Banks detect only 2% of financial crime while spending billions on compliance
Illicit financial activity reached $4.4 trillion globally in 2025, up $1.3 trillion in just two years. Banks are losing the fight against financial crime despite allocating 10% to 15% of their workforce to Know Your Customer (KYC) and Anti-Money Laundering (AML) activities.
This gap between spending and results-the "compliance trap"-stems from a fundamental mismatch. Traditional rule-based systems and manual review cannot match the speed and scale of criminal operations. As payment systems accelerate and crime industrializes, the old model breaks down.
The answer, according to leading compliance technologists, is agentic AI: systems that operate autonomously to detect, investigate, and report financial crime without human analysts driving every step.
The limits of rule-based systems
No-code compliance platforms dominated the past decade. They let compliance teams build detection rules without engineering support. But they created a new bottleneck: the analyst.
In traditional AML operations, up to 95% of alerts are false positives. Building a single Suspicious Activity Report takes four or more days. No-code tooling cannot scale to meet current crime volumes.
Unit21's 2026 relaunch signals the industry shift. The platform moved from a rules engine to an agentic system where AI agents tune detection logic and investigate cases with minimal human intervention.
How agentic AI works in practice
Agentic AI differs from earlier AI tools in a crucial way. Generative AI summarizes data. Analytical AI finds patterns. Agentic AI plans, executes, and adapts sequences of actions toward a goal-like a digital worker investigating a case rather than a chatbot writing a summary.
When an alert enters the system, an AI Investigation Agent follows a structured workflow:
- Signal gathering: The agent retrieves transaction history, entity profiles, risk scores, and watchlist matches across multiple screens.
- Workflow orchestration: The agent executes modular steps matching the bank's procedures-checking alert history, running open-source intelligence searches, cross-referencing sanctions lists.
- Findings assembly: The agent produces a narrative, evidence logs, and a recommended disposition with explicit reasoning.
A human analyst still makes the final call. But instead of starting from scratch, they review and approve the agent's work. This keeps humans accountable while multiplying their capacity.
The engineering challenge: context, not prompts
Building effective AI agents for compliance is harder than writing better prompts. The real work is context engineering-feeding the model exactly the right evidence without overwhelming it.
Large language models process tokens through transformer architecture, where every token attends to every other token. As context grows, the model's attention becomes scarce. Effective agents curate high-signal information to maximize accuracy.
Unit21 uses seven years of human investigations to determine the optimal context for different tasks. They then test agent outputs against investigations completed by top analysts, using a secondary AI model to check quality before humans see the work. This "LLM-as-a-judge" layer catches errors early.
The system also validates citations-verifying that agent claims come from retrieved data, not from the model making things up.
Three ways AI agents fail
Most early deployments collapse not because models are weak, but because guardrails are poor.
The hallucinating investigator: Too much context and open-ended prompts lead models to fill data gaps with plausible fiction. The fix is narrow "atomic agents" with tight decision boundaries.
The over-suspicious agent: Pattern-driven training without grounding produces false escalations-flagging routine internal transfers as money laundering. Agents need context questions built into their logic to avoid jumping to fraud conclusions.
The black box agent: Conclusions that regulators cannot defend. Accurate outputs without a chain of evidence create liability. Agents must pull data deterministically and document findings in structured format.
The speed problem
Instant payment systems move money in seconds. Legacy rule-based systems cannot keep pace. Banks that adopt agentic AI gain a measurable advantage: these systems detect and investigate at machine speed, not analyst speed.
The productivity gain is substantial. Agentic AI can deliver 20-fold improvements over manual work.
Leading institutions are starting with pilot programs to prove impact before scaling. The stakes are clear: the cost of inaction-$4.4 trillion in undetected crime annually-is too high.
The shift from manual execution to AI supervision is not optional. It is the path to compliance that actually works. Learn more about AI Agents & Automation or explore AI for Finance to understand how these systems reshape financial operations.
Your membership also unlocks:
