Our broker-dealer is starting to use AI agents that draft client messages, summarize calls and run back-office workflows. Do SEC Rule 17a-4 and FINRA's recordkeeping rules apply to what the agents produce, and can we keep those records in an audit-trail system instead of WORM storage?

What should I check, step by step?

1. List every agent and what it produces: client emails or chat replies, call or meeting summaries, notes written to the CRM, trade or account workflow steps, and internal messages about the business. Mark which ones are records under Rule 17a-3 or business communications under 17a-4(b)(4).

2. For each record type, note the retention period: at least three years for 17a-4(b) records, the first two easily accessible, and six years under FINRA Rule 4511 where no other period is set. Some 17a-3 records have longer periods; check the rule for each one.
3. Find where each output lives today. A vendor app, an observability tool or the agent framework's own log usually isn't your 17a-4 archive, and many expire or let an admin delete entries.
4. Route record outputs into a system that meets 17a-4(f): WORM, or an audit trail that keeps every change and deletion with date, time and who made it, and can re-create the original.
5. Write down whose action each agent record is. The rule asks for the identity of the individual where applicable, and an agent isn't a person, so record the agent's own identifier plus the account or employee it acted for and who approved it.
6. Keep the inputs that explain a record: the prompt or instruction, the model and version used, and the approval, as FINRA's report suggests for monitoring. Store them with the record or link them by id.
7. Check the vendor: confirm the archive or agent vendor can meet the rule, and that the contract says so. FINRA's 2026 report lists weak due diligence of third-party recordkeeping vendors as a finding and suggests testing them by asking for records the way an examiner would.
8. Run the drill: pick an agent-drafted client message from months ago and produce it, with its audit trail, in a human-readable and a usable electronic format, as Rule 17a-4(f)(2)(iv) requires.

What has FINRA said about AI agents, supervision and recordkeeping at broker-dealers?

Regulatory Notice 24-09 (June 27, 2024) says FINRA's rules apply when member firms use generative AI just as they apply to any other technology, whether the firm builds the tool or uses AI embedded in a third-party product. It names supervision, communications, recordkeeping and fair dealing as areas the use can touch.

• The 2026 Annual Regulatory Oversight Report has a section on emerging trends in AI agents. Its listed risks include autonomy without human validation and approval, agents acting beyond the user's actual or intended scope and authority, and multi-step reasoning that makes outcomes difficult to trace or explain, complicating auditability. Its considerations include how to monitor agent system access and data handling, where to have human-in-the-loop oversight, how to track agent actions and decisions, and how to set guardrails that limit agent behavior.
• On monitoring, the report says this may include storing prompt and output logs for accountability and troubleshooting and tracking which model version was used and when. A FINRA blog post of January 27, 2026 lists the agent types firms are exploring: conversational, software development, fraud detection and prevention, trade and AML surveillance, process automation and optimization, and trade execution agents.
• In Regulatory Notice 25-07 (April 14, 2025; comments extended to July 14, 2025) FINRA asked members what recordkeeping challenges they face with AI-generated communications, giving AI chatbots and AI-generated transcripts or summaries of meetings as examples. Asking for comment doesn't change the current rules, so the existing requirements still apply.

What does SEC Rule 17a-4(f) require for an audit-trail recordkeeping system instead of WORM storage?

Under Rule 17a-4(f)(2)(i), an electronic recordkeeping system must either (A) preserve each record for its whole retention period with a complete time-stamped audit trail, or (B) preserve records exclusively in a non-rewriteable, non-erasable format. The audit trail must include all modifications to and deletions of the record or any part of it, the date and time of actions that create, modify or delete it, the identity of the individual doing so where applicable, and anything else needed to keep security, signatures and data so the record stays authentic and reliable and the original can be re-created if it's changed or deleted.

• Either way the system must also verify automatically the completeness and accuracy of its storage and retention processes, be able to download and transfer a record and its audit trail in both a human-readable and a reasonably usable electronic format, and include a backup system or other redundancy so records stay available if the main system goes down.
• The firm must be able to produce records immediately on request, and must file undertakings with its examining authority signed by a designated executive officer or a designated third party, who commits to provide the records. The SEC's adopting release says the audit-trail option lets firms use the business systems they already run, provided those systems meet these tests.
• In practice, an append-only log isn't enough on its own. The test is whether the system keeps the original and every change, shows who and when, and can hand an examiner the record and its history in a usable format for the whole retention period.

Why don't AI agent logs usually meet broker-dealer recordkeeping requirements?

The record lives only in the agent vendor's app or in a tracing tool with a short retention window, and nobody exported it to the firm's archive. FINRA's 2026 report lists not retaining, archiving and reviewing non-email electronic communications sent through firm-approved channels as a finding; agent-sent messages are a newer version of the same gap.

• The log can be edited or trimmed by an administrator, or an agent rewrites its own memory or notes, and no history of the change survives. That fails the audit-trail test even if the latest version looks right.
• The record exists but nothing ties it to who acted: which agent, for which employee or account, under whose approval, with which model version. Without that, the record can't answer the questions an examiner or a supervisor will ask.
• The firm can't produce it fast. If getting an agent's past messages needs a vendor ticket and a custom export, run that drill now, before an examination.