What 1M live AI alerts tell us about how fincrime teams adopt AI

Everyone wants AI agents to close alerts for them. Almost nobody has turned that on. That gap opened The AI Files, Unit21's virtual summit on AI in fraud and AML, where CEO Tyler Allen shared what about a year of live AI alert reviews shows about how financial crime teams actually adopt AI. The short version: the technology is ready, the regulatory confidence is not, and the teams getting the most value go deep on one workflow before they go wide.

‍

Tyler has a close view of this. He was Unit21's first hire, wrote much of the core product as a software engineer, and brought Unit21's AI products to market as head of AI. Unit21 has run transaction monitoring, fraud monitoring and case management for eight years, for about 200 enterprise customers in 90+ countries, and about 5% of US SARs are filed through Unit21 today. So the data covers the full life of an alert: detection, the L1 investigation, escalation and, for some, the SAR. It is aggregated across customers with no customer PII shared.

‍

Here are the five things the data shows, and the three steps Tyler recommends before you deploy.

‍

1. Everyone wants auto-close. Almost no one uses it.

Automation in alert review works like automation in cars. Lane assist and cruise control are automation too, long before the car drives itself. Unit21 sees the same levels in fraud and AML:

• Assist: AI does the research and an analyst decides. About 84% of usage.
• Auto-review: the agent recommends a decision for a human to approve. About 16%. Teams have to watch for bias here, since a recommendation can anchor the reviewer.
• Auto-close: the agent closes the alert on its own. About half a percent.

‍

Demand for auto-close is real, and so is the value. Adoption is slow because of regulatory uncertainty, which is the theme behind the next finding. For more on moving up these levels safely, see progressive autonomy: the trust ladder for AI in compliance and how to safely auto-close alerts with AI.

‍

2. Accuracy is not what's holding AI back

A year or two ago, the worry was hallucination. Models have improved, and so have the products built around them. In the data Tyler shared, AI agreed with human reviewers 99% of the time on sanctions alerts and 95% of the time on transaction monitoring alerts. Those are strong numbers even next to your best analysts.

‍

What holds teams back from deploying AI at scale is fear of how regulators will view it. That is a solvable problem, and the third step below is how to start solving it. We covered how Unit21 handles the hallucination question in 4 ways we've engineered around the AI hallucination problem.

‍

3. Deep adoption beats wide adoption

At Money20/20, Tyler heard a tier one bank's CTO describe a "let a thousand flowers bloom" approach: give everyone AI, measure adoption, skip ROI. He then asked the same bank's AML lead about AI on the show floor. The AML team hadn't deployed any.

‍

That pattern is common. General-purpose tools such as Copilot or chat assistants aren't designed for AML or fraud work, so financial crime teams get left behind while the company reports broad adoption. Unit21's data points the other way: teams that focus on one workflow first get much more value, and that early win proves to the rest of the organization that AI can work and pay off. If you're choosing that first workflow, read where to start with AI in AML.

‍

4. AI isn't just more efficient. It's more effective.

This is Tyler's favorite finding. The business case for AI is usually efficiency: fewer manual tasks, fewer false positives, smaller backlogs. The data shows it also finds more bad activity. AI-worked alerts found true positives about 2x as often as human-only alerts overall, about 15x in fraud workflows and about 7x in enhanced due diligence reviews.

‍

The reason is simple. AI isn't racing a backlog and doesn't get alert fatigue. Tyler shared the story of one organization reviewing check deposits against a simple threshold, with a false positive rate around 99%. After a week of clearing roughly a thousand false positives, the team missed the one real fraud in the pile. An agent works the thousandth alert as carefully as the first, which is better for the institution and for the people on the other end of the fraud.

‍

‍

5. Same agent, completely different jobs

AI agents aren't set-and-forget, and they aren't one size fits all. A lot of investigation work is shared across use cases. Adverse media checks show up in sanctions, fraud and EDD reviews alike, and structuring checks recur across AML workflows. Those building blocks get teams about 90% of the way out of the box.

‍

The last 10% is yours: your SOPs, your risk tolerance, your products and your customers. The teams that succeed are the ones who can update their own agents as fraud schemes change, new advisories come out and new products launch, the same way you iterate with any AI assistant. More on that in why configurable AI is the only AML AI compliance teams will trust.

‍

Three steps before you deploy AI agents

Baseline one queue

A common mistake is skipping a record of how the program performs today. Pick one queue, such as EDD, and answer the basics: how analysts perform, whether SOPs are current and followed, where the data lives and how long reviews take. Without that baseline you can't show before and after, and AI layered on a broken process won't fix it.

‍

Build a fixed test set

Set aside about 100 verified alerts per use case, enough for meaningful results. This golden eval set lets you test any vendor or internal build consistently, and re-test every time you update an agent, adopt a new model or add an investigation step. We walk through it in how to build your first AML AI agent and prove it works to an examiner.

‍

Prepare your AI evidence file

Since regulatory uncertainty is the real blocker, know what you'll show examiners before they ask: how the AI works, what data it accesses and which models run under the hood. That file can come from your vendor, your own team or both. For the bigger picture, see building the accountability framework for AI in financial crime.

‍

What comes next

The keynote opened a day of sessions that followed the same arc, from spotting a real product behind a great demo to what examiners will ask once AI is live. The last session turned to the people behind the alerts; read the takeaways in block it, flag it, share it: frontline tactics for pig butchering and sextortion.

‍

The full findings, broken down by industry, are coming in The State of Agentic FinCrime.

‍

Frequently asked questions

How many fincrime teams let AI auto-close alerts?

Very few. In the live alert data Tyler presented, about 84% of AI usage was assist, about 16% was auto-review, where the agent recommends a decision, and about half a percent was auto-close.

‍

Is AI accurate enough to review AML and fraud alerts?

In the data presented, AI agreed with human reviewers 99% of the time on sanctions alerts and 95% of the time on transaction monitoring alerts. The bigger blocker to scaling AI is regulatory uncertainty, not accuracy.

‍

Does AI find more fraud or just work faster?

Both. AI-worked alerts found true positives about 2x as often as human-only alerts overall, about 15x in fraud workflows and about 7x in EDD reviews, because AI doesn't get alert fatigue or run out of time.

‍

What should a team do before deploying AI agents for AML or fraud?

Baseline one queue, build a fixed set of about 100 verified alerts per use case for testing, and prepare the AI evidence file you will share with examiners.

‍

‍

Figures are from Unit21's analysis of live AI alert reviews over about a year, aggregated across customers, as presented at The AI Files on November 3, 2026.

‍