Key Takeaways
- FIS partnered with Anthropic to launch a Financial Crimes AI Agent powered by Claude, cutting anti-money laundering investigation times from days to minutes for early adopters BMO and Amalgamated Bank.
- Agentic AI systems move beyond rules-based fraud detection by autonomously analyzing transaction histories, behavioral patterns, and communication logs in real time.
- The Monetary Authority of Singapore is running a proof-of-value program with five banks and government agencies to train shared AI models for pre-emptive scam detection. FIS has built an AI agent that can cut anti-money laundering investigations from a multi-day slog to a matter of minutes, and two real banks are already running it. The Financial Crimes AI Agent, built on Anthropic’s Claude and announced this week, is live at BMO and Amalgamated Bank, with FIS signalling that fraud prevention is next on the roadmap.
How Agentic AI Reshapes Fraud Detection
Traditional fraud detection runs on static, rules-based logic: flag the transaction after it happens, generate a lot of false positives, inconvenience legitimate customers, and move on. Agentic AI breaks that pattern. Rather than handing investigators an alert to act on, these systems work like autonomous analysts, pulling transaction histories, behavioural patterns and communication logs simultaneously and acting on what they find.
The core technique is behavioural baselining. The system builds a model of normal activity for each customer or entity, then watches for deviations: an unusual payment destination, a login from an unexpected location, a vendor invoice that doesn’t match prior patterns. That kind of signal is hard for a rules engine to catch cleanly; it’s exactly what a trained model can surface at scale. Trustmi’s behavioural AI platform takes this approach to ACH fraud specifically, cross-referencing historical payment data, vendor communications and invoice patterns to catch authorised-push-payment scams before the money moves.
The Monetary Authority of Singapore recently launched a proof-of-value initiative with the Government Technology Agency, the Singapore Police Force, and several banks to pool anonymised historical transaction data and train shared AI models for pre-emptive scam detection. The goal is identifying higher-risk transactions across institutions before customers lose money, rather than reconciling losses after the fact.
The “AI vs. AI” Fraud Battleground
The same generative AI capabilities banks are deploying defensively are being used offensively. Fraudsters are producing convincing deepfakes, synthetic voices and forged documents at scale, making social engineering attacks faster and cheaper to run. If you want to understand the financial exposure this creates, the scale of deepfake-linked fraud losses in 2025 is sobering reading.
Banks are responding by training models specifically to spot synthetic content, analysing metadata inconsistencies, formatting anomalies and document provenance rather than just the face value of what’s submitted, but detection alone isn’t enough. In regulated environments, every blocked transaction needs to be auditable. That’s why “explainable AI” is becoming a hard requirement rather than a nice-to-have: compliance teams need to reconstruct why a system flagged something, not just accept that it did.
Legislative pressure is building alongside the technical response. The House Financial Services Committee recently advanced several bills targeting AI-enabled financial crime, including the Bank Fraud Technology Advancement Act of 2026, which calls for studies on advanced fraud detection technologies. JPMorgan Chase has also committed nearly $14 million in philanthropic investments to consumer protection efforts, including an AI-powered platform designed to detect and report text-message scams in real time.
What Comes Next for Financial Security
The direction of travel is toward what the industry is calling a “unified risk view”: deep neural networks and knowledge graphs that assess customer behaviour across every product line simultaneously, producing a single risk score per interaction rather than siloed alerts from separate systems. It’s architecturally more complex, but it’s the only way to catch fraud that deliberately moves across product boundaries.
Agentic AI in payments is already past the proof-of-concept stage. Mastercard recently demonstrated a live AI-driven transaction using its Agent Pay authentication system, pointing toward a near-term future where AI agents initiate, authenticate and settle payments autonomously. That puts the security burden directly on the network layer, authentication can’t be an afterthought when the agent is the one approving the payment.
The harder problems are governance and bias. Improved accuracy and faster response times are real wins, but AI models trained on historical data can encode historical biases, and the speed of agentic systems means errors propagate faster too. Human oversight isn’t optional here, it’s the control layer that keeps the whole system accountable. For builders integrating these workflows, the Anthropic Model Context Protocol is worth understanding as a practical foundation for connecting LLMs to live financial data sources. For more on AI agents and automation tools, visit our AI Agents section.
Originally published at https://autonainews.com/fis-anthropic-ai-agent-cuts-bank-aml-investigations-to-minutes-2/
Top comments (0)