Supervising the Machine

Brokerage firms have a new kind of colleague, and FINRA has started describing it. Its 2026 Annual Regulatory Oversight Report, published in December 2025, gives generative AI a section of its own for the first time and, also for the first time, addresses AI agents directly: "systems or programs that are capable of autonomously performing and completing tasks on behalf of a user."

The report's catalog of what firms are actually doing is the clearest picture yet of how the technology has entered the industry. A year earlier, FINRA described three common uses: summarization, analysis and policy retrieval. The 2026 report lists more than a dozen, from conversational question-answering and translation to drafting, classification, coding, workflow automation and synthetic data. "Summarization and information extraction" remains the most common.

The risks FINRA attaches to agents are specific. An agent may act without human validation or approval. It may act "beyond the user's actual or intended scope and authority." Its multi-step reasoning can make an outcome hard to trace or explain. It may mishandle sensitive data, lack domain expertise, or pursue a poorly designed objective, and it carries every existing GenAI risk — hallucination, bias, privacy — along with it. The effective practices that follow are equally concrete: a governance framework with documented policies, pre-deployment testing, logging of prompts, outputs and model versions, human-in-the-loop review, and guardrails that limit what an agent is allowed to do.

What the report doesn't do is invent a new rulebook. FINRA's rules are intended to be "technologically neutral," a position it set out in Regulatory Notice 24-09 and repeats here. A message drafted by a model is still a communication under Rule 2210. A workflow run by an agent still needs to be supervised under Rule 3110. A prompt and its output may still be a record under Rule 4511 and Exchange Act Rule 17a-4. Generative AI, in FINRA's words, "can implicate rules regarding supervision, communications, recordkeeping and fair dealing." The obligations stay with the firm no matter which tool does the drafting.

This turns the familiar question around. The common question is how much work a model can take over. The regulatory question is which actions need a person, and whether the firm can show afterward who approved what. For firms building with AI, that means designing for supervision from the start, which is cheaper than adding it later. Models can draft, sort and route. The firm still answers for whatever goes out under its name.