FINRA Exam Request List for AI: Evidence to Prepare by Requirement

How AI requests reach a firm, and how to organize for them

FINRA makes examination document requests through Request Manager, an application accessed through FINRA Firm Gateway, which keeps an audit trail of requests and productions. Once examiners are on-site and have a sense of the examination strategy, they can make more specific and granular requests around prioritized review areas. A firm that holds evidence addressing an exception not yet provided can offer it at the Exit Meeting or shortly after, for staff to review before the exam closes. Evidence that is assembled in advance and indexed to match your own production log can be located and produced quickly.

Regulatory Notice 24-09 reminds members that FINRA rules, which are intended to be technology neutral, and the securities laws continue to apply when firms use Gen AI and large language models. That holds for tools developed in-house, for third-party technology, and for AI features embedded in existing third-party products. The 2026 Annual Regulatory Oversight Report adds that GenAI use can implicate supervision, communications, recordkeeping and fair dealing. Those four areas are a reasonable spine for an anticipated request list, with the caveat that the list below is a planning aid and not an official FINRA checklist.

The most useful distinction for preparation is between documentary evidence and operational evidence. Documentary evidence includes inventories, procedures and approvals. Operational evidence is a per-interaction record of what an AI system was permitted to do and what it actually did. This is a working framework, not a regulator-defined taxonomy, but it explains why many firms can show a policy and committee minutes yet struggle to show how a control operated on a given day.

Inventory, supervisory procedures and vendor oversight

Inventory. An examiner is likely to ask which AI tools the firm uses and which supervised activities each supports. The inventory should include proprietary tools, third-party tools and embedded features in products already in use, since FINRA rules reach all three. Record the business owner, the supervised activity, the data the tool can access, and the approval date. The usual gap is embedded features that never went through a procurement review, and AI used in compliance functions themselves, such as surveillance summaries prepared for human reviewers.

Supervisory procedures. Rule 3110 requires a supervisory system reasonably designed and tailored to the firm's business, and Rule 3110(b) requires written procedures that members establish, maintain and enforce. Prepare procedures that name each AI use, define permitted and prohibited actions, and identify who reviews outputs. The typical weakness is procedures written at the level of general principle. An examiner who asks how a procedure was enforced needs records of enforcement, not only the procedure.

Vendor and third-party oversight. FINRA sent firms a November 2023 questionnaire on third-party vendor use, including vendor-supported AI, and follow-up outreach is ongoing. Hold due diligence files, contract terms, change notifications and periodic review records for each AI vendor. The common gap is oversight that stops at onboarding, with no evidence of what the vendor tool did inside your workflows afterward.

Communications review, recordkeeping and testing

Communications. The 2026 Report lists ensuring that GenAI-assisted customer communications comply with federal securities laws and FINRA rules. Expect requests for the communications types where AI assists, the review process applied to them, and samples showing that review occurred. Documentary evidence covers the review procedure and reviewer assignments. Operational evidence covers the specific output, the prompt or context that produced it, and the review outcome. Firms often have the first and not the second.

Recordkeeping. Examiners may ask which AI interactions are retained, in what form, for how long, and whether the records can be produced intact. Confirm retention periods against the applicable books-and-records rules with counsel, then verify that the AI systems actually retain what the schedule says. The usual gap is a retention schedule that covers email and chat but leaves prompts, outputs and tool calls in vendor consoles with short or unknown retention.

Testing and monitoring. Model risk management frameworks such as SR 11-7 describe development, implementation and use, validation, and governance, policies and controls. The NIST AI RMF organizes AI risk work into four functions, including GOVERN, and is not a checklist. Neither is a FINRA rule, so use them as organizing references rather than as stated examination requirements. Translated into evidence, they point to validation records, monitoring results, and per-interaction records of controls firing. The gap here is evidence that testing happened before deployment with nothing showing behavior in production.

Requirement-by-requirement evidence map

Use this map to build the inventory. Each row pairs the likely request with documentary evidence, operational evidence and the usual gap.

• AI inventory: Documentary: tool register with owner, activity and approval. Operational: usage records confirming what is actually in use. Gap: embedded and unsanctioned tools.
• Supervisory procedures: Documentary: written procedures per AI use. Operational: records of procedures being enforced. Gap: principle-level procedures with no enforcement trail.
• Vendor oversight: Documentary: due diligence, contracts, review cadence. Operational: records of vendor tool behavior in your workflows. Gap: onboarding-only oversight.
• Communications review: Documentary: review procedure and assignments. Operational: output samples with review outcomes. Gap: outputs not linked to reviews.
• Recordkeeping: Documentary: retention schedule. Operational: retained prompts, outputs and tool calls that can be produced intact. Gap: records held in vendor consoles.
• Testing and monitoring: Documentary: validation plans and results. Operational: policy decisions, blocks, redactions and escalations in production. Gap: no production-time evidence.

Where runtime enforcement can generate this evidence

Some operational evidence can be generated continuously as a by-product of operation, while other evidence has to be reconstructed after the fact. Reconstruction depends on vendor consoles, user recollection and sampling, which is weakest when an examiner asks about a specific interaction. A control placed in the AI execution path can record each decision as it is made.

Trussed enforces policy in the execution path across models, agents, tools, MCP servers, AI sites and coding assistants. Policy is enforced at runtime, not reviewed after the fact: an interaction or tool call is allowed, redacted, blocked, rate limited, or escalated for human approval. Each decision yields audit-ready evidence for that individual interaction or tool call, showing what the system was permitted to do and what it did. The platform deploys self-managed in the customer cloud, with configurable retention and data residency, which bears on the durability and retention criteria above.

The scope is specific. This covers individual interactions and tool calls, not transaction-level governance or end-to-end tracing of agentic workflows. Trussed has run continuously for 18+ months in a regulated enterprise, and its customers are in insurance, pharma, healthcare and education. It does not provide FINRA-specific mappings. Runtime evidence supports the operational rows of the map above, while inventories, procedures, vendor files and approvals remain documents your firm must maintain.