Interactive Brokers invites about 3,000 clients to let an AI agent trade on its own

On October 15, 2026, Interactive Brokers held its third-quarter earnings call. Between the revenue lines, management described a new product called Autonomous AI trading. We read the full call transcript posted on IBKR's investor site and kept to what executives said on it. Press summaries were used only to find the account count.

What was launched

The prepared remarks put it plainly: the product "allows eligible clients to connect the AI agent of their choice to their IBKR account and authorize it to trade autonomously." It works with any compatible MCP client or AI platform. MCP, the Model Context Protocol, is an open standard that lets an AI application talk to outside tools and data.

Clients create and fund a dedicated partition for the agent inside their existing account, in a single session, without opening a new account. Per the transcript, the agent can trade global stocks and options, using order types and sessions that include Day, Overnight, and Overnight plus Day. The rollout is underway in several countries, with more to follow.

How small it is right now

In the Q&A, CEO Milan Galik called it "a very limited rollout." IBKR invited only clients who had already used its human-in-the-loop AI Connector, and he said approximately 3,000 clients were invited over the last several days. He added that the number is "going to greatly increase" and that eligibility will widen over the next few weeks.

For scale, a Zacks summary of the call puts IBKR at 5.58 million customer accounts. Dividing 5,580,000 by 3,000 gives roughly one invitation for every 1,860 accounts. That is our arithmetic, and it is rough: invitations went to clients, and accounts are not the same unit. The point is the order of magnitude. This is a pilot-sized group.

Invited is also not active. The transcript does not say how many of the 3,000 connected an agent, funded a partition, or placed a trade.

The step before it: the AI Connector

Autonomous trading sits on top of a tool IBKR says has been popular. In the prepared remarks, management described the AI Connector as "one of the fastest-adopted features we have ever introduced." This quarter IBKR added Perplexity as an official partner, extended support to any MCP-based tool including coding and desktop assistants, and added account-level profit and loss, alert management, and watchlists. It also added two-leg options strategies.

No adoption figure was given for the Connector either, so "fastest-adopted" is a company description, not a number we can check.

What management expects

Asked about the long-run effect, Galik said he expects clients to become "more comfortable with AI making trading decisions on their behalf," that good investors "will become much more productive," and that IBKR therefore expects higher trading volumes from AI adoption over time. That is a forecast from the firm selling the product.

It helps to know how the firm earns money. In the same call the CFO said commissions rose 12% from the year-ago quarter to $602 million.¹ More activity per account is good for that line. We are not saying that drives the product. We are saying the expectation of more trading comes from a business that is paid per trade.

What the transcript does not say

We searched the transcript for safeguards and found none described. It does not say what limits an agent faces on order size or frequency, whether there is a kill switch, who is responsible when an agent misreads data or a prompt goes wrong, or how the firm supervises agent orders. It describes a funded partition but does not spell out whether the agent can touch anything outside it.

Those may well exist in IBKR's client documentation. We have not seen it, so we are not claiming either way.

How this compares

The same day, Schwab said it blocks agents that log in for clients and keeps money movement off limits for connected models. Axi and eToro have also opened MCP routes, covered in our broker roundup. IBKR is now the large broker furthest along on one specific axis: letting the agent, not the human, send the order.

If you are testing agents on any broker

The useful habit is to read the permission model before the feature list. Ask what the agent can do, with how much money, and what stops it. Separately, an agent that is right in a demo can still be wrong with live data, which is why we wrote about lookahead bias in AI backtests. This article is about a product announcement, not a performance claim.