client disclosures, and the reviewer's conclusions need separate evidence. Records at the level of interactions and tool calls also do not amount to transaction-level governance, model lineage or end-to-end tracing of agentic workflows, and the review should not describe them that way.
A useful framing for the CCO: enforcement evidence strengthens the effectiveness prong, while adequacy, disclosure and judgment remain documented human work.
Running the review: a practical sequence
The steps below fit within a normal annual review cycle. Documenting the review in writing is sound practice, and counsel should confirm the current documentation requirements that apply to your firm.
- Inventory every AI assistant use case, including research and client-facing workflows, and tie each to a written policy.
- Confirm that AI-related policy changes made during the year are documented and carried into the review.
- List current evidence sources, such as vendor dashboards, prompt histories and attestations, and mark which cannot show the applied policy or the decision.
- Sample interactions and tool calls across the period, and check that each record shows the policy version, the decision and the outcome.
- Review escalation and exception records for approver, decision and outcome, and assign ownership of AI evidence to the CCO function.
From policy to reviewable evidence
Written policy
Versioned rules for what the assistant may do.
Control point
Where the policy is applied to an interaction or tool call.
Decision record
Allow, redact, block or escalate, tied to the policy version.
Annual review
Sampled records tested against adequacy and effectiveness.
Check whether your records could reconstruct what the AI did
Walk through how policy enforcement in the execution path produces audit-ready evidence for sampled interactions and tool calls.
See How Evidence Is Produced