Dark pool trading prices your order and an AI agent's differently

Around 30% of daily US stock volume now comes from people like us trading our own accounts, per The TRADE’s January read on the year. Most of that flow never touches a public exchange. It goes to a wholesaler, a firm that fills the order itself, off-exchange, and we see a fill price that is usually a hair better than the quote on the screen. The standard story stops there: retail gets price improvement, everyone is happy, free trading works.

The part that goes unsaid is that the wholesaler did not give every order the same deal. Before filling anything, it asked one question about the order in front of it: how likely is it that whoever sent this knows something I do not? That single question, scored in microseconds, is the hinge the entire off-exchange market turns on. And it is about to matter more, because the EU bans paying brokers for order flow from 30 June, two weeks out as I write this during a heavy central-bank week, per IC Markets’ calendar for the week of 15 June.

What dark pool and off-exchange venues score before they fill your order

Here is the mechanism, plainly. When a marketable order arrives at a wholesaler, the firm is about to take the other side of it. If the sender turns out to be right about where the stock is going, the wholesaler loses on the position it just took on. So it estimates that risk in advance. The industry word for it is toxicity: the probability that the order is informed and will move against the firm that fills it. A retail market order to buy 40 shares of a megacap at lunchtime scores as very low toxicity. A burst of correlated orders that consistently precede a move scores as high.

Low-toxicity flow is cheap to fill, so the wholesaler competes for it by offering a small price improvement inside the public quote, a tenth of a cent here, a fraction of the spread there. High-toxicity flow is expensive to fill, so it gets less improvement, gets internalized only partly, or gets handed back out to a public exchange where the wider crowd absorbs the risk. A 2026 peer-reviewed model by Umut Cetin in Mathematical Finance puts a fine point on it: the broker is implicitly comparing wholesalers on toxicity-adjusted terms, which means the score is not a side detail, it is the input.

The SEC’s own economists describe the downstream effect: pulling the uninformed orders off to wholesalers leaves the public exchange holding a more informed, more adverse mix, which widens spreads there. So the segmentation does not just sort our orders. It changes the price on the lit exchange that everyone, including us, references as fair.

When retail order flow starts looking toxic

For two decades, segmentation was a stable, almost boring arrangement. People traded like people, scored as harmless, collected the small improvement. What is new is that a growing share of retail accounts are starting to trade through agents, and an agentic strategy that actually works is, by definition, the thing the toxicity engine was built to detect: orders that precede moves, correlated across accounts, arriving in patterns.

The uncomfortable implication is that the price improvement we have come to expect is conditional on looking uninformed, and a good agent stops looking uninformed. The same flow that earned a rebate last year can be re-scored and quietly lose it, with nothing on the fill confirmation to explain the change.

The EU ban sharpens this. Removing the payment does not remove the sorting. Optiver, a firm that makes markets for a living, flagged the workaround over a year ago: brokers are building single-market-maker venues with an affiliated firm and routing orders there exclusively, which they describe as offering even less competition for retail flow than the system it replaced. The segmentation survives the ban. It just moves inside a closed loop where the broker owns the venue. Banker on Wheels measured what that costs on the European versions of these venues.

Read your order-routing disclosure as your fills shift

The decision lens here is execution quality, with a risk-surface edge for anyone wiring an agent into a brokerage account. The score is invisible, but its inputs are not. We can read our broker’s order-routing disclosure, the quarterly document that names where our orders went and how much price improvement they earned, and watch whether that improvement holds steady or drifts as our trading pattern changes. If an agent is placing the orders, the honest assumption is that its flow will not be priced like a sleepy retail account forever, and any backtest that assumed retail-grade fills is overstating the edge by exactly the improvement that will disappear.

The price improvement we treat as a perk is really a verdict. It says the venue has decided we do not know anything. The day that verdict changes, the fill changes, and nothing on the screen tells us why.

What I keep turning over is the symmetry of it. The better an agent gets at finding an edge, the more it trips the one filter designed to make that edge expensive to execute. The signal and the toxicity score are reading the same orders. I am not sure a strategy can be informed enough to be worth running and uninformed enough to keep its retail fills, and that tension may matter more to the math than the model ever will.