AI Won't Replace Your MLRO. It Will Redraw the Job
AI in financial crime is not a productivity upgrade for the compliance team — it is a shift in what the MLRO is actually accountable for, from doing the checks to owning the machine that does them.
The MLRO carries some of the clearest personal accountability in a Jersey firm, and AI is arriving in that role faster than in most others. The risk is not that AI gets the financial-crime work wrong. It is that it gets it right often enough that nobody keeps watching.
Anti-money-laundering work is where AI meets one of the most demanding parts of Jersey's regulatory reality. Customer due diligence, transaction monitoring, sanctions screening — these are not efficiency problems a firm can quietly optimise. They are the controls on which the island's reputation as a well-regulated international finance centre rests, and the MLRO carries personal weight for them.
The JFSC's July 2026 guidance addresses AI in compliance processes directly, and its message to the financial-crime function is not “slow down.” It is that standards must be maintained or improved, with human oversight proportionate to risk and a named person responsible for final decisions in higher-risk cases. This piece is about what that redrawn job looks like.
Jersey’s financial-crime function has always operated under a particular kind of scrutiny: the island’s standing as a well-regulated centre depends on it getting these controls right, and the MLRO carries that weight personally. AI arrives into that seat not as a neutral tool but as a change to how the control itself works — which is why the guidance treats AI in compliance processes as something to be governed, not merely adopted.
The efficiency that quietly lowers the bar
AI is very good at the volume work of financial crime: clearing the obvious, ranking the ambiguous, surfacing the patterns a human reviewer would take hours to find. That is real value, and I have seen it turn an overwhelmed alert queue into a manageable one. The risk is what happens next — the queue gets quieter, the team gets smaller, and the assumption sets in that the model is simply better than the people it replaced.
A financial-crime control that gets more efficient has not necessarily got more effective. The JFSC guidance names exactly this risk in its fraud-detection scenario: a model degrades, false negatives climb, and nobody notices because the absence of an alert looks identical whether the system is working or failing. Efficiency is visible immediately; a lowered detection bar is visible only in the case that gets through.
What the guidance actually requires of an AI-assisted AML control
The JFSC does not ban AI from AML work — it sets the conditions under which it is acceptable. Its illustrative onboarding scenario is a template a Jersey firm can adopt directly: AI flags missing information and red flags, and the firm keeps the final decision. Around that sit specific, checkable requirements.
1. A skilled human on the higher-risk path
The guidance requires higher-risk cases to be routed to a suitably skilled person, with a competent person remaining responsible for high-impact or unusual cases. AI can triage; it cannot be the decision-maker on the cases that matter most.
2. Live data behind the model
It requires sanctions lists and watchlists to be kept current and source documents validated. An AI screening tool running against a stale list is not a control — it is a false reassurance, screening confidently against the wrong reality.
3. A record that survives challenge
It requires records of inputs, outputs and overrides. When a regulator or auditor asks why a case was cleared, “the system did it” is not an answer. The override log is where the MLRO's accountability becomes evidence.
The two errors that cost differently
Financial-crime AI makes two kinds of mistake, and treating them as equivalent is a governance error in itself. A false positive flags a clean customer — an efficiency cost, a frustrated relationship, wasted review time. A false negative lets a case through that should have been stopped — a control failure that can become a reportable event.
The guidance's answer is to monitor false positives and false negatives and recalibrate on performance degradation. In practice this means the MLRO must own a metric most firms do not yet track: not how many alerts the model raised, but how many it should have raised and did not. The first number is easy and comforting. The second is hard and is the one the job now turns on. The guidance’s language is deliberate here — it asks firms not just to run AI in compliance but to maintain or improve standards, which is a bar a firm can only claim to clear if it is measuring the failure it cannot see.
The dependency the whole control now rests on
There is a second-order risk in AI-assisted financial crime that the MLRO owns whether or not they chose the tool: the control now depends on a system that can fail without warning. The JFSC guidance treats a third-party AI outage as a live scenario — the onboarding tool goes down, red flags stop surfacing, and the process runs on regardless, with the AML control quietly switched off.
For the financial-crime function this reframes resilience as a compliance obligation, not an IT one. The guidance's remedy is to treat the tool as a critical dependency with a manual fallback, and to pause or limit the activity when it is unavailable rather than let onboarding continue blind. The question the MLRO must be able to answer is not “is the tool reliable?” but “what is our control the moment it is not?” A function that cannot describe its manual fallback has not adopted AI safely — it has simply moved its single point of failure somewhere it can no longer see.
The redrawn job: from reviewer to model owner
The MLRO's accountability does not shrink when AI arrives — it moves up a level. The old job was to ensure the checks were done well. The new job is to ensure the machine doing the checks is understood, monitored, and challengeable — and to remain the named human who owns the decisions it informs. Table 1 sets out that shift so a board can see what it is actually asking of the role.
Table 1 — How AI redraws the financial-crime seat: the same accountability, exercised over a machine.
| The MLRO still owns | But now exercises it by | Guidance anchor |
|---|---|---|
| Final decisions on risk | Owning the human-review path for higher-risk cases | Suitably skilled person on final decisions |
| The quality of the control | Monitoring false negatives, not just alert volume | Monitor false pos/neg; recalibrate |
| The integrity of the data | Ensuring lists and watchlists stay current | Keep sanctions/watchlist data current |
| The evidence trail | Keeping records of inputs, outputs, overrides | Records that survive audit and challenge |
| The response to drift | Setting escalation thresholds and retraining | Recalibration on performance degradation |
The pattern I return to is that AI does not make the MLRO's job easier. It makes it more senior. The person who once could point to a completed checklist now has to be able to explain, monitor and defend a model — which is a harder skill, and one the guidance's people clause insists the firm must actually resource, not assume.
None of this is an argument against AI in financial crime. Used well, it lets a stretched function focus its scarce human judgement on the cases that most need it — which is exactly what the guidance’s risk-based approach intends. The argument is that the MLRO’s accountability is the one thing AI cannot absorb, and that a function which lets efficiency quietly become the goal has misread what the tool is for. The point of automating the volume is to make room for better human decisions on the cases that matter, not to remove the humans who make them. That is the distinction between a financial-crime function that has adopted AI and one that has quietly abdicated to it, and it is a distinction the MLRO cannot delegate to the model.
If your AI stopped flagging the cases it should stop tomorrow, would your financial-crime function notice from a metric it already watches — or only from the case that finally got through?
Next in the series: the front office — what AI-assisted advice and suitability mean when the customer is on the other side of the decision.
Sources: JFSC, Guidance on the use of AI in Jersey's financial services sector (July 2026) — Principle 3 (AI in compliance processes: CDD, transaction monitoring, sanctions screening), the illustrative AML onboarding scenario, and the fraud-detection model-drift scenario; Government of Jersey et al., Jersey AI Playbook (2026).
Professional disclaimer: The views expressed are my own and are shared for general informational purposes. This content is intentionally non-confidential and does not include proprietary, client-specific, or employer-sensitive information. It is not legal, regulatory, investment, or professional advice; readers should seek appropriate advice for their circumstances.
Financial services leader (20+ years) | Lean Six Sigma Green Belt | London School of Economics (LSE) AI Leadership Accelerator
See where your firm stands
The free diagnostic scores your firm across the twelve operating-model instruments in about eight minutes.
Take the diagnostic →A version of this article was first published on LinkedIn as part of the Sequence body of work. This is the canonical version.
The views expressed are the author's own and shared for general information. This content is intentionally non-confidential and does not include proprietary, client-specific or employer-sensitive information. It is not legal, regulatory, investment or professional advice. © 2026 Paul Soares. Sequence is the trading name of Sequence Advisory Limited (in formation) · hello@sequenceadvisory.com