Reg BI and AI-generated recommendations.

Reg BI and the advisers fiduciary duty are different standards in different rulebooks, and a large share of what is written about AI and best interest blurs them. If you are a registered investment adviser giving advice as an adviser, Reg BI is not your rule. Section 206 is.

If you are a dual registrant, both are live, and which one applies turns on the capacity you are acting in at the time rather than on the software. That has a practical consequence worth designing for: one AI workflow can sit under Reg BI in one interaction and under Section 206 in the next. Firms that document the tool but never document the capacity are the ones who cannot answer this in an exam.

This is the question the rule actually hinges on, and it is decided on facts and circumstances the same way it was before any of this existed. The test is whether the communication would reasonably be viewed as a call to action, and whether it was individualized enough that the customer would reasonably rely on it.

Applied to tooling, that produces a gradient rather than a line. A model that tidies the grammar of a paragraph is nowhere near it. A model that summarises a fund's prospectus is closer. A model that takes a customer's holdings and surfaces three specific alternatives is close enough that you should assume it has crossed and design accordingly.

The operational test that works: if the customer acted on the output and lost money, would you be comfortable telling a regulator it was not a recommendation? If the answer requires a long explanation, it was one.

The Care Obligation requires you to consider reasonably available alternatives. That obligation was written on the assumption that a person decides what to look at. When a tool decides which alternatives ever reach your screen, part of the obligation has moved into software, and you are still answerable for it.

This is the specific thing to ask a vendor, and almost nobody asks it: what universe does the tool search, what does it exclude, and on what basis. A screen that quietly drops share classes, or ranks by a factor nobody at your firm chose, is making a Care Obligation decision on your behalf.

Three things worth having in writing before you rely on one:

• The universe the tool considers, and what is outside it. Not a marketing claim about coverage — the actual exclusions.
• The ranking or filtering logic at a level you could describe to a customer. You do not need model weights; you need to know what it optimizes for.
• Whether the vendor has any economic relationship with what the tool surfaces. That is a Conflict of Interest question and it belongs in your diligence file either way.

Reg BI does not require a particular form of documentation, which firms often read as good news and should not. The Compliance Obligation requires policies reasonably designed to achieve compliance, and an examiner reconstructs whether that happened from whatever you kept.

For a recommendation an AI was in the path of, the defensible file shows what the customer's situation was, what alternatives were considered, and that a named person exercised judgment over the output rather than forwarding it. The last part is the one that goes missing, and it is the one that matters most.

Reg BI is Exchange Act Rule 15l-1, adopted in Release 34-86031 and in force since June 2020. The obligations described above are the four the rule sets out. Last reviewed at source: September 2026.

1. 17 CFR 240.15l-1, Regulation Best Interest (checked 2026-10-27)
2. SEC Release 34-86031, Regulation Best Interest: The Broker-Dealer Standard of Conduct (2019) (checked 2026-10-27)

This is not legal advice and it is not a substitute for your compliance counsel. It is how a practitioner who implements these systems reads the rule, written down so you can check it. If something here is wrong, tell me and it gets corrected.