At the start of the year, I believed every alert artificial intelligence touched in financial crime compliance needed human approval. AI could gather the evidence, summarize the case and draft the narrative. An analyst would review the result and click to close or escalate.
Compliance runs on accountability. A named reviewer on every case made that accountability easy to show.
Faster Teams, Supportive Regulators and a Better Measure of Oversight
Three things changed my mind.
Customer teams moved faster than the approval step allowed. Once AI took over documentation, one payments company cut the time it spends writing case narratives by 60%. A wealth management firm with more than 1 million customers cut SAR filing time by 25%. The approval click stayed on every case, but on the thousandth low-risk alert in a queue, a click carries very little judgment.
Regulators have stopped resisting AI in compliance. FinCEN and the European Union’s new anti-money laundering authority now explicitly encourage advanced technology in AML/CFT programs, while accountability for every AI decision stays with the institution. That puts the design question in the industry’s hands.
Override rates gave us a better measure of oversight. When investigators overrule an agent more often than expected, the agent’s performance is slipping. Because an agent documents every step, quality assurance (QA) teams can also review every case, compared with the 5% to 10% sample most programs check today. A click on each alert tells you someone looked. An override trend tells you there is a problem. A full QA review shows whether the system works.
Humans on Highest-Consequence Decisions
Now we are placing people where the consequences are highest. Humans own the decisions that change a customer’s financial life, such as account closures and SAR filings. Agents handle evidence gathering, documentation and lower-risk work.
We treat each AI agent like a new hire. It follows the institution’s own procedures, it logs every step it takes, and a person controls the outcome of every investigation. AI quality checks can review every case against those procedures.
We give every customer the same advice. Document the full investigative workflow, start with the simplest use case and expand from there. Compliance teams own the configuration themselves, so they can adjust limits without waiting on engineers.
Taking the click off routine alerts frees analysts for the decisions that need them, such as account closures, SAR filings and emerging threats. That is where human judgment matters most.
