Compliance Guide

    AI-Generated Drafts and Broker-Dealer Books and Records

    As of November 2026, the question is interpretive. FINRA says its technology-neutral rules apply to generative AI, and the books and records rules cover communications relating to a firm's business as such. FINRA has also said AI use may lead to the creation of new records. Whether an unsent draft, or the prompt behind it, falls inside that scope turns on content, use and disposition. Firms need a documented position and controls that can enforce it per interaction.

    Evaluating controls at the AI execution layer

    • Policy is enforced per interaction and per tool call, with allow, redact and block outcomes, not only on chat text.
    • Each interaction carries a disposition tag, such as unsent, exported or sent, that can drive retention class.
    • Evidence records the user, prompt, output, model and policy decision, and is retrievable per interaction.
    • Evidence is stored in WORM or audit-trail form and exports in a searchable, sortable format.
    • Retention period and data residency are configurable per record class.
    • Retrieval has been tested against a simulated regulator request, and supervisory review can reach the same records.

    The tests that decide whether something is a record

    SEA Rules 17a-3 and 17a-4 set minimum requirements for the records broker-dealers must make, how long they must keep them and the formats allowed. FINRA's 2026 oversight report pairs them with FINRA Rules 2210(b)(4), 3110(b)(1), 3110.09 on retention of correspondence and internal communications, and 4511.

    For communications, the operative test is business as such. FINRA describes the obligation as creating and preserving, in an easily accessible place, originals of communications received and copies of communications sent that relate to the firm's business as such. Its examples are emails, instant messages, text messages, chat messages and interactive blogs. Per Regulatory Notice 25-07, this includes inter-office communications and communications with the public.

    FINRA Rule 4511 requires members to make and preserve books and records as required under FINRA rules, the Exchange Act and applicable Exchange Act rules, in a format and media that complies with SEA Rule 17a-4. Where no retention period is specified elsewhere, the default is six years, running from account closing for account-related records and otherwise from when the record is made.

    Supervision is a separate obligation that attaches to the same content. FINRA rules require supervision of internal and external communications relating to the investment banking and securities business.

    What the guidance says and where it stops

    FINRA states that its rules are technology neutral and continue to apply when member firms use generative AI, whether the firm builds the tools or uses a third party's, including embedded features in existing third-party products. That puts assistants, copilots, agents and vendor-embedded features all in scope for analysis.

    FINRA's AI report adds that the use of AI applications may lead to the creation of new records, and that firms should review their AI tools and systems for compliance with recordkeeping obligations. Regulatory Notice 25-07, a request for comment published in April 2025, asks how recordkeeping challenges vary by type of AI-generated communication, such as chatbot output or AI-generated transcripts and summaries of meetings. It is not a rule or an interpretation, and it does not resolve the drafts question.

    The open point is visible in industry correspondence. In a September 2026 comment letter, the ASA asked FINRA to confirm that the Rule 4511 and SEC recordkeeping obligation extends only to the final published communication, not to AI-generated drafts, prompts and iterations. That is a trade association's request, not a regulator position. Until FINRA or the SEC speaks directly, each firm is making a judgment that an examiner may test.

    Three retention scoping approaches and their tradeoffs

    A scoping decision should be explicit, reasoned and approved by named owners with counsel review. These approaches are design options, not regulatory categories, and this guide does not conclude that any one is sufficient.

    • Final communications only: Lowest volume and simplest retrieval. It depends on the view that drafts, prompts and iterations are outside scope, which is the position the ASA asked FINRA to confirm. If an examiner takes a broader view, the firm has gaps and little evidence of what the AI produced along the way.
    • Interactions that become communications: Retain the interaction record when an output is sent, shared or exported, and tag unsent ones by disposition. This ties retention to use and keeps volume bounded. It requires reliable disposition tracking across every AI surface.
    • All interactions on in-scope surfaces: Broadest coverage and the strongest answer to an examiner request. It creates large volumes of ephemeral output, higher storage and retrieval burden, and more content to protect and eventually dispose of by policy.

    What an examiner could reasonably ask to see

    No regulator has published a list of AI interaction data elements a firm must retain. A field list is therefore a firm-level design choice, and it should trace back to the business as such test. For a single interaction, a defensible set usually answers five questions: who the user was, what was asked, what the model returned, which model produced it, and what policy decision applied to the interaction.

    Format requirements are better settled. SEA Rule 17a-4(f) permits electronic recordkeeping systems that meet either the non-rewritable, non-erasable (WORM) requirement or the audit-trail alternative, which allows recreating an original record if it is altered, over-written or erased. The amendments took effect January 3, 2023, and firms preserving records electronically had to file updated third-party access undertakings reflecting Rule 17a-4(f)(3)(v). For AI interaction records, that points to immutable or versioned history.

    Production matters as much as preservation. Electronic records must be producible in a reasonably usable electronic format that regulators can search and sort, so raw model logs alone are a weak answer. Retention periods should be configurable per record class, because the six-year default in Rule 4511 applies only where no other period is specified, and communications follow the Rule 17a-4(b)(4) period.

    Where Trussed fits

    Trussed is runtime control and evidence infrastructure for regulated enterprises that must show what their AI was allowed to do and what it did. It enforces policy in the execution path across models, agents, tools, MCP servers, AI sites and coding assistants. Decisions are allow, redact, block, rate limit or escalate for human approval, applied to individual interactions and tool calls. Policy is enforced at runtime, not reviewed after the fact.

    That enforcement generates continuous audit-ready evidence. Trussed deploys self-managed in the customer's cloud, with configurable retention and data residency, which maps to the per-class retention and storage criteria above. Production use includes 18+ months of continuous operation in a regulated enterprise, 99.95% uptime on live production traffic, 100+ models under one policy plane and 30+ GB of audit evidence generated per day. Customers span insurance, pharma, healthcare and education, and insurance is where Trussed is deepest.

    Trussed does not decide your retention scope. That position belongs to your compliance function and counsel. Trussed supplies a layer that can enforce the position per interaction and produce the evidence an examiner would ask to see.

    Questions to answer for each AI interaction

    Is it a communication?

    Does the content relate to the firm's business as such?

    What happened to it?

    Was the output sent, shared, exported or discarded?

    Which record class applies?

    Retention period, format and storage location follow the class.

    Can it be produced?

    Searchable, sortable and retrievable for a regulator request.

    Turn a retention position into enforced policy and evidence

    Review how policy enforcement and audit-ready evidence work for individual AI interactions and tool calls in a self-managed deployment.

    Talk to an Expert