AI financial crime investigation means using software agents to gather evidence, assess alerts and draft decisions on suspected money laundering, fraud or sanctions breaches. The technology is arriving quickly. The harder question is who signs off on the result.

Two launches in October 2026 answered that question in different ways. One agent will not close a case until an analyst confirms it. The other is reported to reach a decision in 17 seconds, and has not said publicly how much human review sits around it. For a founder building in this space, that difference matters more than the model underneath.

What AI financial crime investigation involves

When a transaction-monitoring system raises an alert, an investigator has to work through four jobs:

1. Assemble the data: transactions, customer history, documents.
2. Assess it: is this normal for this customer?
3. Decide: close the alert, escalate, or file a report.
4. Record why: so another person, or a regulator, can follow the reasoning.

FIS cites an estimate that analysts spend 40 to 60 percent of investigation time assembling data before analysis can begin. That is why this category attracts so much attention. It is the same evidence problem we covered in AI in payment dispute management.

Two designs for AI financial crime investigation

FIS made its AML Investigations Agent commercially available to US banks on 15 October 2026, with Amalgamated Bank as design partner. According to its release, every output is a recommendation, and no case can be closed without explicit analyst confirmation. Each statement links back to its source record, and the recommendation, the reasoning and the analyst’s decision are logged. Amalgamated’s Sean Searby says “the investigator remains accountable for the final determination.”

Starling Bank has deployed an agent built into its core banking engine. The Paypers reported on 21 October 2026 that it takes an alert through investigation, audit trail and decision in a single pass, with a decision within 17 seconds. The article says Starling has not made public how much human review surrounds those decisions.

Neither design is wrong by default. A confirmation step costs time. A native decision costs some of the ability to challenge it.

The Sign-Off Test for AI financial crime investigation

Whether you are building or buying, test any AI investigation agent with four questions.

1. Source: can the reviewer see where each statement came from?

A summary the reviewer cannot trace is a claim, not evidence. FIS says each statement links to its source record. Ask any tool you evaluate to show the same.

2. Record: is the reasoning logged, not only the result?

An audit trail that stores “case closed” is thin. A useful one stores the recommendation, the reasoning and the decision, so the case can be reconstructed later.

3. Override: who can overrule the agent, and how?

Someone with authority should be able to reverse or reject an output without a support ticket. If no one can, the agent is the decision-maker, whatever the marketing says.

4. Answerable: who answers to the regulator?

The institution is generally the one that answers for its compliance decisions, not the software vendor. The product should make that person’s job easier, not obscure it.

Applying the test to what is public

Question	FIS agent	Starling agent	Tangos
Source	Stated: statements link to sources	Not stated	Not stated
Record	Stated: recommendation, reasoning and decision logged	Audit trail reported	Not stated
Override	Stated: analyst confirms every case	Not public	Review possible, not stated as required
Answerable	Stated: investigator accountable	Not public	Not stated

“Not stated” does not mean absent. It means the public material does not say. For a buyer, that gap is a question to ask. For a founder, it is a gap to close in your own materials.

What founders should take from this

• Build reviewability in from day one. Source links and an audit trail are how an analyst can sign with confidence. They are not extras.
• Know which buyer you are selling to. Some institutions will want a confirmation step. Others will want native decisions. The product, the pitch and the pilot all change.
• Choose your route. You can sit inside an incumbent’s platform, inside an institution’s core, or alongside both. Each decides who you sell to.
• Treat projected savings as a hypothesis. FIS projects more than 90% less time on first-level alert review, about 98% on second-level investigation and more than 80% on report narratives. These come from Amalgamated’s analysis of its workflows, not measured results.
• Name the signer early. Whoever is accountable for the call shapes your product more than your model does. Our guide to identifying the right buyer shows how to map them, and our piece on what makes an AI pilot commercially ready covers why pilots stall without one.

Where to start

Few bootstrapped founders can build a full investigation platform. A narrower entry point passes the Sign-Off Test more easily:

• Evidence assembly for one alert type, with every item linked to its source.
• Case-note drafting that an analyst confirms before it is saved.
• Audit-trail tooling that records reasoning for tools a bank already runs.

Pick one entry point for AI financial crime investigation, then speak to five people who do this work. Our guide to customer discovery for an AI startup covers how.

Frequently Asked Questions