How HSBC uses AI to find suspicious money flows

HSBC’s Dynamic Risk Assessment helps investigators find suspicious banking activity. The bank piloted the system, co-developed with Google, in 2021. Its published deployment results describe a smaller queue of alerts containing more useful leads.[1][2]

Learning what deserves investigation

Google’s AML AI learns from the bank’s customer, account and transaction information, together with records of previous investigations. It produces monthly risk scores for retail and commercial customers, with information about what contributed to each score. The bank supplies the data; Google says the product does not enrich it with Google’s own consumer data.[3][4]

That is a different unit of attention from a single payment. The output concerns a customer’s activity in context. A score can help decide where to look, while its contributing factors give an investigator somewhere specific to begin. Google explicitly cautions that a score between zero and one is not a direct probability that money laundering occurred.[4]

What changed at HSBC

These figures describe HSBC’s monitoring and investigation results, rather than criminal convictions or a controlled experiment. Their practical significance is the combination: more suspicious activity surfaced while fewer alerts required attention.[2]

The queue still needs organising

Google’s investigation guidance describes two ways to turn scores into work. A team can select its highest-ranked cases according to available investigator capacity, or investigate everything above a chosen risk threshold. The latter makes the workload depend on the risk found; the former makes capacity an explicit part of the selection.[5]

A persistent high score creates another ordinary operational problem: the same customer can reappear next month although an investigation is already open. The guidance recommends considering repeat-alert filters when neither the score nor its explanation has changed materially. This prevents a model’s regular output from automatically becoming duplicate work.[5]

Learning from investigations also requires a fair historical test. Google’s process documentation says training months should not overlap the months used to evaluate performance. Otherwise, a model could appear successful by recognising cases it has already learned from. Complete transaction periods and consistent data fields also matter when it begins scoring new activity.[6]

A faster route from detection to action

The next HSBC–Google programme, announced in June 2026, adds generative and agentic AI to financial-crime work. HSBC said it expected the architecture to help it intervene twice as fast after detecting risk. That is an announced objective, separate from the earlier monitoring results.[7]

The useful progression is from identifying a promising lead to helping the investigator resolve it. A better ranking saves attention; a clear explanation focuses the enquiry; a well-organised case process makes the finding usable. Measuring those stages separately shows where AI has improved the work and where another part of the process still determines its speed.[3][5]

Report an error or suggest a correction