A modern dealing desk is a real-time optimisation problem: exposure, toxicity, LP fill quality, and per-client route. Every AI vendor claiming this ground reduces to one of four archetypes. Here is which archetype does what.
The dealing desk is the single densest ROI surface inside a retail broker. Better toxicity scoring alone can shift a firm's P&L by seven figures a year. Which is why every liquidity bridge and RMS vendor now has 'AI' in the deck.
The neutral view: three vendor archetypes each move a specific part of the desk, and the fourth — bespoke build — is where the compounding advantage sits.
Per-order toxicity scoring at sub-second latency.
Fill quality, markout, rejection rate across LPs — actioned, not just dashboarded.
Rules the dealer can approve, not just observe.
Real-time exposure across FX, indices, crypto.
PrimeXM, oneZero, Centroid, Gold-i native paths.
Does the model learn your book, or ship generic?
Bridge-native analytics with light ML add-ons.
RMS-adjacent AI overlays.
Ex-bank / ex-hedge quants selling desk models.
Bespoke AI systems, engineered for the firm's own P&L and regulator envelope.
| Criterion | PrimeXM / oneZero / Centroid built-in analytics | Gold-i / Your Bourse / Match-Trader RMS AI modules | Boutique quant shops | KJ Capital |
|---|---|---|---|---|
| Real-time flow classification | Basic | Basic | Deep | Deep + operator UI |
| LP scorecard actioned | Observation | Partial | Model only | Action + attribution |
| Per-client A/B routing | Manual | Rule-based | Model only | AI recommendations, dealer-approved |
| Model tuned to your book | No | No | Yes | Yes |
The compounding advantage on a dealing desk lives in models tuned to your book, delivered with an operator UI a dealer will actually use during the session. Bridge analytics are the base. RMS add-ons are the quick-win. Boutique quants deliver models. Bespoke build partners assemble the whole desk stack — and are usually the shape a serious broker converges to.
The £120k Dealing Desk AI productised build is designed for exactly this — eight weeks to a live copilot, then continues on the AI Operator retainer.
No. PrimeXM, oneZero and Centroid provide the data layer; AI sits on top of them, not underneath.
Materially. In our experience, per-desk P&L uplift from a firm-specific model runs 15–40% above generic vendor add-ons in the first six months.
Dealers only trust a system that shows its work. Every KJ Capital dealing-desk build ships with per-recommendation reasoning, not black-box scores.
Eight weeks from kickoff to a monitored, dealer-approved system in production.
Firm VPC or firm cloud tenant. Data never leaves your boundary.
The £15k AI Diagnostic is a two-week engagement that produces a costed build plan mapped to your stack, your regulator and your P&L.