AI inside the electronic trading stack: where it fits, what it improves and what controls it requires

Three quarters of financial firms already use AI, yet only two per cent of reported use cases run without a human in the loop. Both numbers tell the same story: AI is entering electronic trading as components inside governed systems, not as an autonomous trader. The practical question for a desk is architectural — where to insert AI, at what latency, and behind which controls.

This paper maps the three insertion points — support tools, input modules and system logic — and explains why the nearest value sits in the input modules: liquidity estimation, hit-rate and fill-time models, price and volume forecasts and pricing support. It shows how the latency budget decides which model families are even eligible, and why the controls that matter — collars, limits, kill functionality, surveillance — stay outside the model, on frameworks trading firms already run.

The full paper sets out the use-case map, the latency tiers, the control stack and the staged path from first signal to bounded production — shadow first, one desk at a time, scaled only while the evidence holds. Download the full breakdown.