A few starting points for fraud teams exploring Gen AI.
Just presented at ACFE Singapore 2026 on leveraging AI for fraud detection.
The room was full of fraud practitioners — investigators, auditors, risk managers — who understand the threat deeply but are still figuring out how to operationalise AI. That gap is real.
Why this matters now
Lately I’ve been seeing fraud detection openings at companies that never hired for fraud before. That alone tells you how bad it’s gotten.
The FBI IC3 numbers back it up. Reported fraud losses were relatively flat for years — and then LLMs showed up. Losses have grown exponentially from 2020 to 2025, hitting an all-time record last year.
Attackers got an unfair advantage faster than defenders could adapt. Deepfake KYC bypass. Voice-cloned exec impersonation. Autonomous scam agents that A/B test their own scripts. None of this was within reach of the average scammer five years ago.
The good news: defenders get the same tools.
How fraud teams can actually use Gen AI
Here’s the slide I shared at ACFE — where Gen AI shifts each stage of fraud detection:
🔍 Fraud Discovery — threat-mapping with LLMs. Move from reactive detection to anticipatory risk intelligence. Prompt: “Based on these product features, map the fraud risks associated with this product.” Surfaces threats you haven’t seen yet.
⚙️ Fraud Logic Development — rule generation from your own data. Feed the model your data signals plus confirmed fraud cases, ask it to suggest five new detection rules to test. Cuts the rule-development cycle from days to hours.
🏷️ Tagging & Centralisation — agent-led labelling. Give an agent access to your data and tools. It picks the signals to check, produces a labelling recommendation with reasoning, and the analyst reviews instead of starting from scratch. Removes the manual-review bottleneck.
📉 Performance Tracking — always-on monitoring with reasoning. Agent runs daily, queries your performance data, alerts when hit rate drops or a rule stops firing — with reasoning attached.
If you’re setting up fraud detection systems from scratch — or rebuilding for the LLM era — these four are the highest-leverage starting points.
Two questions also came up at the talk — ones every fraud team is quietly asking:
Which LLM is best for fraud detection?
The right LLM depends on your specific fraud use case, your data, and your constraints.
The only way to know is to run evals. Define what good looks like for your fraud problem — accuracy on ambiguous cases, reasoning quality, latency, cost — then test models against your actual data. The model that wins your eval is the right model for you.
Can LLMs catch fraud they’ve never seen before?
Think about it this way — if fraudsters are already using LLMs to brainstorm new attack vectors, we as defenders can use the same capability to ask: where are our vulnerabilities before they find them?
LLMs are already well trained on fraud data. Feed in your fraud taxonomy — your product features, your user flows, your existing controls — and ask the LLM to map where the gaps are. It will surface attack patterns your team hasn’t encountered yet because it’s reasoning from a much broader base of fraud knowledge than any single analyst or team has.
The same tool fraudsters use to plan attacks, you can use to stress test your defences. That’s the defender’s advantage — if you use it well.
The gap between how well fraudsters are using AI and how well defenders are using it is real.
Closing that gap is what I’m focused on.