Maintaining a Single AI Audit Trail for FINRA and SEC Recordkeeping in Dual-Registrant Firms
A dual-registrant firm can keep one AI audit trail if it treats the trail as an evidence-design problem. Map each AI artifact (prompt, output, tool call, policy decision) to the Exchange Act, FINRA Rule 4511 and Advisers Act Rule 204-2 record categories. Give every record a registrant entity, a unique individual identifier and a retention class. Make sure the store can recreate altered originals and export records with their audit trail. Evidence generated by policy enforcement in the execution path is easier to defend than evidence reconstructed afterward.
Design a common record schema with registrant-level attribution
- 1
Individual identity
A unique identifier for the person or service that created, modified or deleted the record.
- 2
Registrant entity
The broker-dealer or the investment adviser on whose behalf the activity occurred. This is the primary filter for production.
- 3
Interaction and tool call
The prompt, output and each tool invocation as discrete records, so every step captured is attributable.
- 4
Policy decision and action
Which policy applied and whether the result was allow, redact, block or escalate for human approval.
- 5
Timestamps and change history
Time of creation plus time and actor for every modification or deletion.
- 6
Retention class
A label that drives the applicable retention period, set by the legal mapping and changeable only by named owners.
- Scoping matrix signed off by compliance and counsel, with each AI artifact mapped to a regime and registrant.
- Documented choice of WORM or audit-trail route per system, with proof that an altered or deleted record can be recreated.
- A sample record exported with its audit trail in human-readable and reasonably usable electronic formats, retrieved using locating information.
- Production filtered by registrant entity, returning no records from the other registrant.
- Retention classes with named owners and an audited change process.
- Agent and multi-step activity tested to confirm which steps are captured at each stage.
Start with scope: which AI artifacts are records
Dual registrants operate under overlapping but non-identical regimes. The broker-dealer is subject to Exchange Act Rules 17a-3 and 17a-4, which set minimum requirements for the records a broker-dealer must make, how long it must keep them and the format they may be kept in. FINRA Rule 4511 requires members to make and preserve books and records as required under FINRA rules, the Exchange Act and the applicable Exchange Act rules. The investment adviser is subject to Advisers Act Rule 204-2, which enumerates specific record categories, for example brochures, brochure supplements, Form CRS and the dates they were given to clients.
FINRA states that its rules are technology neutral and apply when member firms use generative AI, just as they do for any other tool. Regulatory Notice 24-09 creates no new requirements and relieves firms of none. The practical consequence is that AI evidence belongs inside existing supervision and recordkeeping frameworks. Whether a particular prompt, output or tool call is a required record under a given rule is a determination the firm must make with counsel, and it should be documented with a named approver.
The working artifact is a scoping matrix. Rows are AI artifacts: user prompts, model outputs, tool calls, policy decisions and any human approvals. Columns are Rules 17a-3 and 17a-4, FINRA Rule 4511 and Rule 204-2. Each cell records whether the artifact falls under that regime, which registrant it belongs to and who signed off. Many artifacts will land in more than one column, and that is the case the single store has to handle.
Reconcile retention and storage inside one repository
Choose the storage route per system. If you use WORM, the broker-dealer rules call for an audit system that accounts for the inputting of records and any changes to every original and duplicate. If you use the audit-trail alternative, the test is whether the original can be recreated after alteration or deletion. That turns on the design of the audit trail and not on a specific storage product. A versioned, tamper-evident store with change history is one design that fits that description, though you should validate it against the full rule text before relying on it.
For retention, the retention class field lets one repository apply different periods to different records. Where a record falls under both regimes, many firms choose to apply the longer applicable period and document the decision. Changes to a class should themselves be audited actions.
Deployment location matters here. A repository that runs in the firm's own cloud, with configurable retention and data residency, keeps the retention clock, access controls and export tooling under the firm's direct control. That supports clear ownership of the controls during an examination.
Generate evidence at execution, not afterward
FINRA's 2026 oversight report lists auditability and transparency as a concern for generative AI. It notes that complicated, multi-step agent reasoning tasks can make outcomes difficult to trace or explain. Reconstructing a record after the fact from application logs is where most gaps appear: missing tool calls, ambiguous identity, no record of what was prevented.
Enforcing policy in the execution path changes the evidence. Each interaction and tool call passes through a decision point before it completes. The decision (allow, redact, block or escalate for human approval) is made at runtime and written with the registrant, identity and timestamp already attached. A blocked request and an approved escalation then produce records as well, which matters for supervisory review and model risk management.
Be explicit about limits. Records captured at the level of individual interactions and tool calls address the tracing concern only for the steps they capture. Governance documents should state which steps are recorded and which are not, rather than implying complete coverage of multi-step agent behavior.
Where Trussed fits
Trussed enforces policy in the execution path across models, agents, tools, MCP servers, AI sites and coding assistants, with actions that include allow, redact, block, rate limit and escalate for human approval. The scope is individual interactions and tool calls. Audit-ready evidence is generated by that enforcement and not assembled afterward. Trussed deploys self-managed in the customer cloud, with configurable retention and data residency.
Trussed's evidence is currently mapped to HIPAA, NAIC, NYDFS, the EU AI Act and NIST. Insurance, a regulated financial sector, is where Trussed is deepest. Mapping to FINRA and SEC record categories remains the firm's own legal analysis.
Operational proof points: 18+ months of continuous production 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. Trussed has customers in four regulated verticals: insurance, pharma, healthcare and education. It holds SOC 2 Type II, HIPAA, GDPR, FERPA, ISO 27001 and NIST AI RMF alignment.
One evidence store, two examiners
Scope
Map prompts, outputs, tool calls and policy decisions to each regime.
Schema
One record format with registrant and individual attribution.
Retention
A per-record retention class inside a single repository.
Production
Export the record and its audit trail for either examiner.
See how enforcement produces audit-ready evidence
Review how policy enforcement in the execution path generates evidence for individual interactions and tool calls, and how it deploys in your own cloud.
See How Evidence Is ProducedHow does your AI governance program compare?
See where your program has gaps in less than 2 minutes.