LeoVegas SVP says AI trading without human oversight creates risk

Andre Zammit made the argument plainly at SBC Summit earlier this month: the operators buying the fastest data feeds on the market and stopping there are building something that will eventually break. Zammit, Senior Vice President of Sportsbook at LeoVegas Group, was on a panel about live data and odds pricing when he said something that cut against the prevailing pitch most technology vendors make to the industry.

The prevailing pitch is speed. Get the data first, suspend the market faster, reprice before the arbitrageurs move. Zammit's position is that speed without the infrastructure to process it creates friction, not advantage. That is a different claim, and it has structural implications for how operators are pricing the value of AI in their trading operations.

What he described at LeoVegas is a layered system: automated pricing engines, automated liability engines, and what he called smart bet delays, all sitting between the raw data feed and the customer-facing product. The point is not to slow things down. The point is that unprocessed speed is a liability rather than an edge, and operators who have not built the back-office tooling to match the front-end data investment are carrying risk they have not priced.

The part of Zammit's argument that the industry has been slower to accept is the human element. The firms that have moved furthest on AI trading have tended to present it as a replacement story: fewer traders, faster decisions, lower overhead. Zammit inverted that framing. The investment he described is in retraining traditional traders as model operators and risk strategists, people who understand what the machine is doing and can intervene when it is wrong. That is a more expensive model than pure automation, and it implies that operators who have cut trading desks to the bone may have optimised for the wrong variable.

I have seen this pattern before in a different market structure. When automated systems are given full execution authority over positions with no human review layer, the errors they make are not random — they are systematic, because the model's blind spots are consistent. The damage compounds before anyone notices. A human watching the book does not need to be faster than the algorithm. They need to be present when the algorithm encounters something outside its training distribution, which in live sports betting happens more often than the clean backtests suggest.

The question the industry has not fully resolved is how to price that human layer against the margin compression that automated trading is supposed to deliver. Zammit did not give a number. No one does, because the counterfactual, what a fully automated book loses in a correlated shock event, is not in anyone's public reporting.

LeoVegas operates across the UK, Germany, and the US, three jurisdictions with meaningfully different regulatory requirements around in-play betting and responsible trading obligations. The infrastructure argument Zammit made is also a compliance argument: a trading engine that moves faster than its own risk assessment does not satisfy regulators who want to see bet acceptance decisions made with adequate controls. That pressure will increase as more jurisdictions formalise requirements around automated decision-making in wagering.