What Singapore's AML Leaders Are Getting Wrong About AI - And How to Fix It

There's a lot of excitement about AI in compliance right now. Vendors are promising it will transform your AML operations. Consultants are writing papers about it. And compliance leaders across Singapore are under real pressure to show they're keeping up.

But here's what's actually happening on the ground: most firms aren't failing at AI because the technology doesn't work. They're failing because they're implementing it in the wrong order.

Here are the three mistakes we see most often - and what to do instead.

Mistake #1: Treating AI as the solution to a process problem

When compliance teams are drowning in alerts, the instinct is to reach for AI to reduce the volume. Understandable. But if your underlying screening process is broken - wrong data, inconsistent application, no clear escalation path - AI doesn't fix it. It accelerates it.

We've seen firms deploy sophisticated models on top of messy data and fragmented workflows, only to find themselves with a faster version of the same problem. The alerts are still noisy. The decisions are still inconsistent. And now there's a machine involved that nobody can fully explain to an auditor.

The fix: before you automate, standardise. Get your screening process into a single, consistent workflow. Define what a good decision looks like. Then let AI accelerate that - not replace it.

Mistake #2: Forgetting that MAS cares about outcomes, not inputs

It's easy to get caught up in the sophistication of your AI stack. But regulators aren't scoring you on technology. They're asking a simpler question: can you demonstrate, for any given customer or entity, exactly what was checked, what was found, and what decision was made - and why?

That's an audit trail question. And it's one that a lot of AI implementations make harder, not easier. Complex models can feel opaque. If your compliance team can't explain a decision in plain language, that's a problem when MAS comes asking.

The fix: design for explainability from day one. Every alert, every review, every decision should be captured in a structured case record that any auditor can follow. The AI should surface the risk. The human - and the documentation - should own the decision.

Mistake #3: Underestimating how much false positives cost you

Most teams know false positives are annoying. Fewer realise just how much they're costing.

When analysts spend the majority of their day dismissing irrelevant alerts, two things happen. First, the real risks get less attention - buried under noise, genuine hits get slower reviews and sometimes get missed entirely. Second, your best people burn out. Compliance analyst turnover is high, and alert fatigue is a major driver.

The root cause is usually poor entity matching. Legacy tools flag anything that looks remotely like a name on a watchlist, without the context to distinguish a genuine match from a coincidence. Common names, transliterated characters, and outdated records all generate noise that humans then have to clean up manually.

The fix: invest in entity resolution before you invest in anything else. If your screening tool can link names, aliases, jurisdictions, and related entities before surfacing an alert, your analysts only see the matches that actually matter. That's not just an efficiency gain - it's a risk management improvement.

What getting it right looks like

The compliance teams in Singapore that are navigating this well share a few things in common. They have a single platform for their daily operations - not four tools stitched together with spreadsheets. They've defined clear human accountability for every decision AI supports. And they've built their audit trail into the workflow from the start, not bolted it on afterwards.

The good news is this isn't out of reach for most firms. It doesn't require a massive technology overhaul. It requires the right foundation: consistent data, structured workflows, explainable decisions, and a screening tool that reduces noise rather than generating it.