How AI Will Replace AML Analysts: The Job Regulation Built

On March 11, 2026, the Netherlands Court of Audit published the findings of an investigation into what Dutch banks actually get for the money they spend hunting money launderers. The number that matters is not the cost, though the cost is remarkable: the eight banks in the audit deployed roughly 13,000 full-time employees on anti-money-laundering work in 2024, at an expense of about 1.6 billion euros — up from 1.16 billion in 2021. The number that matters is the one the auditors could not produce. After years of escalating spend and headcount, the Court of Audit found that the benefits of the regime are, in its own framing, unknown, while the consequences for ordinary citizens wrongly caught in the net are serious and documented.

Sit with that for a moment, because it is the whole article. A national audit body looked at an occupation employing 13,000 people, costing 1.6 billion euros a year, and reported that nobody can demonstrate it works.

Every occupation this series has examined so far was created by demand. Somebody wanted a policy priced, a book closed, a contract reviewed, a lead qualified. The work existed because a buyer valued the output, and the automation question was always the same: can a machine produce output that the buyer values as much, for less. The anti-money-laundering analyst is a different animal entirely. The job was not created by demand. It was created by enforcement. Banks did not build compliance armies because customers wanted them or because the work paid for itself. They built them because regulators fined them until they did.

That origin changes the displacement mechanism completely, and it changes it in the direction of speed. When an occupation exists to satisfy a demand, the machine has to be good enough to satisfy that demand. When an occupation exists to satisfy a mandate, and the mandate has never been shown to produce the outcome it was written to produce, the machine does not have to be better than the human at catching criminals. It only has to be cheaper at producing the same unmeasured output, with a paper trail that survives an examination.

That is a far lower bar. It is being cleared right now.

The Occupation: What an AML Analyst Actually Does

Strip away the titles — financial crime analyst, KYC specialist, transaction monitoring analyst, sanctions screening officer, SAR investigator — and the work is a queue.

An automated monitoring system watches transactions against rules and models. It throws alerts. A human being picks up each alert and decides whether it is nothing or something. To decide, the analyst pulls the customer file, reads the know-your-customer record, looks at the transaction history, checks the counterparty, searches adverse media, considers whether the pattern has an innocent explanation, and writes down a conclusion. Most alerts are nothing. The analyst closes them with a rationale and moves to the next one. A small fraction escalate. A smaller fraction still become a suspicious activity report filed with the financial intelligence unit.

The defining feature of this work is not difficulty. It is volume against a template. The analyst is not exercising rare professional judgment on each case; the analyst is applying a documented procedure to a case file and producing a written record that the procedure was applied. The output is a decision plus an audit trail, and the audit trail is arguably the more important half — the bank is being examined on whether it looked, not primarily on whether it found.

Consider what that means. The product of the occupation is documentation of diligence. That is precisely the shape of work that large language models handle well: read a bounded set of documents, apply a written policy, produce a structured, reasoned, well-cited narrative. This series has repeatedly found that the automation-resistant core of a job is the part where a human absorbs ambiguity and takes responsibility for a judgment that cannot be derived from the record. In AML alert review, that core exists — but it is a thin band at the top of a very tall pile.

Here is the size of the pile. The most rigorous public numbers remain the Bank Policy Institute's 2020 analysis of 2017 data, and they are stark: large banks reviewed roughly 16 million alerts and filed roughly 640,000 suspicious activity reports. That is about 96 percent of alerts resolving to no report at all. Then the funnel narrows again — the same analysis found banks received law enforcement feedback on a median of about 4 percent of the reports they filed. Europol's earlier European study found roughly 10 percent of suspicious transaction reports were further investigated, and roughly 1 percent of criminal proceeds were ultimately confiscated.

I want to be careful here, because this statistic is routinely abused. You will see vendors claim that 90 to 95 percent of AML alerts are false positives, cited as if it were a measured 2026 figure. It is not. I could not find a credible primary source for a current false-positive rate, and most of what circulates is marketing recycling the same decade-old inference. What the BPI data actually supports is narrower and sufficient: about 96 percent of alerts produced no report. That is an alert-level fact, not a claim about how many of the people flagged were innocent.

But the narrower fact is enough. An occupation in which 96 percent of the work resolves to "close this, nothing here, and write down why you looked" is an occupation defined by clearing a queue of mostly-negatives against a procedure.

The Money Is Already Moving

The Dutch banking sector is where this is happening first and most visibly, and it is worth being precise about what has actually been announced, because the secondary coverage has badly garbled it.

In October 2025, at a finance conference in Driebergen, executives from ABN Amro, ING, Rabobank and ASN collectively indicated they expect roughly 2,600 money-laundering-check positions to disappear within about two years. That is the cleanest single datapoint in this story, and it is worth naming what it is: an expectation voiced by the executives who run those functions, not a sum of formally announced redundancy plans.

Around it, the individual announcements are real but narrower than the headlines suggest. ING's own March 2026 investor presentation projects a decrease of approximately 1,250 Operations full-time employees in 2026 — Operations, not a company-wide cut — with a footnote specifying that this lands "among others in KYC and associated activities, which account for approximately 6,000 FTEs." ING does not claim to have a 6,000-person AML division; it says KYC and associated activities employ about that many. The bank targets roughly 350 million euros of incremental cost efficiencies in 2026. Separately, and often wrongly merged with that figure, ING filed with the Dutch benefits agency UWV in October 2025 that up to 950 positions in the Netherlands may be eliminated by the end of 2026, attributed to digitalisation broadly, with AML not named.

ABN Amro announced 5,200 cuts, about 20 percent of its workforce, by 2028. The widely repeated "35 percent of AML staff replaced by AI" line is a distortion: the roughly 35 percent figure spans customer service, operations and money-laundering checks together, and "replaced by AI" is an inference rather than the bank's stated cause. ASN Bank is cutting up to 950 further roles and Triodos 250 to 270 by 2028 — both general reorganisations, neither AML-attributed.

I am belaboring the sourcing because the sloppy version of this story — "banks announce mass AML layoffs, blame AI" — is not what the record shows, and the accurate version is more interesting. What the record shows is a sector that has stopped growing its compliance headcount for the first time since the fines started, executives who expect thousands of those seats to be gone inside two years, and an audit body publicly reporting that nobody can prove the function works. That is a function being quietly repriced, not a dramatic firing.

The broader labor signal is consistent. Challenger, Gray and Christmas reported that in March 2026, AI led all stated reasons for US job cuts for the first time, accounting for 15,341 of the month's 60,620 announced cuts — about 25 percent. By May 2026 AI had led all reasons for a third consecutive month, at 38,579 cuts, bringing the 2026 year-to-date AI-attributed total to 87,714 — already above the 54,836 that Challenger attributed to AI across the whole of 2025, when AI accounted for about 5 percent of cuts. In banking specifically, Morgan Stanley's survey of roughly 35 lenders, reported at the start of 2026, projected more than 200,000 European banking jobs — about 10 percent — going by 2030, concentrated in central services including risk and compliance. That estimate was reportedly revised sharply upward during 2026.

There is a second-order effect worth flagging. Debasish Patnaik, a senior partner who leads QuantumBlack at McKinsey, told Fortune in June 2026 that banks are cutting junior analyst classes by as much as two-thirds while sourcing roughly 62 percent of their AI talent from those same cohorts. Treat that as his assessment rather than a published dataset — no methodology accompanies it. But the shape of the claim matters: the entry-level compliance seat has historically been where banks trained the people who eventually exercise the senior judgment that everyone agrees must stay human. Automating the bottom of a pyramid you are still recruiting the top from is a structural problem that shows up in about a decade, long after the cost savings have been booked.

How the Machine Does the Work

The technical claim I am making is narrow: an agent can execute the alert-review loop end to end on routine alerts, and produce a better audit trail than a tired human on their fortieth alert of the day.

None of those five steps is speculative in 2026. The one that used to be the blocker is step three. Rules engines have always been able to fire an alert; what they could not do was read an unstructured customer file, weigh a plausible innocent story against a suspicious pattern, and write down a defensible reason. That was the human moat, and it was a real one for two decades. It is the moat large reasoning models crossed.

The vendor field has organized around exactly this. On May 4, 2026, FIS announced a Financial Crimes AI Agent built with Anthropic, with BMO and Amalgamated Bank named as institutions developing with the agent, and general availability planned for the second half of 2026. The stated claims are the expected ones: AML investigations moving from days to minutes, reduced false positives, better SAR narratives.

I would ask you to discount those claims heavily. That is a press release issued before general availability, not a measured outcome, and there are no published results behind it. What makes the announcement informative is not the performance claim. It is the shape of the product and the identity of the buyers. When a core banking infrastructure vendor of that scale builds financial crime as its opening agentic use case, and names two banks as development partners, the category has been chosen. Vendors do not lead with a use case they think is hard to sell.

There is one line in that announcement that deserves more attention than the speed claim, and I will come back to it: "Investigators will remain in control of every decision."

The Buyer's Side: A Negative Business Case

Automation happens when someone with budget authority decides to pull the trigger. In every other occupation in this series, that decision involved a tradeoff — cheaper, but is it as good? Here, the buyer has been handed an argument that is close to unanswerable.

Jaap van der Molen, who leads financial crime work at ABN Amro, has described the regime bluntly as a negative business case. The arithmetic behind that phrase is brutal: Dutch authorities seized on the order of 400 million euros in criminal assets in 2024, against roughly 1.4 to 1.6 billion euros a year that banks spend looking. Set aside for a moment whether asset seizure is the right measure of the regime's value — it is genuinely not the only one, and deterrence is real if hard to count. The point is that the person who runs the function is publicly saying the numbers do not work, and the national audit office is publicly agreeing that nobody can show they do.

Notice what is absent from that case: any claim that the machine catches more criminals. It does not need to. The function is being graded on documented diligence and cost, and the machine wins on both. This is why I think AML review clears faster than underwriting did. The underwriter has a loss ratio. The bookkeeper has a closed book that either reconciles or does not. The AML analyst has an output that a national audit body just said cannot be evaluated. You cannot defend a job on the quality of an output nobody can measure.

The Moat Everyone Assumes — And Why It Points the Wrong Way

Ask anyone in compliance why their job is safe and you will get the same answer: the regulator requires a human. This is the same liability moat this series found protecting paralegals, where the argument genuinely has teeth because an unauthorized-practice rule names a licensed human. In AML, the moat is weaker than the people standing behind it believe, and the 2026 evidence is actively cutting against it.

Start with the most striking case. In litigation between the Dutch neobank bunq and the Dutch central bank DNB, a court held that DNB was wrong to require the bank to use manual processes for transaction monitoring. Read that again, because the direction is the opposite of the assumed one: a regulator was overruled for insisting on humans. The ruling is from 2022, not 2026, and I would want the specific holding checked before anyone leans on it in a policy argument — but the precedent that a supervisor cannot simply mandate manual review as a matter of preference exists, and it exists in the same jurisdiction now shedding 2,600 compliance seats.

Then look at what the regulators are actually saying in 2026. FinCEN issued a notice of proposed rulemaking on April 7, 2026, with comments closing June 9, that shifts toward an effectiveness-based standard. It is explicitly not an AI mandate. But its fact sheet states that institutions which responsibly experiment with innovative technologies will not incur additional enforcement risk merely for doing so. That is not a green light for autonomy. It is the removal of a specific fear — that adopting AI is itself a finding — which is exactly the fear that kept compliance the last function in the bank to modernize. Meanwhile the EBA transferred its AML powers to the new EU authority AMLA on December 31, 2025, and FATF's horizon scan on AI and deepfakes stresses human oversight and explainability without prohibiting automated decisioning.

So the regulatory picture is not "a human is required." It is "you must be effective, you must be able to explain yourself, and someone must be accountable." None of those three requirements is a headcount.

Every item on that list is a reason to keep one accountable signer. Not a floor of analysts. This is the pattern this series keeps finding, and AML is its clearest expression yet: governance requirements do not preserve occupations, they specify the shape of the single oversight seat that survives. "Investigators will remain in control of every decision" is entirely compatible with a tenth as many investigators, each in control of ten times as many decisions.

The Evidence That Cuts Against Me

An honest version of this argument has to confront the numbers that do not fit, and there are two.

The first is the US Bureau of Labor Statistics. Compliance officers, the closest large occupational category, numbered about 418,000 in 2024 with a median wage around 78,420 dollars, and BLS projects roughly 3 percent growth through 2034 — about average, not collapse. Financial examiners, a smaller and more specialized category at about 65,100 with a median around 90,400 dollars, are projected to grow about 19 percent through 2034. That is much faster than average. BLS is not forecasting the end of this work. It is forecasting expansion.

I take this seriously, and it is the same trap the radiology debate fell into. The confident 2016 prediction that AI would end radiology aged badly precisely because the demand for imaging grew faster than the automation of reading it, and because regulators and insurers never approved autonomous interpretation. If you want a reason to think I am wrong, that is the best one available.

But I think the BLS categories mislead here, for a specific reason. Compliance officer and financial examiner are broad buckets that include bank examiners working for regulators, ethics and licensing officers, and environmental and safety compliance staff — categories that grow when the rulebook grows. The transaction-monitoring alert reviewer is a sub-population inside a bucket that is expanding for unrelated reasons. An occupational category can grow overall while its highest-volume, most-templated tier is hollowed out, and the aggregate number hides it completely. That is precisely what the Dutch data captures and the BLS categories cannot: not "compliance shrinks," but "the queue-clearing tier of compliance shrinks while the accountable tier does not."

The second piece of counter-evidence is that no one has published measured results. Every performance claim in this space traces to a vendor. The FIS and Anthropic agent was not generally available when this was written. American Banker ran a headline in 2026 stating flatly that bankers say AI is not eating jobs yet. The word doing the work in that sentence is the last one.

Implementation: How This Actually Rolls Out

The sequence is predictable, because it is the sequence every regulated function follows.

Phase 1 is where most institutions are now, and it is worth being clear that Phase 1 already destroys jobs — quietly, through attrition and unbackfilled requisitions, which is exactly how ING's roughly 1,250 Operations FTE reduction reads. Nobody announces a layoff. The queue just needs fewer people, and the people who leave are not replaced.

Phase 2 is the pivotal one and it is a governance decision, not a technical one. The moment an institution accepts that sampling agent dispositions is adequate supervision — the same standard it already applies to models that make credit decisions affecting far more people — the linear relationship between alert volume and headcount breaks. Everything after that is arithmetic.

Impact Assessment: Who Is Actually Affected

The people at risk here are not the chief compliance officer or the seasoned investigator who knows what a trade-based laundering scheme smells like. They are the analysts in the first two years of the career, sitting in the alert queue, doing the work that was always described internally as paying your dues.

Concentrate the risk properly and it lands on: junior transaction monitoring analysts, KYC refresh and remediation staff, sanctions screening first-liners clearing name-match noise, and the very large offshore and outsourced operations built specifically to clear alert volume cheaply. That last category is the most exposed of all, because it was created for exactly one reason — labor arbitrage on a queue — and an agent is a cheaper arbitrage than a lower-cost geography.

The macro numbers give a sense of scale without pretending to precision. If Morgan Stanley's roughly 200,000 European banking jobs by 2030 is even directionally right, and risk and compliance is one of the named concentrations, the compliance share alone runs to tens of thousands of seats. The Dutch 2,600 is a small country with a large banking sector giving an early, honest reading of a number every European bank is computing privately.

And there is a real human cost on the other side of the ledger, one the Court of Audit named explicitly: the citizens wrongly caught by the regime — debanked, frozen, investigated on a false positive — bear serious documented consequences. If agents genuinely reduce the noise rate, some of that harm goes away. That is a real benefit, and it is not the one the banks are buying. They are buying the cost line. The reduction in wrongful harm is a side effect nobody is being graded on, which is precisely why I would not count on it materializing.

Benefits and Challenges

The benefits are real. Consistency is real — an agent applies the procedure identically at hour eight of a shift. Auditability is real, and probably better than the human baseline, because a model that cites its sources every time beats an exhausted analyst writing a two-line disposition. The speed is real. And if the noise rate genuinely drops, fewer innocent people get debanked.

The challenges are equally real, and I want to name the one that actually worries me. It is not that the agent will miss a launderer — the humans miss them too, at a rate no one can measure, which is the whole problem. It is that automating the alert queue at scale entrenches the regime's core pathology. Right now the cost of the AML system is the only thing forcing anyone to ask whether it works. The Court of Audit report exists because 1.6 billion euros a year is a number that demands justification. Make the same activity cost 200 million and the question stops being asked. The regime becomes cheap enough to never evaluate. We will automate our way past the audit rather than through it, and the citizens wrongly caught in the net will stay caught, more efficiently.

That is a strange thing for a displacement article to conclude, but it follows directly from the evidence. The occupation was built to satisfy a mandate nobody has validated. Automating it does not validate the mandate. It just removes the last constituency with an incentive to challenge it.

What Survives

The accountable investigator survives, and their job gets better. The genuinely hard cases — the ones where a typology is novel, where the innocent explanation is sophisticated, where the answer is not in the file — go to a smaller number of more senior people who now have an agent doing the assembly work that used to eat their day. That is a good job. There will be far fewer of them.

Model governance and typology design survive and grow. Somebody has to decide what the system looks for, validate that it looks correctly, and defend both to a supervisor. This is where financial examiner headcount growth probably shows up, and it is a plausible reading of the BLS number that is consistent with everything else here: the people who examine and govern the systems grow while the people who feed the queue do not.

If you are in the alert queue today, the move is to get to the other side of the sampling decision before it is made. Learn the model governance framework your institution already applies to credit models, because that framework is what will be extended to cover agent dispositions, and the people who understand it will be the ones writing the standard rather than being measured against it. Learn to defend a methodology to an examiner rather than to defend a disposition. The durable skill in this occupation was never reading the alert. It was being the person willing to sign underneath the conclusion and answer for it — and the 2026 evidence says the institution still needs exactly one of those per decision, not one per alert.

The desk clears. The signature stays.

Further Reading

• The straight-through mechanism this article describes, worked through in a neighboring regulated occupation: how AI will replace insurance underwriters
• The strongest version of the counter-argument, where a licensing rule genuinely does protect the occupation: the paralegal liability moat
• The same hollowing pattern in a back office graded on documentation: how AI will replace bookkeeping and accounting clerks
• My falsifiable claim on where compliance headcount lands: AML analyst headcount and the sampling decision by 2027
• The reporting behind the Dutch numbers: the audit that found 1.6 billion euros of unknown benefit