AI Agents Can’t Fight Fraud They Can’t See

The FIS-Anthropic partnership signals a turning point for financial crime investigations. But agentic AI is only as accurate as the identity data underneath it. Ken Michie, CTO of Fideo Intelligence, explains how AI agents could help or hinder financial crime investigations, how fraudsters are using them, where current detection stacks fall short, and what banks and fintechs need to get right before putting them to work.

Q. The FIS-Anthropic partnership is getting a lot of attention. What does it actually signal for where the industry is headed?

Ken: It signals that the industry has moved past debating whether AI belongs in financial crime workflows. Agentic AI is going to be a game changer. These investigations are time consuming, and regulatory requirements are deep, so it is logical that the industry will move toward a model where agents do the heavy lifting, and human in the loop make the final decisions on the cases that matter.

The move itself is not surprising. We are moving from a period of deploying AI analytics to one of automated AI operations. FIS, and everyone else, has a mandate to create its own version of agentic operations. The fact that FIS is doing it with Anthropic is big news and a strong indicator they are going to make rapid progress. Previously, AI in financial services was used mainly for scoring and recommendations. This signals a transition toward agentic infrastructure.

That is a meaningful operational shift. But it also raises a question that is not getting enough attention: What happens when an AI agent works from incomplete or fragmented identity data? An AI agent is only as good as the intelligence it works from.

Q. How does this change what “good” looks like for AI in financial crime investigations?

Ken: The opportunity is to combine rich data with agent reasoning so these systems can make sense of what they are seeing. The moat is not around the agent as much as it is around the data and capabilities the agent can access. Agents are notorious for hallucinating, so the goal shifts to deterministic data that grounds their reasoning in fact, not fiction. Humans need to be in the loop. You do not want agents making decisions on every single AML investigation.

Q. Where does Fideo fit in this shift to agentic AI?

Ken: We see the opportunity to leverage our deterministic identity graph to power investigations beyond AML and into fraud detection. That means expanding investigators’ toolkits so they can see patterns and connections within the identity graph, while also building compliance-related agents that help tie together the facts.

Our rich identity backbone, spanning public data through to hidden dark web sources, gives customers a wide range of intelligence for fraud detection. Agents make this faster, simpler, and more effective by iterating through APIs and MCPs until they arrive at a conclusion to present.

Q. What does “incomplete identity data” actually mean in practice, and why does it matter so much for AI agents?

Ken: Incomplete identity data means relying on static attributes like name, address, SSN, and date of birth without the dynamic signals that show how an identity actually behaves in the real world. It is the difference between looking at identity in isolation and looking at it in relation to the networks, devices, accounts, and behaviors it connects to, often based on onboarding data that was accurate 18 months ago and never enriched. For human investigators, that makes decisions slower and harder.

For AI agents acting autonomously, it means being confidently wrong at speed and scale, clearing risky cases and flagging legitimate ones. The quality and continuity of identity signals is a performance issue, not a hygiene issue. If you feed agents stale, fragmented data and ask them to move faster, you simply accelerate the consequences. Strong identity data means real time behavioral signals, network level relationships, dark web intelligence tied to real customer profiles, and risk attributes that travel with an identity across onboarding, payments, and compliance.

Q. How does Fideo address this, and where does it fit relative to what FIS and Anthropic and others are building?

Ken: The FIS Anthropic model is a powerful reasoning and investigation engine that can pull evidence, evaluate it against known typologies, and compress case review. That is real and valuable. What it needs is an identity intelligence layer comprehensive enough to make that reasoning accurate, and that is what Fideo provides. Verify operates at the upstream edge, delivering sub-second identity risk decisions at onboarding and payments by focusing on identity integrity rather than static checklists.

Signals operates in the middle office, giving human and AI investigators deeply contextual, real-time intelligence with structured risk attributes, relationship indicators, exposure insights, and network connections. Our dark web monitoring turns fragmented adversary intelligence into structured signals tied to real customers, providing early warning on synthetic ID campaigns, account takeover activity, and mule recruitment. FIS and Anthropic are building the agent’s reasoning capability. Fideo is the identity intelligence that makes that reasoning accurate. You cannot logic your way to the right fraud decision if you are missing the signals that show what is actually happening with that identity.

Q. What should banks and fintechs be thinking about as they start operationalizing AI in fraud and compliance workflows?

Ken: First, separate the agent question from the data question and answer the data question first. Many organizations are focused on how to get AI agents to do the right thing, but they are jumping to that question before they’ve solved for the quality and continuity of the identity signals those agents will rely on. Without a strong, deterministic data foundation—clean, structured, verifiable identity signals—AI agents will be prone to inaccuracy no matter how well-designed the workflow. You cannot get an agent to reason correctly if the inputs it’s reasoning from are fragmented, unreliable, or the agent has room to hallucinate.

Equally important is capturing industry-specific knowledge within the agent workflows themselves. Generic automation tuned to basic tasks will underperform in fraud and compliance contexts because these domains require nuanced judgment about patterns, risk signals, and regulatory intent. The agents need to be built with that domain expertise baked in, not bolted on afterward.

Next, organizations must move beyond an onboarding mindset. Account opening checks are table stakes, but they do not address the coming wave of synthetic identity fraud. The institutions that are ahead are building continuous identity intelligence, a dynamic, relationship-based view of customers that updates with every interaction across every channel, so agents can reason accurately about payments, account compromise, and coordinated activity.

Finally, stop treating fraud and compliance as separate problems. Synthetic identities are both fraud and compliance failures, and mule accounts are both fraud vectors and AML exposures. The intelligence layer has to work across both domains at once. Over the next 18 months, institutions that build the right identity intelligence foundation will pull away from those that bolt agents onto fragmented data and wonder why the results fall short.

—

Before you deploy an AI investigator, see what your identity intelligence stack is missing. Fideo works with banks, fintechs, and payment platforms to strengthen identity intelligence and improve fraud stack ROI. No roadmap required.

Talk to the team ->