AI Record Retention Schedule for Broker-Dealers: Mapping Prompts, Outputs, and Policy Decisions to 3-Year and 6-Year Rules

Retention is a classification problem before it is a storage problem

Most AI retention efforts start by capturing everything and deciding later. That produces large volumes of data without answering the examiner's question: which of these records does the firm have to keep, for how long, and can it produce them?

SEC Rule 17a-4 sets retention by record type, not by technology. Records required under 17a-3(a)(1) to (3), (5), (21) and (22) are preserved for at least 6 years. Records required under 17a-3(a)(4), (6) to (11), (16), (18) to (20) and (25) to (31) are preserved for at least 3 years. Rule 17a-4(b)(4) separately requires 3 years for originals of communications received and copies of communications sent, "and any approvals thereof," including communications relating to the firm's business as such. FINRA Rule 4511(b) adds a 6-year default for FINRA-required records where no other period is specified.

FINRA states that its rules are technology neutral and continue to apply when firms use generative AI, including supervision, communications and recordkeeping. Its guidance is not a new rule. It does tell firms to review AI tools for recordkeeping compliance before testing and deployment. As of November 2026, the retention period for an AI artifact therefore follows the record it supports, and deciding which record that is remains a determination for the firm and its counsel.

Mapping framework: artifact to category, period and note

Treat the table as a starting hypothesis for counsel review, not a conclusion. The period attaches to the underlying record, so the same artifact type can land in different rows depending on the business function.

• Prompts and outputs that form a business communication: Category: communication relating to the firm's business. Candidate period: 3 years under 17a-4(b)(4). Note: confirm with counsel whether the artifact is a communication, and whether public-communications rules also apply.
• Human approvals of AI-assisted communications: Category: approval of a communication. Candidate period: 3 years, from the "and any approvals thereof" language. Note: link each approval to the communication it approved so both can be produced together.
• Outputs and tool calls supporting a 17a-3 record: Category: the underlying 17a-3 paragraph. Candidate period: 6 years or 3 years, depending on the paragraph. Note: inventory use cases by business function first, because the underlying record drives the period.
• Enforcement decisions (allow, redact, block, escalate): Category: supervisory and control evidence tied to the interaction it governed. Candidate period: the period of the linked record, or the 6-year default if no period is specified. Note: retain the link to the prompt, output and approval.
• Model risk explainability summaries: Category: model risk management documentation, such as a written summary of key input factors and output rationale. Candidate period: set by counsel. Note: FINRA's AI guidance describes such summaries as part of model risk management.

Trigger events, the 6-year default and retrieval tiers

A period is only usable with a start event. FINRA describes the 6-year period as a default: absent contrary guidance, records pertaining to an account are kept 6 years after the account is closed, and other records 6 years after they are made. Document which trigger applies to each row. Where a firm chooses a 3-year classification over the 6-year default, record the reasoning, since the longer default applies when no period is specified.

Retrieval has two parts. Under Rule 17a-4, both the 3-year and 6-year tiers require the first 2 years to be in an easily accessible place. Design and test retrieval for that window regardless of tier, then confirm the artifact remains retrievable for the rest of its period. FINRA Rule 4511(c) also requires FINRA-required records to be preserved in a format and media that comply with SEA Rule 17a-4. Any store that holds AI artifacts, including a third-party recordkeeping arrangement, should be assessed against that standard using the full rule text.

Drafting the schedule: fields, owners and procedures

A schedule that an examiner can follow has the same columns for every artifact class. The fields below are a suggested design, not a regulatory requirement. Align written supervisory procedures to the schedule, since FINRA Rules 3110 and 3120 call for supervisory systems and procedures reasonably designed to comply with recordkeeping obligations. Where Rule 3110.07 applies to review of AI-assisted communications, evidence of that review belongs in the schedule, with its period confirmed against the full rule text.

• Artifact type and record category: The AI artifact and the communication, 17a-3 record or FINRA record it supports.
• Governing rule and period: The 17a-4 paragraph or FINRA rule, the 3-year or 6-year period, and the counsel-approved rationale.
• Trigger event: Record creation, account closing or another defined start point.
• Named owner: One accountable owner per artifact class, tied to supervisory procedures.
• Storage and retrieval standard: Location, format and media, easy access for the first 2 years, and retrieval for the full period.
• Legal hold and disposal: How holds suspend disposal and how expired records are removed and the removal documented.

Evidence from the execution path, and the tension with minimization

To reconstruct an AI interaction, an examiner needs to see who acted, what input was given, what output was returned, what policy decision applied and what human approval followed. FINRA's 2026 oversight report refers to storing prompt and output logs for accountability and troubleshooting, and to data management that addresses retention. Prompts and outputs alone do not show what the system was permitted to do. The enforcement decision does, and it should be retained and linked to the interaction it governed, for the same period.

This is where runtime enforcement matters for evidence. Trussed enforces policy in the execution path across models, agents, tools and MCP servers, applying allow, redact, block, rate limit or escalate for human approval to individual interactions and tool calls. Because the decision is enforced at runtime, not reviewed after the fact, the audit-ready evidence is generated by the enforcement itself and can be assigned to a schedule row. Retention is configurable, and self-managed deployment in the customer cloud gives the firm control over data residency. Whether any evidence store meets Rule 17a-4 format requirements is a question for the firm's counsel and technology teams.

Retention also conflicts with data minimization when prompts and outputs contain customer or nonpublic personal information. Redaction before storage means the retained record may differ from what the user typed. Decide in advance, with counsel, whether the retained record is the redacted version, the original or both, who can access each, and where each is stored. Write those rules into the schedule so the answer to an examiner does not depend on one engineer's configuration.