Agentic AI: The New Weapon Against Financial Crime Faces Costly Hurdles
Agentic AI could slash KYC processing time by 70%, but high human oversight costs and governance issues slow adoption, says McKinsey.
The battle against financial crime is entering a new era as banks turn toward agentic AI financial crime prevention tools to streamline their operations. According to a recent report by McKinsey & Company, this advanced form of artificial intelligence has the potential to drastically reduce the time spent on know-your-customer (KYC) processes by anywhere from 30 to 70 per cent.
Currently, the burden of compliance is immense. The report reveals that many banking institutions dedicate between 10 and 15 per cent of their entire workforce solely to anti-money laundering (AML) and KYC activities. While agentic AI offers a way out of this labor-intensive cycle, the path to full adoption is blocked by significant governance and financial obstacles.
The Accuracy Gap and Governance Struggles
One of the primary concerns for financial institutions is the ability to measure and trust the reasoning of AI agents. McKinsey notes that banks often lack the necessary metrics to evaluate how these agents behave and make decisions, even though such evaluations are relatively low-cost because they do not require labeled data.
To illustrate this, the report cites a bank that tested its customer onboarding process through ten different evaluations. The bank focused on four key pillars: accuracy, processing time, cost, and consistency across repeated runs. The institution set a performance floor of 80 per cent for accuracy and 70 per cent for consistency.
The results were mixed. While consistency ranged from 72 to 100 per cent, the weaker end of the spectrum revealed a worrying trend: in some cases, roughly one out of every four runs produced a different answer to the exact same input.
The Hidden Cost of Human Oversight
While the technology itself is powerful, the cost of keeping it "safe" is staggering. The report highlights a surprising disparity in the operational costs of AI-driven banking services. In a typical customer service task:
- Token costs (the actual cost of running the AI) account for only 20 to 25 per cent of variable-run costs.
- Human oversight accounts for a massive 70 to 75 per cent of the cost.
Furthermore, the report points out that 60 per cent of the cost associated with an agentic task is spent on refining the answers rather than generating the initial response. This suggests that the AI's first attempt is often insufficient for the high-stakes environment of banking.
Strategies for Optimization
McKinsey suggests that these costs could be slashed by more than half through strategic optimization. Proposed measures include:
- Deploying cheaper AI models for simpler, low-risk tasks.
- Removing duplicate operations within the workflow.
- Minimizing data transfers between different AI agents.
- Moving from a model of checking every single output to a sample-based quality review system.
The Compliance Dilemma
However, the shift toward sample-based reviews is not a simple technical upgrade. In the highly regulated world of banking, reducing the frequency of quality checks involves complex risks. The report warns that such decisions must be driven by compliance and risk management frameworks rather than a mere desire for technological efficiency.
As banks move toward more autonomous systems, the traditional "human-in-the-loop" model may no longer be sufficient for supervisory purposes, making reliable oversight an increasingly difficult challenge.
"In financial-crime prevention, agentic AI has moved from concept to production faster than most institutions expected—and the value is real. But we are still at the beginning of the process."