A shorter alert queue proves little unless suspicious activity is still caught. Three respondents explain how to test AI against a baseline, log every decision and check what the model suppresses.

Transaction monitoring teams have long drowned in alerts, most of which turn out to be harmless. AI promises to shrink the transaction monitoring queue by scoring alerts before an analyst sees them. However, every alert a model suppresses is one a regulator may later ask about. So the real question is not whether AI cuts the backlog, but whether a firm can prove it still catches crime. Our earlier piece on AI fraud models found a similar gap between scale and judgement.

We asked one question: has AI reduced the alert backlog in your transaction monitoring team, and how do you prove its decisions to auditors and regulators? We wanted AML compliance officers, financial crime operations leads and RegTech founders, with alert or false positive figures. First-hand results mattered more than vendor pitches. Three people answered in this round, and we used all three. None of them runs a transaction monitoring team, so we’ve set their advice against published results and guidance.

What AI has done for transaction monitoring so far

The best-known transaction monitoring result comes from HSBC. Using Google Cloud’s AML AI, the bank cut alert volumes by more than 60% while detecting two to four times more suspicious activity, Google said at launch in June 2023. That comparison set reports filed under AI against a period under HSBC’s older rules. Still, Google sells the product, and the figures are now three years old.

US policy is moving in AI’s favour too. FinCEN’s proposed rule on AML programmes, issued on 7 April 2026, would let its director weigh whether a firm uses “innovative and effective compliance tools, including the use of artificial intelligence” before major enforcement, WilmerHale explains. The rule is not yet final, but it signals that AI in transaction monitoring is welcome if it works.

Measure detection, not just alert volumes

Omer Malik is chief executive of ORM Systems, which provides risk services and IT hardware.

“AI can reduce transaction monitoring backlogs by helping compliance teams prioritise higher-risk alerts, spot recurring false positives and recognise suspicious patterns that rule-based systems may miss. But a lower alert count isn’t enough evidence of success. The real test is whether investigators can work cases more efficiently without missing suspicious activity or weakening oversight.

A practical approach is to compare AI-assisted monitoring against an established baseline on three measures: alert volumes, investigator handling time, and the share of confirmed suspicious cases. A system that cuts alerts by 30%, for example, may look effective, but that gain is questionable if it also filters out genuinely suspicious transactions. Historical case testing, independent validation and ongoing monitoring are how you catch that risk.

My rule is to automate prioritisation but preserve accountability. Every AI-assisted decision should leave a traceable record showing the transaction data considered, the model version, the reason for the risk rating, and whether an investigator reviewed or overrode it. Teams should also retest the model after significant changes in transaction patterns, and document the results for auditors. A smaller backlog means little unless you can show detection still works and every important decision can stand up to scrutiny.”

Omer Malik, CEO, ORM Systems

His three measures are the right ones for transaction monitoring, because together they show whether fewer alerts means less noise or less detection. His 30% example is hypothetical, but it makes the point well. The HSBC result is persuasive only because the bank tracked confirmed cases alongside alert volumes. Also, his audit-trail list fits the Wolfsberg Group’s guidance below. However, ORM Systems does not run transaction monitoring, and he shares no figures of his own.

Explain every suppressed alert

Kuldeep Kundal is founder and CEO of CISIN, a technology services firm that builds custom software and analytics systems.

“AI cuts the alert backlog most effectively when it’s used as a pre-scoring layer to filter out noise, as long as every decision is backed by explainable AI that auditors can check. Replacing rigid rules with machine learning models can significantly reduce initial alert volumes, so teams stop wading through thousands of low-level false positives and focus on signals that point to real financial crime risk. But that efficiency gain only counts if the decision logic survives a rigorous regulatory audit.

To prove decisions to auditors, the systems we build use explainable AI that produces a feature-importance score for every automated outcome. When a regulator asks why a transaction was suppressed or flagged, the system has to produce a structured log showing which data points, such as historical velocity, entity reputation or unusual geography, carried the most weight. That moves the process away from a black box towards a transparent, repeatable method with a clear audit trail. Building those explanations into the workflow is the surest way to satisfy legal and risk officers.

In AML compliance, defensibility comes first and efficiency second. The strongest set-ups add a human review step for a random share of suppressed alerts. That feedback loop shows regulators the AI isn’t just clearing a backlog to hit a target, but is correctly identifying the highest risks.”

Kuldeep Kundal, Founder and CEO, CISIN

His most useful point is the random check on suppressed alerts. A model that hides alerts can also hide its own mistakes, so sampling what it drops is how a firm finds them. In addition, feature-importance logs give an examiner a reason for each decision rather than a score alone. His original answer said this was the only way to satisfy risk officers, which went too far, so we’ve softened it. CISIN builds systems rather than running transaction monitoring itself, so we’ve also reworded his “we rely on” line. He gives no alert or false positive figures.

Start with better data

Marc Pamatian is a finance and bookkeeping specialist and founder of Chief Bookkeeping Officer. He focuses on the data behind transaction monitoring.

“Balancing regulatory compliance with operational efficiency is always a challenge. Over the years, I’ve noticed that the balance between detecting financial crime and managing alert volumes is where many organisations struggle. High false positive rates drain resources and pull attention away from genuine risks. What’s worked well in my experience is a data-driven approach that focuses on refining the algorithms and improving the quality of the data going in. Collaboration between AML compliance teams and technology providers matters too, so the tools meet practical needs rather than just ticking regulatory boxes.

I’ve also seen success with cross-functional teams that combine tech expertise with compliance experience, which often leads to more precise and useful insights. Regulatory frameworks keep changing, so firms need approaches that scale and adapt without adding needless complexity. Tools that cut false positives while improving detection have become essential.”

Marc Pamatian, Founder, Chief Bookkeeping Officer

His focus on input data is a fair point. A model fed poor or incomplete customer data will score alerts badly, however advanced it is. Mixed teams of engineers and compliance staff also fit the Wolfsberg Group’s call, below, to balance model risk against financial crime risk. However, his answer stays general. He gives no alert or false positive figures and names no tools. We could not verify his firm or its work in AML either. Chief Bookkeeping Officer appears to be a bookkeeping business, so treat his view as general advice on transaction monitoring rather than first-hand results.

Regulators want AI they can follow

Industry guidance now spells out the bar. In August 2025, the Wolfsberg Group of global banks published its second statement on effective monitoring for suspicious activity. It rests on three pillars: transition and validation, balancing model risk against financial crime risk, and explainability. Meanwhile, the EU’s new AML authority plans direct supervision from 2028 and has started collecting data to test risk models, FinTech Global reports, citing vendor Napier AI. Therefore, firms whose transaction monitoring cannot explain a suppressed alert will struggle, however short their queue.

What works for AI in transaction monitoring

The answers and guidance point to a clear order of work. First, set a transaction monitoring baseline for alert volumes, handling time and confirmed cases before switching anything on. Next, clean up the input data, as Pamatian argues, then run the model alongside existing rules and compare outcomes, as Wolfsberg’s transition pillar suggests. Then log the data, model version and main drivers behind every decision, plus any human override. Finally, sample a random share of suppressed alerts for human review, and retest after major changes in customer behaviour. Ownership checks feed these systems as well, as our roundup on beneficial ownership rules shows.

What to watch next

Regulation will set the pace for AI in transaction monitoring, and so will evidence from transaction monitoring teams themselves. Watch whether FinCEN finalises its AML programme rule, and how far it credits AI tools in practice. Also watch the EU authority’s testing of transaction monitoring models as it prepares for direct supervision. In addition, fresh results from banks would help, since the most cited figures date from 2023.

Above all, transaction monitoring with AI works only when a firm can show its reasoning. Fewer alerts are welcome, but the proof regulators want is that the right ones still reach a human.

FintechBits covers payments, banking and fintech developments for readers in the US and UK. This article is for information only and is not financial or legal advice. Views expressed are those of the FintechBits editorial team.