Supervising AI Client Onboarding Under Reg BI and FINRA Rule 2090
Treat each AI interaction and each tool call that reads or writes customer profile data as a distinct control point. Decide in advance whether it is allowed, redacted, blocked or escalated to a human, and enforce that decision while the AI operates. The enforcement layer then records the policy applied and the action taken, which gives supervisors durable evidence. Best-interest determinations, principal review, written supervisory procedures and testing remain firm responsibilities.
Control points in an AI-assisted onboarding workflow
- 1
Prompt handling
Inputs that carry personal or financial data, or that ask the AI to go beyond onboarding scope, are checked before they reach the model.
- 2
Response handling
Outputs are checked before the customer or an adviser sees them, including statements that could read as a recommendation.
- 3
Profile and CRM writes
Tool calls that create or change a customer profile are the highest-consequence step and usually warrant the strictest policy.
- 4
Sensitive data handling
Identifiers and financial details are redacted or blocked where the task does not need them in clear form.
- 5
Inferred versus stated values
Values the AI inferred or summarized are kept distinguishable from values the customer stated.
- 6
Human review triggers
Incomplete profiles, conflicting answers or low-confidence inferences route to a person before anything is written.
- Map Reg BI's disclosure, compliance and customer care components and Rule 2090 to each onboarding step.
- List every AI interaction and tool call that reads or writes profile data, and mark which pass through the enforcement path.
- Assign allow, redact, block or escalate to each, and record the rationale in written supervisory procedures.
- Define the evidence retained per decision, the retention period and who can access it.
- Test policies on a documented schedule, review escalation outcomes, and update procedures as the tools change.
Start with the rules, not the tooling
Regulation Best Interest was adopted by the SEC on June 5, 2019, under the Securities Exchange Act of 1934. It applies when a broker-dealer makes a recommendation to a retail customer, and it requires the firm to act in the customer's best interest at the time of the recommendation without placing its own interests ahead of the customer's. The general obligation has three components: disclosure, compliance and customer care. The compliance component requires written policies and procedures, including policies reasonably designed to identify and, at a minimum, disclose or eliminate conflicts of interest. FINRA Rule 2090, the Know Your Customer rule, was adopted into the consolidated FINRA rulebook alongside Rule 2111 on suitability. As of November 2026, neither is reported as amended or repealed.
These rules are generally described as technology neutral. The same supervisory, recordkeeping, disclosure and investor-protection duties apply whether a task is performed by a person, conventional software or an AI system. An AI assistant that collects customer information or drafts an investment profile does not create a separate regime. It adds a new place where existing duties must be met.
The rules do not prescribe control points for AI-assisted onboarding. Everything in the next sections is a firm design choice, built down from the regulatory text so that the firm can explain why each control exists and which obligation it supports.
Choosing allow, redact, block or escalate
For each control point, the firm decides which of four outcomes applies and documents the rationale in its written supervisory procedures. Allow suits routine, low-risk steps such as asking a customer a scripted question. Redact suits cases where the AI needs the interaction to proceed but does not need the raw sensitive value. Block suits actions the firm has decided the AI must never take, such as writing an inferred value into a profile field as if the customer had stated it. Escalate for human approval suits steps where a registered person must decide before anything proceeds.
The tradeoff is friction against assurance. Escalating too many steps removes the benefit of automation and trains reviewers to approve without reading. Escalating too few leaves consequential writes unreviewed. A workable approach is to start strict on profile writes and conflict-sensitive outputs, review the escalation volume and outcomes, and loosen policy only with documented reasoning.
Consistency matters as much as strictness. If one model or tool applies a policy and another does not, the firm cannot say its procedures were followed. Because the same supervisory standard applies regardless of the AI component, the same policy set should apply across every model and tool in the onboarding path. Trussed enforces policy in the execution path across models, agents, tools and MCP servers, so one policy can apply to individual interactions and tool calls regardless of which model handles them.
Why enforcement at runtime changes the evidence
After-the-fact review samples what the AI already did. Policy enforced at runtime, not reviewed after the fact, acts before the profile write or the customer-facing response takes effect, and the act of enforcing produces the record. That record shows what the AI was permitted to do, not only what it produced.
For each decision, a supervisory program should retain the interaction or tool call involved, the policy version in force, the action taken (allow, redact, block or escalate), a timestamp, and, for escalations, the identity of the approver and their decision. Keep this evidence separate from the AI system's own output so one cannot silently alter the other. Trussed generates continuous audit-ready evidence as a by-product of enforcement, and the same pattern applies in any design: evidence is a result of the control operating, not a separate reconstruction exercise.
Retention needs its own decision. The firm should confirm retention periods against the broker-dealer books-and-records rules that apply to it, and decide who controls access. Trussed supports self-managed deployment in the customer cloud with configurable retention and data residency, which lets the firm hold the evidence in its own environment.
For design reference, the SEC's 2023 proposal on predictive data analytics described a written description of a technology's material features before deployment, periodic updates, and an annual documented review of policies and their implementation. That was a proposal, not a requirement, but the structure is a sensible model for the firm's own documentation.
What stays with the firm
Runtime enforcement supports supervision. It does not perform it. The determination that a recommendation is in a customer's best interest, the review by a principal, the written supervisory procedures, and the testing of those procedures remain the firm's responsibility. An enforcement layer can route a step to a registered person and record the decision. It cannot make the decision, and it does not establish compliance with Reg BI or Rule 2090. Trussed makes no claim of certification under either rule.
Scope limits also matter. Enforcement applies to the interactions and tool calls that pass through its path, so the firm should map which onboarding steps occur outside it, such as manual data entry or a separate vendor tool, and cover them by other means. Trussed enforces policy on individual interactions and tool calls. It does not provide transaction-level governance across agents or end-to-end tracing of agentic workflows, so firms that need that view must supply it separately.
Finally, AI onboarding tools can introduce conflicts of interest. Because the compliance component requires policies reasonably designed to identify and disclose or eliminate conflicts, the firm should assess whether the tool's prompts, defaults or outputs favor particular products, and treat that assessment as part of its own procedures.
From rule text to runtime control
Rule
Reg BI and Rule 2090 set the supervisory starting point.
Control point
Each AI interaction and profile tool call is a policy decision.
Enforcement
Allow, redact, block or escalate while the AI operates.
Evidence
Policy version, action and approver retained for review.
See what the AI was allowed to do and what it did
Review how policy enforced in the execution path produces audit-ready evidence for AI-assisted onboarding in a regulated enterprise.
Request a DemoHow does your AI governance program compare?
See where your program has gaps in less than 2 minutes.