Financial Services Recordkeeping

    Capturing Copilot Studio Agent Conversations for FINRA and SEC Rule 17a-4 WORM Retention

    First decide which agent prompts, responses and tool calls the firm treats as business records or communications. Then capture them at the point of execution as complete, attributed, time-stamped records. Write them to storage that meets either the WORM or the audit-trail requirement of Rule 17a-4(f), and prove you can produce them in a reasonably usable electronic format. As of November 2026, the amended rule is in force and the firm owns the classification decision and the storage.

    A capture-and-retain pattern with WORM handoff

    1. 1

      1. Interception at the execution path

      Place a capture point where prompts, responses and tool calls pass, so every interaction is seen before and after the model or tool acts.

    2. 2

      2. Record normalization

      Write one structured record per event with agent identity, user identity, channel, tool name, parameters, result, policy decision and a trusted timestamp.

    3. 3

      3. Integrity metadata

      Add sequence numbers and content hashes so completeness and accuracy can be verified when the record is written and again at retrieval.

    4. 4

      4. Handoff to compliant storage

      Deliver records to object lock-based storage or a designated third-party archive, configured to the elected WORM or audit-trail route and the retention period.

    5. 5

      5. Indexing and retrieval

      Index by agent, user, date and conversation so a record and its audit trail can be located and exported in a reasonably usable electronic format.

    6. 6

      6. Supervision feed

      Route the same records to supervisory review so monitoring of prompts, responses and outputs produces retained evidence.

    Criteria for examiner readiness

    • Route: the firm has documented whether it elects WORM or the audit-trail alternative, and the storage meets it.
    • Completeness: prompts, responses, tool calls and policy decisions are all captured, with a way to verify nothing was dropped.
    • Attribution: each record identifies the agent, the user, the channel and a reliable timestamp.
    • Retention: configurable to at least six years, with the first two years easily accessible.
    • Redaction and residency: sensitive data handling before retention is a documented decision, and records stay in the firm's chosen environment.
    • Production: a test request returns a record and its audit trail in a reasonably usable electronic format.

    What the rules ask the firm to preserve

    Start with the requirement, not the platform. Under the October 2022 amendments to Exchange Act Rule 17a-4, which had a compliance date of May 3, 2023, a broker-dealer using an electronic recordkeeping system must meet either the audit-trail requirement or the WORM requirement. Before the amendments, WORM (non-rewriteable, non-erasable) was the only electronic format. It remains an option, and the audit-trail alternative now sits beside it. The firm should document which route it elects.

    The rule also requires the firm to furnish a record, and its audit trail where applicable, preserved on an electronic recordkeeping system, in a reasonably usable electronic format when an SEC representative requests it. That makes production a design requirement, not an afterthought.

    On retention, Rule 17a-4(a) sets a minimum of six years for the listed Rule 17a-3 records, with the first two years in an easily accessible place. FINRA Rule 4511 sets a six-year default for books and records with no specified period. FINRA also describes Rule 17a-4(b)(4), Rule 3110(b)(4) and Rule 4511 as requiring preservation of communications relating to the firm's business as such, and its examples include chat and instant messages. FINRA's 2026 oversight report lists ongoing monitoring of GenAI prompts, responses and outputs as a control, which points to capturing those elements.

    Classify agent interactions before designing capture

    None of this text names AI agents, so classification is a firm decision. A Copilot Studio agent that answers client-facing or advisor-facing questions about firm business produces prompts and responses that look much like chat messages. Tool calls are less clear. A call that retrieves an account record or submits a request may be a required record in its own right, evidence of what the agent was permitted to do, or both.

    Write the classification down by interaction element: user prompt, agent response, tool invocation and parameters, tool result, and policy decision. Name the approver, and state the retention class for each. Compliance, records management and security should sign it together, because the capture design inherits every ambiguity left in the policy.

    FINRA's communications obligation also reaches non-firm or third-party channels used for firm business. Unmanaged agents or unmanaged channels are therefore a recordkeeping exposure even when the sanctioned deployment is well controlled.

    Test platform-native transcripts against the same criteria

    Native conversation transcripts and platform storage are convenient, but convenience is not the test. Evaluate any native store against the requirements above. Is retention configurable to the longest applicable period, with the first two years easily accessible? Can records be altered or deleted by administrators, and if so, does an audit trail capture the change? Do the stored records include tool calls and their parameters, or only the visible conversation? Can records be exported with attribution and timestamps intact, and in volume?

    If a native store cannot meet the elected route on its own, it can still be a convenience copy. The authoritative record is then the one written at capture and delivered to storage the firm controls. Document that decision, because an examiner will ask which copy is the record.

    Runtime enforcement as a source of evidence

    A retained conversation shows what was said. Examiners also want to know what the agent was allowed to do. Trussed enforces policy in the execution path on individual interactions and tool calls: allow, redact, block, rate limit, or escalate for human approval. Each enforcement decision generates audit-ready evidence at runtime, not after the fact, so the record includes the policy applied and the outcome as well as the content.

    Redaction needs a deliberate decision. Removing sensitive data before retention reduces exposure, but it changes what is preserved. The firm should decide, as part of classification, which fields may be redacted before the record is written and which must be kept intact, and it should retain the redaction event itself.

    Trussed deploys self-managed in the customer cloud, with configurable retention and data residency, so records can stay in the firm's own environment. Trussed does not by itself confer Rule 17a-4 compliance. WORM or audit-trail storage and the final records-classification decisions remain the firm's responsibility. Trussed is deepest in insurance, a regulated financial sector, and its scope is individual interactions and tool calls, not end-to-end tracing of agentic workflows.

    Capture-and-retain path for agent interactions

    Classify

    Decide which prompts, responses and tool calls are business records.

    Capture

    Record each interaction at the point of execution with attribution and timestamps.

    Enforce

    Apply allow, redact or block decisions in the execution path and keep the decision.

    Retain

    Write to WORM or audit-trail compliant storage in the firm's environment.

    Produce

    Furnish the record and its audit trail in a reasonably usable electronic format.

    Define your agent record capture pattern

    Review how in-path enforcement produces audit-ready evidence for agent interactions and tool calls, and how it fits your WORM or audit-trail design.

    Talk to an Expert
    Next step

    How does your AI governance program compare?

    See where your program has gaps in less than 2 minutes.