Patience separates the successful projects from the failed ones. Banks keep data scientists in place and let them train models the slow way, rather than deploying a chatbot and calling the job done.
Sven Stumberger, the author of The Next Wave of Global Anti-Money Laundering Enforcement, has worked in financial crime since 2003, across more than 150 institutions in 60 countries, and now heads the Americas practice at Forward Global. He raised the pattern in conversation in Conductors - AI Against Financial Crime podcast, for its second episode: banks are running the same script with artificial intelligence that they ran with transaction-monitoring systems two decades ago.
In the early days of transaction monitoring, the pitch from every vendor followed the same script, a polished presentation and a promise of deployment within three months. The typical result took two and a half years, and half the detection scenarios did not work as designed. Each time a regulator flagged the system as inadequate, the bank upgraded to a newer one that produced a larger volume of output. Better results did not always follow.
Stumberger has applied machine learning to financial crime work since around 2015, before the industry started calling it AI. His team called it automation.
One firm launched a proprietary AI tool for internal audit twice, in 2024 and again in 2025, under two different chief executives and, by Stumberger's account, almost the same press release both times. He said he does not believe anyone still uses it. Months of investment produce a model that checks the Office of Foreign Assets Control (OFAC) sanctions list for updates, a task most banks' existing sanctions-screening vendors already handle through an automatic feed.
What the successful projects have in common is patience. Banks keep data scientists in place and let them train models the slow way, rather than deploying a chatbot and calling the job done. Stumberger's clearest example is related to network detection: identifying which small share of a bank's customer base transacts with which other customers, and surfacing those networks faster than a human analyst would. That is an improvement to the back end of a system already generating large volumes of false positives from customer screening and transaction monitoring. It speeds up the analysts' work rather than replacing it.
Stumberger defines judgment as a decision made against a small number of rules and principles. Once those rules are defined clearly enough, a model can be trained to apply them. What remains unresolved is responsibility. He compares the situation to the debate around self-driving cars, where the open question concerns regulation. Who is liable when the system, rather than the driver, makes the call?
A separate obstacle has nothing to do with model performance. A head of anti-money laundering compliance at a large bank can be responsible for a team in the hundreds, a position with real weight inside the bank's hierarchy. Overseeing two AI systems instead carries less weight, regardless of how well those systems perform. Stumberger points to that incentive, alongside the difficulty of winning regulatory buy-in for a new model, as reasons full adoption moves slower than the technology allows.
Asked to score, from one to five, the statement that AI in compliance is still mostly hype, Stumberger gave it a three: part of the industry has moved past the hype phase and is deploying the technology properly. He gave a two to the statement that AML investigators will work exclusively with AI within three years, citing the timeframe and the incentive to protect headcount.
Discover customer stories, case studies, and insights.
