An AI Use Policy for RIAs

Why this policy exists

Most advisers have adopted AI faster than they have governed it, and the least-covered control is the one covering AI a firm did not choose to buy. In the 2026 Investment Management Compliance Testing Survey reported by InvestmentNews, 80% of firms had formally adopted AI tools while 30% had policies addressing third-party AI use.

A policy is also the first artefact an examiner asks for. Per Goodwin's summary of the SEC's 2026 priorities, examiners will test the policies and procedures firms have for supervising AI. An absent policy is not a neutral state.

Section 1 · Scope and inventory

The policy covers every system that generates, summarises, classifies or recommends using a model — including AI features inside tools adopted for another purpose.

Maintain a written inventory with, for each entry: the tool, the business purpose, the categories of client data it can access, the vendor, whether outputs reach clients, and the date of last review. The inventory is the document that makes the rest of the policy enforceable; without it, every other section applies to an undefined set.

A practical test for completeness: walk the firm's existing vendor list and ask, for each, whether that vendor shipped an AI feature in the last two years. The inventory that results is usually longer than the one built by asking staff what AI they use.

Section 2 · Permitted and prohibited use

State which data categories may be entered into which tools, and name the prohibited uses explicitly.

Recommended prohibitions to state in writing: entering client personally identifiable information into a consumer AI product not on the inventory; relying on AI output as the sole basis for a recommendation without documented human review; using AI to generate client-facing marketing claims about the firm's own AI capability without compliance review.

That last one is unusual in template policies and belongs in this one. The SEC's AI washing priority makes the firm's own marketing an examinable surface.

Section 3 · Human review before client delivery

No AI-generated output reaches a client without a named human reviewing it, and the review must leave a trace.

Only 37% of surveyed firms had procedures to test or validate AI outputs before client delivery, and 48% had human-in-the-loop oversight policies. The gap between a policy that says a human reviews and a record showing a named human did review is the gap an examination tests.

Specify: who may review, what they are attesting to, how the attestation is recorded, and what happens when the reviewer disagrees with the output.

Section 4 · Records and retention

Treat AI-generated transcripts, summaries and drafts as records unless a documented analysis says otherwise.

Cooley's analysis of Rule 204-2 concludes that the rule's definition of a record captures AI-generated transcripts whether firms realise it or not. A policy that stays silent on retention is making a retention decision by default, and it is the decision hardest to defend.

Address: what is retained, for how long, where, in what format, and how it is produced on request.

Section 5 · Model version and prompt logging

Record which model version produced which output, and when.

The FINRA 2026 Annual Regulatory Oversight Report describes storing prompt and output logs for accountability and troubleshooting, and tracking which model version was used and when. Most firms cannot do this today because most vendors do not expose it. Where a vendor cannot supply model version metadata, record that limitation in the inventory rather than leaving the field blank — an examiner reading a blank cannot distinguish an unasked question from an unanswerable one.

Section 6 · Vendor diligence

Before adoption and annually after, test whether the vendor can meet the recordkeeping obligations the firm carries.

The FINRA report describes simulated regulatory examinations as the mechanism for testing third-party vendors against Rules 17a-3 and 17a-4. A fuller question set is set out in our note on third-party AI vendor due diligence.

Section 7 · Incident response

Define what counts as an AI incident and who is called.

14% of surveyed firms had incident response plans updated for AI — the lowest-covered control measured. AI incidents do not look like security incidents: a model that produces a confidently wrong figure in a client summary is an incident with no breach, no alert and no log entry unless the firm has defined one.

Name the triggers: material factual error in client-facing output, client data entered into a non-approved tool, vendor model change without notice, output that cannot be reproduced.

Section 8 · Review cadence

Review the inventory quarterly and the policy annually, and record the date. Vendor AI features ship faster than annual policy cycles. A quarterly inventory review is the minimum that keeps Section 1 true.