.png)
Most financial crime leaders I speak to are not short of enthusiasm for AI. They have seen the headlines - JP Morgan's 95% reduction in AML false positives, NatWest's 135% improvement in scam detection. They know their Boards want an AI story, and the regulators are watching.
And yet, when I ask what specific problem they are trying to solve, too often there is no clear answer.
If you’ve been mandated to adopt AI, start with structured ideation before discussing the specific AI solution or vendor.
The first step is understanding where AI could realistically make a difference in your organisation, because the answer is probably not everywhere. You need to identify the problems most worth solving, where AI is likely to work, and where results can be measured.
The approach we take with clients starts before any conversation about specific technology.
It begins by walking the actual journeys, including onboarding, screening, transaction monitoring, investigation and reporting, with the people who run them. Not the process as it appears in a policy document, but as it is carried out day to day.
In practice, we run this as a structured sprint with the client team, typically in a handful of days. Working through each part of the process, we look for a consistent set of signals.
- Where are analysts spending time on tasks that do not require their expertise?
- Where does quality vary depending on who is doing the work?
- Where is time being spent gathering and assembling information rather than making a decision about it?
- And critically, where is the control weakest, rather than merely the process slowest?
The output is a prioritised view of where AI would make the most material difference, given the organisation's processes, data and governance maturity. Some opportunities will align with common industry patterns. Others will reflect how the firm is structured, where there has been historic underinvestment, or where regulatory pressure is most acute.
One principle matters throughout, and it's having a clear understanding of what measurable benefits the organisation is trying to achieve.
AI does some things extremely well, and better than humans at scale. These include processing and summarising large volumes of information, identifying patterns and anomalies across datasets, extracting and structuring information from unstructured documents, drafting standardised content, and supporting repetitive decision tasks consistently.
Knowing these strengths should point you towards some problems more strongly than others.
The next step is to test your assumptions. If AI can fix your problem, someone else will probably have done something similar already. Look for evidence, either in financial crime or in a related area.
This is the right place to start if you want to implement AI seriously and avoid getting distracted by hype.
Screening alert resolution is a clear use case for AI, particularly where the volume problem is acute. AI can triage alerts before a human looks at them, resolve clear false positives through context analysis, and generate closure narratives for the audit trail.
- Banco do Brasil recently reported reducing false positives by over 90% through AI-assisted screening.
- Société Générale reportedly saved the equivalent of 200 FTE, with 60% of cases requiring no human intervention.
- Santander UK reduced its onboarding cycle from 12 days to 2.
KYC onboarding and document processing is similarly mature. Natural language processing can extract and cross-reference information from structured and unstructured documents, verify identity against registries, and flag anomalies for review.
For compliance teams, SAR narrative drafting is yielding positive results. Generative AI can create a draft SAR from case data, transaction patterns and risk indicators, with an analyst reviewing and finalising rather than starting from scratch. That can improve speed and consistency, provided every draft is reviewed before submission and hallucination risk is actively managed.
QA automation also has significant potential, particularly where AI reviews completed work against defined standards at far greater coverage than traditional sample-based testing.
There are other credible areas too. Horizon scanning, obligations mapping, regulatory change assessment, customer risk scoring, document extraction and transaction monitoring optimisation have all seen success stories.
There are questions around governance to overlay across all of this.
- How can the outcomes be explained?
- How can the model be governed?
- How can the use case be tested with enough confidence to move into production rather than remain a promising pilot?
A useful proof of concept is only useful if there is a credible path into day-to-day operations.
You may still have too many candidates, even after testing for business value and AI fit. At that point, apply a basic scoring exercise. Assess each idea by value, feasibility, risk, data availability, explainability, governance complexity and ease of moving into production.
Then choose one, two or three use cases maximum. Each should have a clear problem statement, expected benefit, accountable owner, success measure and route to BAU.
The best starting point may not be the most exciting use case. It could be the one you have the greatest confidence you can deliver in the shortest time.
That is how risk-aware financial crime leaders can make progress with AI without betting the house on the wrong challenge.
-3.png)
