IBKR removed the per-order approval button

The approval button is gone. Interactive Brokers removed it on October 21, and outside AI models can now send orders alone.

IBKR launched Autonomous Trading on October 21, 2026. Clients link an outside AI and let it trade.

The scope is wide. The AI scans over 75,000 stocks, ETFs and options.

It places orders with no human sign-off. That is the change, and nothing else about it is novel.

CEO Milan Galik gave the rollout numbers on the third-quarter call on October 15. About 3,000 clients received invitations.

Wider access follows in weeks. IBKR holds $964.7 billion in client funds across 5.58 million accounts.

The core construct is the AI Partition. It is a walled-off slice of a client’s account.

Cash goes in. The AI trades only inside that wall.

Funding is capped three ways. Up to 10% of account equity, a $1,000 floor, a $25,000 ceiling.

The connection runs through MCP, the Model Context Protocol. It works like a plug, with one end fitting the model and the other fitting the broker.

ChatGPT, Claude, Grok and Perplexity all connect. The client picks the model.

Now the permission split, which is the heart of it. The model gets read access to the partition.

It also gets write access to place orders. Read-and-write is what separates this from every assistant product shipped this year.

Order types are deliberately plain. Market, limit, stop and stop-limit, with day, overnight, and overnight plus day.

The product list is guarded. Stocks, ETFs and listed options are in.

Penny stocks are out. So are short sales, margin and leveraged products.

The gate is a probation period. A client must first run review mode: ten trades of $100 or more across two weeks, each one human-approved.

Only then does full autonomy unlock. Ten approvals buys an unlimited licence inside the wall.

One line in the risk notice stands out. IBKR says it will not monitor what the AI does.

The legacy model had a human gate at the end. The AI drafts an idea, a person reads it and clicks submit.

That click did three jobs at once. It checked the logic, it carried the liability, and it slowed everything down.

The autonomous model keeps only the setup. The human writes the rules in advance, and the AI spots and sends.

Two things improve. Speed, because the approval step costs nothing to skip.

And coverage, because nobody watches global markets at 3 a.m. Overnight order types exist in this product for exactly that reason.

Now where the machine case breaks, and IBKR has priced each one into the design.

Risk sizing is the first. The $25,000 cap is not a product decision, it is an admission that nobody knows how these models behave under stress.

Leverage is the second. Margin and short sales are banned from the partition, which means IBKR will not let a model trade borrowed money.

Containment is the third, and it is the only hard guarantee here. If the partition runs dry, trading stops, and the model cannot reach the rest of the account.

The fourth failure mode has no control attached. Correlation.

Thousands of partitions running four or five public models on the same market data react together. The cap bounds one account’s loss, not the aggregate behaviour.

Four pools move here.

The first is IBKR’s own, and it moves up. The broker earns $2.59 per cleared trade, per its third-quarter filing.

An agent that trades more generates more commission. That is the business case, stated plainly.

The second is the advisory layer. Advisors charging 1% of assets to pick stocks now compete with a model that costs tokens and never sleeps.

The third is the robo-advisor. A preset rebalancing model is a static product facing a live one.

The fourth is the product the robo replaced. Each displacement compresses the layer beneath it, and this is the second round.

The adoption evidence is already in. Robinhood opened agentic trading in May 2026.

By the third quarter it reported 200,000 signed-up clients and $1.6 billion in trades. That is from its October call.

IBKR is larger and more global. If 5% of its accounts fund a partition, that is roughly $7 billion of AI-directed cash.

One scoreboard note matters for how this gets counted. AI-managed fund AUM does not track retail AI partitions.

The line between a managed AI fund and ten thousand funded partitions is a label, not a function. The metric will understate this for as long as the naming holds.

IBKR built the first sandbox where an outside model trades real cash with no per-order gate. The cap is low and the guardrails are real.

The door is still open. What shifted is not the amount of money, it is the default.

Per-trade human approval was standard at every broker. It is now a setting with a toggle next to it.

The ten-trade probation is the piece to watch. It is the only remaining human checkpoint, and it is a two-week formality rather than an ongoing control.

Watch the cap. Twenty-five thousand dollars is a research budget, and the first increase tells you IBKR has seen data it likes.

Watch the monitoring line harder. A broker saying it will not observe the AI is a liability position, and the first contested loss will test whether it holds.

Territory: +machines on execution, +incumbents on the perimeter.

Per-trade human approval became a setting, while the cap, the banned products and the partition wall stayed with the broker.