The app for independent voices
I just published my latest deep dive on Agentic AI in financial crime fighting
After breaking down how these systems actually work in production environments, it is clear that the shift is not about better models but about rebuilding the entire operating system behind fraud and AML.
Most financial institutions still run on static rules, fragmented data pipelines, and human-led investigations that do not scale, which is exactly why false positives dominate workflows, analysts spend time on navigation instead of decisions, and emerging fraud patterns are detected too late to matter.
What is changing now is the move toward agentic systems that do not just score risk but execute the investigation itself, where AI agents gather signals across transaction history, entity profiles, linked accounts, sanctions data, and external intelligence, then follow structured workflows that mirror how senior analysts think, and finally produce a fully assembled investigation with traceable reasoning and a clear recommendation.
This forces a different architecture.
It is no longer about adding AI on top of rules engines. It is about:
• Context engineering instead of prompt engineering, where the system decides what data matters before the model generates anything
• Workflow orchestration instead of isolated models, where each step in the investigation is modular, auditable, and aligned with SOPs
• Human-in-the-loop decisioning, where agents do the work but accountability remains with operators
• Continuous rule adaptation, where systems learn from false positives and network-level signals rather than static thresholds
The most important shift is structural.
Legacy systems optimize for detection. Agentic systems optimize for resolution.
That changes how risk teams are built, how decisions are audited, and how institutions think about liability, because once an agent can recommend or execute actions, the question is no longer whether fraud is detected but who is responsible for the outcome.
At the same time, new failure modes emerge.
Agentic attacks, synthetic identity amplification, and adversarial manipulation of decision systems create a new surface area where attackers target the logic of the system itself rather than just transactions.
This is why evaluation frameworks, citation validation, and model oversight become core infrastructure, not optional layers.
The result is a redefinition of financial crime operations.
