Robinhood opened its brokerage to AI agents this week. Customers can fund a separate account, hand the keys to Claude or ChatGPT, and let the agent trade stocks plus make credit card purchases on their behalf. FINRA flagged it as a new risk area within hours. The asymmetry between agent capability and agent accountability has finally landed at the SEC's doorstep.

On 27 May 2026, Robinhood announced that customers can now connect AI agents to a dedicated trading account, fund it with a specified balance, and let the agent execute trades on its own. The agent connects via Anthropic's Model Context Protocol. It can buy and sell across the brokerage's full universe of equities. It can also make credit card purchases on a paired account. The customer sets a cap. After that, the agent acts.

The agent makes the trade. The agent has no fiduciary duty. The customer does.

The mismatch is the story. FINRA has already named the category a new risk area. The regulator's existing rulebook assumes a human broker, a human customer, and a relationship between them governed by suitability rules, fiduciary obligations, and the broker's registration. Robinhood just introduced a third party that has none of those things.

This is the MM Liability Gap applied to brokerage. Three questions matter: who decides, who acts, and who pays when it goes wrong. Robinhood has answered the first two without answering the third.

What Robinhood actually shipped

The implementation matters because it is more clever than the headline suggests.

Robinhood did not bolt an AI feature onto the existing customer account. It created a separate, ring-fenced account that the AI agent operates inside. The customer funds it with a fixed amount of capital. The agent's authority is scoped to that capital. If the agent loses the balance, the loss is bounded by what the customer chose to put in. If the agent makes money, the customer keeps the gains.

The mechanism is the Model Context Protocol, Anthropic's standard for letting agents read and act on external systems. Robinhood ships an MCP server. The customer's chosen agent, Claude or otherwise, connects to that server. The connection includes scope: which account, what spending limits, which permissions. The agent then operates inside that scope.

That sounds careful. It is, at the account-design level. What it does not solve is the regulatory question one layer up.

What FINRA's rulebook actually says

FINRA Rule 2111 requires brokers to recommend only securities suitable for the customer, given their financial profile, risk tolerance, and investment objectives. Rule 3110 requires supervision of any registered representative making recommendations. Rule 4512 requires the broker to know the customer.

These rules are built around a model where a human at the firm makes a recommendation to a human customer. Robinhood, like every modern self-directed broker, has been navigating these rules for a decade by arguing that its platform does not make recommendations, only executes customer-directed trades.

An AI agent breaks that argument. The agent is making recommendations. The agent is executing trades based on its own analysis. The customer authorized the existence of the agent, but the customer did not authorize each individual trade. That puts FINRA in an awkward spot. The platform looks like it is offering recommendations through a third-party intermediary, except the third party has no FINRA registration, no Series 7, no fiduciary duty, and in some cases no readable training data.

As Jason Mikula has been documenting in Fintech Business Weekly across multiple FINRA enforcement actions this year, the regulator's posture has shifted from suitability questions toward platform-level supervision questions. The Robinhood announcement walks straight into that shift.

The MM Liability Gap, brokerage edition

Three questions, three answers, one mismatch.

Who decides. The AI agent. The customer set up the account and chose to deploy the agent. The customer did not pick the trade. The customer may not even know the trade happened until it shows up in the daily statement.

Who acts. Robinhood. The agent connects to Robinhood's MCP server. Robinhood's systems execute the trade. The settlement chain runs through Robinhood as the broker-dealer of record. The agent has no direct relationship with any clearinghouse, any exchange, or any custodian.

Who pays when it goes wrong. The customer. If the agent loses money on a trade, the loss is the customer's. If the agent makes a discriminatory or biased decision based on patterns in its training data, the customer cannot reasonably sue Anthropic or OpenAI for the trade outcome. If the agent gets prompt-injected by adversarial content on the public internet and trades into a manipulated position, the customer carries the loss.

The asymmetry is the gap. Three parties, three roles, three different exposures, and a regulatory framework that was built before any of this was possible.

We covered the adjacent pattern in Finix plugged three frontier models into its processor, the liability layer is still empty in April. The Finix story was payment processing. The Robinhood story is brokerage. Different rail, same gap.

What FINRA is about to do

The regulator has options. Three are likely.

The first is a clarifying notice within the next 60 days. FINRA does this when it wants to put platforms on notice without going through full rulemaking. A notice would name AI-agent-initiated trades as recommendations within the meaning of Rule 2111, require platforms to supervise the agents as if they were registered representatives, and lay groundwork for future enforcement.

The second is a sweep. FINRA periodically asks every platform a structured set of questions on a particular topic. An AI agent sweep would ask: who can connect, what data the agent receives, how the platform supervises the agent's behavior, what disclosures the customer gets, and what records the platform maintains. Sweeps usually precede enforcement by six to nine months.

The third is enforcement. If a platform's AI agent causes a measurable customer loss in the next year, FINRA will use the case to set the precedent the industry needs. Robinhood's bet is that the ring-fenced account design will keep it out of the first wave of enforcement. The bet may pay off. It may not.

The SEC will be slower but eventually weigh in on the fiduciary question. The agent is not a fiduciary. The customer accepted the agent's actions. Whether Robinhood owes the customer a fiduciary duty for permitting the connection is the unresolved question that lands at the SEC, probably in 2027.

What this means for every other broker

The category opens. Within ninety days, expect Charles Schwab, Fidelity, and E*TRADE to announce equivalent capabilities. Interactive Brokers has been quiet about its agent strategy but has the engineering capacity to ship something competitive within a quarter. SoFi and Public are the most likely fast-followers among the modern brokers. The Robinhood window for being the only agent-friendly broker is short.

The differentiation will come from the supervision layer, not the connection layer. Connecting an agent via MCP is now table stakes. Detecting agent behavior that looks like market manipulation, detecting prompt injection in real time, attributing decisions back to the agent's training data, producing audit-grade records of the agent's reasoning. That is where the next round of platform investment lands.

Theodora Lau wrote about agentic AI in banking last quarter and the operational learning curve incumbents accumulate by moving early. The same logic applies here. The broker that ships the supervision layer first sets the procurement standard for the rest. The brokers that wait will integrate against somebody else's vendor.

The bigger question, the one nobody on a US broker earnings call wants to answer yet: when the agent loses money it should not have lost, who pays the customer back? Robinhood has answered the technical question of how the agent gets connected. The financial question of who carries the loss is still where it always was. With the customer.

Sources

The agent can trade. The agent has no fiduciary duty. The customer does. Three parties, three roles, three different exposures, and a regulatory framework that was built before any of this was possible. Whose loss is whose?

Charlie Major is a Product Development Manager at Mastercard. The views and opinions expressed in Major Matters are his own and do not represent those of Mastercard.