For Banks and Fintechs, Start With the Alerts Your Rules Already Produce

Most AI advice for financial services assumes a vendor, a budget line and a year. Here are three a capable internal team can start on now, with data the institution already holds. ## 1. Cut false positives in transaction monitoring Take three months of alerts your existing rules produced and label what actually happened to each one. Most institutions find a large share of alerts came from a small number of rule patterns that almost never catch anything real. You do not need a new system to act on that. You need to know which rules are generating noise, and then tune them. If you want to go further, a model trained on the labelled set will rank alerts by likelihood so analysts work the top of the queue first. The measure is simple. Same number of analysts, more genuine cases caught. ## 2. Triage collections instead of treating everyone the same Pull twelve months of delinquent accounts and what eventually happened to each. Segment by what worked: a reminder, a restructure, escalation, or nothing at all. Then score current accounts against those patterns. The point is not to predict perfectly. It is to stop spending the same effort on an account that will self cure as on one that needs intervention today. ## 3. Cluster your complaints before you automate anything Take the last quarter of complaints and cluster them by underlying cause rather than by the category somebody ticked. Most institutions find a handful of process failures generating a large share of contact volume. Fix those processes first. Automating a broken journey only means customers reach the dead end faster. ## The step before any of the three Write down which AI tools are approved and which data may go into them. One afternoon. Circulate it. Without that you have BROAI, which is capable staff quietly using their own tools on regulated customer data because nobody gave them an approved option. No model result is worth that exposure, and in this sector the exposure is a regulatory matter rather than an embarrassment. This is the Scan and Score phase of SWITCH, the framework I use with clients. Establish ground truth before building anything on top of it. Pick one. Give it an owner and a number. Thirty days later, keep it or kill it. If you want this run inside your institution with your own data and constraints rather than as a generic workshop, the [speaking and training page](/speaking) sets out the formats, or reach me at me@martins.com.ng or through the [work with me page](/work-with-me).