AI, AML 5 min read
Taktile leaders share the latest advancements in AI-driven AML with ACAMS Today Magazine
This blog is adapted from an ACAMS Today Magazine feature by Taktile’s Maximilian Eber, Co-founder and CPTO, and Dustin Eaton, Principal of Fraud & AML.
In an ACAMS Today Magazine feature, Taktile’s Maximilian Eber, Co-founder and CPTO, and Dustin Eaton, Principal of Fraud & AML, argue that the shift from rule-based to AI-driven AML systems is quickly becoming a practical advantage.
Today’s AML programs catch less than 1% of the estimated $2 trillion laundered globally each year. And even with financial institutions spending around $206B annually on compliance, results still fall short.
That gap sheds light on why AI is gaining momentum: it can spot patterns better, improve accuracy, and reduce manual effort.
With growing confidence in AI, regulators are now encouraging AI adoption in AML for the first time.
Eber and Eaton lay out a practical roadmap for responsible AI implementation in AML. They explain why AI is uniquely suited for today’s AML challenges and where it is delivering the biggest impact. They also outline what regulators expect from an AI governance perspective, and how teams can reduce adoption risk. Finally, they share the factors that influence the success of most AI initiatives.
Why AI is uniquely suited for today’s AML challenges
Eber and Eaton emphasize to ACAMS Today that the evidence for AI in AML has moved from theory to measurable results in real-world programs.
According to a Guidehouse validation study, 61% of financial institutions reported risk reduction after implementing AI and machine learning in AML.
A major reason is pattern recognition. AI can detect complex behaviors and patterns that human analysts might miss when reviewing cases one-by-one.
AI can also organize structured data like customer account history and unstructured information like adverse media, helping analysts decide which alerts are worth investigating with greater scale and speed.
Eber and Eaton also emphasize AI’s proven ability to reduce the day-to-day workload that compliance teams typically face. For example, deploying AI in transaction monitoring has led to 50–70% lower false positive rates than using traditional rule-based systems. This can create meaningful savings or free up capacity for higher-risk investigations.
From transaction monitoring to sanctions screening, AML teams are gaining efficiency and lowering risk with AI. These teams tend to find success when they treat AI models as governed assets, managed with similar rigor to credit or capital models.
Read the full ACAMS Today Magazine publication.
How AML regulations have expanded to safe AI deployment
As Eber and Eaton outline in the ACAMS Today piece, AI governance in AML mostly builds on familiar model risk management frameworks, rather than introducing entirely new requirements.
A key reference point is the Federal Reserve Board’s SR 11-7 guidance on Model Risk Management. Although issued in 2011 with the OCC, its core ideas still apply to modern AML systems.
The three core components remain the same:
- Conceptual soundness: Understand how the model works, why it should work, and why it fits the use case.
- Ongoing monitoring: Track performance over time to ensure the model keeps performing as expected and spot when recalibration is needed.
- Validation: Require independent review by qualified personnel not involved in development, focused on effectiveness, limitations, and assumptions.
Embedding agentic AI capabilities, however, adds an extra layer of complexity:
- Explainability: AI models are more complex than rule-based systems because they combine many signals at once and weigh them differently depending on the context, rather than following a single “if X, then Y” rule.To provide explanations that support transparency expectations, teams can use techniques such as SHapley Additive exPlanations (SHAP), which estimate how much each input factor contributed to a specific prediction, or Local Interpretable Model-agnostic Explanations (LIME), which approximates the model’s behavior around one specific case with a simpler, easy-to-explain model.
- Dynamic learning: Models that re-train more often need stronger monitoring to manage drift. This is because real-world behavior changes over time, and a model that performed well in the past can become less accurate as fraud patterns, customer behavior, and data quality evolve.
- Data quality: AI performance depends heavily on data quality and representativeness, making data lineage, quality controls, and bias testing central to governance.
Taken together, these governance practices make AI adoption safer, more auditable, and easier to defend to regulators without slowing down innovation.
Moving beyond the limitations of legacy AML systems
Legacy AML programs were built for a different era: lower volumes, slower typology evolution, and systems that could be tuned with rules alone. But today, teams are expected to do more than “check the box.”
Understanding where AI delivers value in AML, and the governance frameworks regulators look for, is only the first step. Implementation is the next challenge: choosing the right use cases, proving performance, building audit-ready controls, and deploying safely without disrupting operations.
In the next part of this series, Maximilian Eber and Dustin Eaton share a practical operational playbook for financial institutions adopting AI-driven AML strategies, including concrete governance practices, validation approaches, and rollout strategies designed to reduce risk.
