around the use of AI tools . 5 It is characterized by periods of optimism , followed by unrealistic expectations , eventually reaching a more balanced state of realistic capabilities . 6 As AI continues to dominate the headlines and interweave itself into our daily lives , we may be close to the “ peak of inflated expectations ,” where expectations outpace reality .
FIs are no exception to this AI hype cycle . AI tools are regularly being adopted by the banks to assist in transaction monitoring , fraud detection and alert triage . The expectation is to remove human surveillance and drastically reduce false positives . The reality , however , is that AI is not a “ cure-all ” and presents a few key concerns . AI tools , like traditional rulebased and manual methods , are prone to generating false positives and false negatives . As AI approaches the peak of inflated expectations , FIs must exercise caution and evaluate this trend of replacing human surveillance with AI .
AI certainly has a role to play in the ongoing battle against financial crime , particularly in pattern recognition . However , a significant concern is the lack of contextual intelligence and nuanced judgment that experienced AFC investigators bring to the table . AI tools simply cannot match the human ability to consider the “ bigger picture .”
For instance , imagine a scenario where a customer , John Doe , is structuring a series of low-value cash deposits , each just under the threshold of $ 10,000 , but spreading them across accounts that appear to be owned by different businesses , such as “ John ’ s Appliances ” and “ Doe Electrical .” At first glance , the transactions may seem legitimate and the entities unrelated . However , an AFC investigator with knowledge of the local business environment and customer behavior would likely raise a red flag , recognizing that both businesses share the same ownership and are located close together . How did AI handle this event ? It failed to alert on this scenario . Why ? AI , relying on historical patterns without considering the deeper context ,
AI certainly has a role to play in the ongoing battle against financial crime , particularly in pattern recognition
missed this connection between the accounts . Score one for the human investigator .
Another challenge with AI is its lack of interpretability . AI algorithms like neural networks 7 are “ black boxes ”; they can achieve high accuracy , but understanding how they arrived at their final decision can be quite challenging . 8 Neural networks aim to mimic the human brain functionality , and they consist of numerous interconnected layers and nodes , each performing calculations on the data . 9 The issue is that these calculations are not human-readable explanations .
In transaction monitoring , a neural network might classify a transaction as suspicious based on subtle patterns in that data such as a high transaction amount , frequent activity or an unusual time of day . Here is the catch : These patterns are buried in the network ’ s weight matrices , making it nearly impossible to explain why the transaction was flagged . This lack of transparency does not exactly win any points with the regulators , particularly in an environment where FIs must stay compliant with ever-evolving regulatory standards . On the other hand , an AFC investigator can explain specific factors that influenced their judgment while triaging alerts ― such as inconsistencies at the account level , breached monetary thresholds , cross-referencing customer records and contacting interested parties . Convinced yet ? Human investigators bring clear , context-driven reasoning that the regulators can trust . Score another for the human .
We have covered the important attributes we risk losing when replacing human surveillance with AI . Does this mean FIs should ditch AI tools and go back to relying solely on human investigators for alert triage ? No . Instead , the ideal solution lies in a hybrid approach where AI tools assist human investigators . This ensures that human judgment and interpretability remain integral to the decision-making process . Augmentation , not replacement , is a strategy that leverages the best of both worlds .
Cara Wick , CAMS , FRM , Global Financial Crimes executive , Charlotte , NC , USA , cjoany @ gmail . com ,
Cameron Taylor , AI and ML Solutions advisor , SAS , NC , USA , Cameron . Taylor @ sas . com ,
1
“ Karl Popper : Philosophy of Science ,” Internet Encyclopedia of Philosophy , https :// iep . utm . edu / pop-sci /
2
Jim Richards , “ OCC Comptroller Talks About AML ‘ False Negatives ’ and Technology ,” RegTech Consulting LLC , February 5 , 2020 , https :// regtechconsulting . net / aml-regulations-andenforcement-actions / occ-comptroller-talksabout-aml-false-negatives-and-technology /
3
Rules to detect suspicious activity in transactions generally meet the regulatory definition of a model , because they transform inputs ( i . e ., data ) into something used to make a decision ( i . e ., file a suspicious activity report ). For more information , see : “ Interagency Statement on Model Risk Management for Bank Systems Supporting BSA / AML Compliance and Request for Information ,” Office of the Comptroller of the Currency , April 12 , 2021 , https :// www . occ . gov / news-issuances / bulletins / 2021 / bulletin-2021-19 . html
4
“ SR 11-7 : Guidance on Model Risk Management ,” The Federal Reserve , https :// www . federalreserve . gov / supervisionreg / srletters / sr1107 . htm
5
“ Explore Beyond GenAI on the 2024 Hype Cycle for Artificial Intelligence ,” Gartner , https :// www . gartner . com / en / articles / hype-cycle-forartificial-intelligence
6
Ibid .
7
“ What is a neural network ?” Elastic , https :// www . elastic . co / what-is / neural-network
8
Ibid .
9
Ibid .
ACAMS Today | March – May 2025 19