Reg BI Audit Trail for an AI Assistant: What You Need to Reconstruct a Recommendation
To reconstruct an AI-assisted recommendation under Reg BI, a firm needs one linked record per interaction. It should show the user input, the client and product context retrieved, the policy version applied, the enforcement decision (allow, redact, block or escalate), any human review or approval, the final output, and timestamps and identities for each step. Application logs alone rarely produce this record. Evidence generated by policy enforcement in the execution path is a stronger basis for showing what the assistant was permitted to do and what it did.
- Can the system produce the prompt, retrieved context, policy version, enforcement decision, human approval and final output together, with timestamps and identities?
- Are enforcement decisions recorded at execution time, or reconstructed from application logs afterward?
- Are individual tool calls and authorization decisions captured as separate events?
- Is human approval a distinct, identity-stamped event linked to the same interaction?
- Can retention be configured to meet the firm's Rule 17a-4 obligations, and is it clear where the evidence is stored?
- Does the vendor state plainly what the trail does not cover, and who covers those gaps?
What Reg BI asks a firm to be able to show
Reg BI's best interest obligation has four components: Disclosure, Care, Conflict of Interest and Compliance. The Disclosure Obligation requires prescribed disclosure before or at the time of the recommendation, covering the recommendation and the relationship between the retail customer and the broker-dealer. The Care Obligation requires reasonable diligence, care and skill. The Conflict of Interest Obligation requires policies and procedures reasonably designed to identify and disclose, mitigate or eliminate conflicts associated with a recommendation. The Compliance Obligation requires firms to establish, maintain and enforce policies and procedures reasonably designed to achieve compliance with Reg BI.
Two features of this framework shape the evidence problem. First, the SEC's adopting release describes the Care Obligation as an objective evaluation that turns on the facts and circumstances of the particular recommendation and the particular retail customer. An examiner will therefore ask about one recommendation, not about average system behavior. Second, the adopting release provides for Rule 17a-4(e)(5) to require retention of all records of information collected from or provided to each retail customer under Reg BI for at least six years after the earlier of account closure or the date the information was last replaced or updated. Confirm the current codified text with counsel.
When an AI assistant contributes to a recommendation, the assistant's inputs and outputs become part of the facts a firm may need to explain. That is the reason for building a per-interaction record.
How the elements map to each obligation
Disclosure is timed before or at the time of the recommendation. A record that holds timestamps for the assistant's output and for any disclosure delivered lets a firm show sequence. This is a design inference, not a stated recordkeeping rule.
For Care, the record should tie the assistant's output to the specific client information it used. Because the standard is facts and circumstances, a supervisor needs the retrieved context for that recommendation, not a summary metric. The mechanics of how a firm documents its care analysis remain a matter for its own procedures and counsel.
For Conflict of Interest, the SEC's 2023 proposal on predictive data analytics linked conflicts to the use of PDA-like technology in making recommendations. That proposal also contemplated written records and an annual review of related policies. It is a proposal, so treat it as a design reference rather than a current obligation. A useful record shows which conflict policy or rule was applied to the interaction and who is responsible for that check.
For Compliance, the key word is enforced. A record that shows the policy version and the enforcement decision at call time is direct evidence that a policy operated on that interaction. It also links a past recommendation to the rules in force when it was made, which matters as policies change. For rollover recommendations to retirement investors, DOL PTE 2020-02 separately requires documentation of why the advice was in the investor's best interest, and the evidence record may need to feed that documentation where it applies.
Where evidence gaps typically occur
Conventional application logs record that a request was made and a response returned. They often omit the policy version, the retrieved client context, and the decision logic that allowed or constrained an action. Where a tool call retrieves client data, the log may show the call but not the authorization decision that preceded it.
Post-hoc review creates a different gap. A reviewer sampling interactions after the fact can assess output quality, but cannot show that a policy was enforced when the interaction ran. The record shows what happened, not what the assistant was permitted to do. This distinction is interpretive, and no regulator statement in this area settles it, but it follows from the Compliance Obligation's reference to policies that are enforced and not only written.
A third gap is fragmentation. Prompts sit in one system, retrieved data in another, human approvals in a workflow tool and outputs in the advisor's communication channel. Reconstruction then depends on manual correlation by timestamp. A single interaction identifier shared across these events removes most of that effort.
There is also a retention mismatch to avoid. A NIST workshop report records one speaker's suggestion of a 30-day baseline for prompt and completion content. That is a workshop view, not a compliance period, and it is far shorter than the six-year period associated with Reg BI customer information records. Whether AI interaction records fall under the six-year period or another one is a question for counsel.
Stating the limits, and where an execution-path approach fits
A per-interaction record is bounded. It covers individual interactions and tool calls. It does not by itself trace an end-to-end agentic workflow across tasks and agents, and it does not document model lineage or dataset provenance. NIST workshop participants discussed broader telemetry, including model and version information and agent states. If your examination exposure includes those areas, plan to cover them with separate controls and say so in your supervisory procedures.
Trussed is one example of an approach that enforces policy in the execution path and generates audit-ready evidence from those enforcement decisions. It enforces policy at runtime, not after the fact, across models, agents and tools, with outcomes that include allow, redact, block and escalate for human approval. Retention and data residency are configurable, and deployment can be self-managed in the customer's cloud. Its scope is individual interactions and tool calls, consistent with the boundary above. Evaluate any such product with the criteria in this guide, whether or not you consider Trussed.
A practical next step is to select one recent AI-assisted recommendation workflow, list the systems that hold each of the six elements, and identify which ones cannot currently be joined by a shared identifier. That gap list defines your evidence requirements before you evaluate any tooling.
Minimum evidence record for one AI-assisted recommendation
Input and context
Prompt and the client and product data retrieved
Policy and enforcement
Policy version applied and the allow, redact, block or escalate decision
Human review
Identity-stamped approval or escalation outcome
Output
Final response, tied to the same interaction identifier and timestamps
Test your evidence record against one recommendation
Review how enforcement decisions, tool calls and human approvals are captured as audit-ready evidence for a single interaction.
See How Evidence Is ProducedHow does your AI governance program compare?
See where your program has gaps in less than 2 minutes.