Twenty Risk Patterns, One Dealing Desk: Why We Let the AI Act First

At Match-Prime we deliberately moved that line further toward the machine. HawkEye, our in-house risk system, now takes both core risk actions on its own - restricting the specific trading activity that breaches the terms of business agreed with a client, and hedging our own book - before a dealer has reviewed the case. The dealer supervises and reviews afterwards rather than approving beforehand, with RMS Monitor keeping that oversight live around the clock. Both actions are taken under the terms agreed with the client, on criteria applied identically to every client. Here is why the market left us no choice - and why the design only holds for a provider that owns its risk infrastructure end to end.

The old division of labour stopped working

The arrangement most dealing desks still run was built for a simpler market: fewer instruments carrying real exposure, lighter flow, a slower pace, a narrower range of prohibited practices to watch for. The machine raised alerts, the human reviewed and acted. The market has since grown more complex on every front at once, and two shifts in particular broke the model.

The first is that the window to act collapsed. Trading that breaches the terms of business once built up over days; it now plays out inside a single session, sometimes within an hour. A traditional review cycle - notification, verification, escalation, senior review, restriction - arrives too late to help: it documents the damage rather than preventing it.

The second is structural: the number of cases has outgrown what any desk can review. In the first half of this year alone, HawkEye generated hundreds of thousands of notifications across twenty distinct risk and conduct patterns - from price-feed integrity faults to latency arbitrage and coordinated account activity. A notification is a signal to be examined, not a finding against a client, and the overwhelming majority are closed without any action at all. But no dealing desk can turn that volume into timely decisions through manual review alone.

Speed and volume together rule out doing this by hand. The two conventional responses each fail: manual review is right but slow, and threshold automation - a rule that trips when a number crosses a line - is fast but blunt. What the market demands is automation that is both fast and selective.

Acting on flow that breaches the agreed terms

Our system earns its authority to act autonomously case by case, inside parameters set and approved in advance. For flow of this kind it goes through three stages, each of which filters out what can be safely set aside.

Surveillance surfaces the sessions matching a defined pattern and filters out ordinary trading activity. A statistical layer then reconstructs the account's recent trading and separates genuine patterns from market noise. Only what survives both reaches the AI agent, which reviews a prepared, interpretable evidence package and applies a restriction at trader-tag level - to the one originating trader concerned, not to the broker's account as a whole - without waiting for approval.

In practice, a restriction only ever lands after three independent checks agree it deserves one, not after a single rule trips on a raw number. That is what keeps legitimate client flow untouched. The agent never rules on raw activity, only on qualified evidence; every stage records why it passed a case forward; and every action is logged, reversible, and open to challenge by the client. Where conduct meets the threshold for market abuse, it is escalated through the regulatory channels that apply to it. We describe the full pipeline in an earlier deep-dive.

Hedging our own risk

The hedge works the same way, and it is deliberately kept separate from execution. A provider can manage that risk in ways the client feels - moving the flow to another book, passing fills to an external venue, or adjusting the terms on which orders are filled. We keep the two apart.

A broker's flow is executed in full against us, on the terms that broker is quoted. Separately, on our own book, HawkEye hedges only the portion carrying market risk we should not hold, at the first line of the book across our own liquidity providers. The hedging decision changes nothing about the price, the speed or the likelihood of execution the client receives: we do not route orders by client profitability, we do not apply asymmetric slippage, and we do not reject or re-quote on the basis of whether a position is hedged.
The broker's execution does not change.

The human becomes the reviewer

Because the machine acts first, the human's role shifts from gatekeeper to reviewer. When the gates are satisfied the action takes effect immediately, and the risk team receives the full case at the same moment - charts, results, and the agent's reasoning - to validate, adjust, or reverse after the fact.

This is the correct trade-off. A perfect decision taken tomorrow is not protection; it is a post-mortem. But review only means something if it can happen whenever the system acts, and this activity does not keep office hours. That is the job of RMS Monitor: a live view of flow and hedging in our dealers' hands wherever they are, so a developing situation is visible around the clock.

Why this needs owned infrastructure

You can't buy any of this off the shelf. Restricting activity requires control of the risk system it lives in; hedging at the first line of the book requires control of the liquidity; monitoring the pricing layer beneath a broker requires being that layer. A surveillance tool connected from the outside can observe and alert, and the better ones do it well, but it cannot take these actions, because it does not own the infrastructure they depend on.

Detection, by now, is something any broker can buy. The questions worth asking a provider are about what happens after the alert: how quickly it acts on flow that breaches its terms, how selectively, and who reviews the action once it has been taken.

Match-Prime, together with the Match-Trade Technologies team, will be at Forex Expo Dubai. If you want to see how this might work on your own flow, come find us at booth 76.

---

This is the first in a series on the systems behind Match-Prime's flow management. The next article looks at how flow is segmented across accounts - a setup arranged together with the broker that optimises risk and profit on both sides.

Vladimiros Spanos is Chief Operating Officer at Match-Prime.

Jarosław Klamut, PhD, is Head of Risk at Match-Trade Technologies, a strategic technology supplier to CySEC-regulated liquidity provider Match-Prime.