Contents
Compliance: A Prime Area for AI
A Matter of Balance: Should We Push AI to 100%?
AI: For Better or for Worse?
Reason #1: Only humans can handle complexity
Reason #2: The “black box” effect versus a key requirement: the need for decision explainability
Reason #3 – Bias: the genetic flaw of algorithms
Reason #4 – Liability in the face of non-compliance risks
Reason #5 – Shadow AI: An Invisible Threat
Augmented Intelligence: The Best of All Worlds
Toward the “Augmented Compliance Officer”!
The boom in artificial intelligence continues: global corporate spending on AI models and platforms is expected to grow by 64% by 2026 according to Gartner, while 93% of audit teams report that they are already using it. In France, more than half of large companies have begun scaling up their use cases.
AI is thus gradually establishing itself as a major driver of transformation. This trend is accelerating with the rise of agent-based AI: according to Salesforce, the average number of active AI agents per company has nearly tripled in one year, particularly in regulated and complex industries.
Compliance: A Prime Area for AI
AML-CFT —and compliance in general—offer an ideal setting for developing AI-based use cases in the face of rapidly growing data volumes.
Tangible Operational Gains
First, for monitoring, using natural language processing (NLP) algorithms that scan guidelines and legal texts to instantly identify the implications for the business. Next, AI is valuable for the analysis of documents, generating reports, and even filling out compliance questionnaires.
Volumes of data that are impossible to manage manually
Finally, to meet KYC requirements, AI can quickly analyze millions of transactions and data points, automatically filter against sanctions lists or lists of Politically Exposed Persons (PEP), perform real-time behavioral analysis to detect suspicious transactions, and automate identity verification, thereby reducing false positives. For compliance teams, AI is a real asset: it boosts their productivity and reduces the risk of errors (false positives). In addition, real-time analysis ensures better compliance and better-prioritized alerts.
A Matter of Balance: Should We Take AI All the Way to 100%?
For compliance teams, AI offers a wide range of benefits: it improves productivity and the reliability of controls, reduces time-consuming tasks, and shortens the time it takes to detect and respond to suspicious transactions… Such opportunities encourage teams to push the boundaries of AI’s applications even further. This makes sense, given that the benefits of AI are tangible and proven. But how far should we go in expanding the role of AI in AML-CFT ? Is it conceivable that an algorithm—no matter how powerful, scalable, user-friendly, and comprehensive it may be—could replace compliance teams, with their tasks managed 100% by AI?
AI: For better or for worse?
AI enthusiasts may dream of it, managers may envision it, compliance teams may want it, and executives may demand drastic reductions in compliance costs to improve their operating results—but they will all run up against the wall of reality. And that reality is clear: total automation is an illusion and will remain so for a long time to come. There are several reasons why it is unrealistic to pursue (nearly) total automation of the “ AML-CFT.”
Reason #1: Only humans can handle complexity
This argument may seem to contradict the very nature of AI: isn’t it said that AI is precisely capable of handling complexity? In reality, this confuses what is complicated with what is complex. A complicated problem can generally be broken down (into processes, causal relationships, etc.), which makes it easier to understand and analyze: in this context, AI is entirely appropriate and faster than the human brain. In contrast, a complex phenomenon consists of a large number of interacting entities, making it impossible to predict its behavior or evolution. Hence the need for human expertise to analyze and manage it. AI alone is not sufficient to analyze and understand a complex phenomenon in its entirety.
Compliance isn't just a technical issue
AML-CFT is clearly a complex system, involving interactions between multiple regulatory requirements, organizational processes at varying stages of maturity, and the ever-evolving behavior of fraudsters. While AI can certainly help—for example, by better identifying suspicious transactions—human expertise remains essential to ensuring the effectiveness of AML-CFT. While AI is essential for detection, only humans can interpret the results. For example,two customers who trigger the same alert but present very different levels of risk.
Many gray areas remain
Let’s not forget that regulations—which are constantly evolving—are not applied in a black-and-white manner: interpreting the context of a transaction and the behavior of fraudsters, assessing intent, or evaluating reputational risk require ethical and contextual judgment that is beyond the reach of an algorithm. Regulatory texts establish a framework, but operational reality often lies in gray areas. An algorithm applies rules or identifies statistical patterns; it cannot weigh the context of a business relationship or assess intent. When regulations come up against complex fraud schemes, atypical business relationships, or ambiguous behavior, only humans—and their business expertise—can strike the right balance to ensure compliance.
Reason No. 2: The "black box" effect versus a key requirement: the need for decision explainability
Despite the good intentions of their designers, AI models are not transparent, and it is unlikely that they will be in the future—if only because of competition among the companies that develop them. AI functions like a “black box,” whose inner workings are difficult to understand. However, for regulators, who verify compliance, an AI model that contains gray areas is unacceptable: compliance requires traceability and full explainability of every check. AI AI Act also strengthens the requirements for governance, documentation, human oversight, and traceability for systems subject to its obligations.
Reason #3 — Bias: The Genetic Disease of Algorithms
Trust in AI is undermined by numerous biases. These arise when models produce results that are harmful, unfair, or out of step with reality. They do not stem from any malicious intent on the part of the machine, but rather from the data it is fed and the way it is designed.
They can be grouped into three main categories:
- Biases Related to Training Data : AI learns from historical data. If this data is of poor quality or incomplete, the model simply replicates and amplifies it. And it is difficult to account in real time for changes in behavior or the emergence of new money-laundering methods, as perpetrators are not lacking in imagination when it comes to achieving their goals.
- Measurement Biases : These arise from the choices made by developers when creating the AI tool. They occur, for example, when the metric chosen to evaluate a concept is flawed or biased, when a single model is applied to groups with very different characteristics, or when the model is based on invalid correlations between two independent events.
- Usage-Related Biases : These stem from humans’ excessive tendency to blindly trust an algorithm’s decision (“ “If the AI says so, then it must be true! ” ”), to the detriment of their own critical thinking. This attitude creates feedback loops : the AI makes a biased decision that generates new biased data, thus creating a vicious cycle.
Reason #4—Accountability for Non-Compliance Risks
Only humans can incur the company’s legal liability before regulatory authorities. An algorithm cannot bear the consequences of a decision to file a suspicious activity report, freeze assets, or address deficiencies in KYC/KYB/KYT processes…
Strict regulatory requirements
Thus, a compliance decision must be understandable, explainable, and defensible before an auditor or regulator. Legal responsibility cannot be delegated to an algorithm, given the increasingly strict requirements for justification, the ever-growing number of ACPR inspections, and the proliferation of internal audits.
Reason #5—Shadow AI, an invisible threat
Shadow AI is “ the informal and covert individual use of an AI solution not approved by the organization, with the employee concealing their practices, operating outside the scope of official processes, and without notifying colleagues or management ,” summarizes an analysis by Inria and Datacraft.
In France, 31% of employees engage in “Shadow AI,” according to a study byOkta. This is certainly lower than in other countries (52% globally). Another survey, published by Qualtrics, paints an even bleaker picture: 80% of employees worldwide say they use artificial intelligence tools not provided by their company. This situation is a cause for concern for the AML-CFT. First, because it undermines—or even jeopardizes—the effectiveness of anti-money laundering efforts, given the lack of clear visibility into who is doing what, when, where, and with what data. Second, there is the issue of sensitive data leaking outside the company. Finally, there is the issue of management’s liability in the face of malicious uses of data that have fed into algorithms.
Augmented Intelligence: The Best of All Worlds
Between the illusion of a “ AML-CFT ” policy entirely delegated to AI algorithms and the continued reliance on cumbersome, time-consuming manual processes, there is a middle ground: augmented intelligence. This is the approach that AP Solutions IO advocates in its “Glass Box” model: making decision-making mechanisms sufficiently transparent, traceable, and explainable so that experts can retain control over the system—a requirement set by regulators to validate compliance. In practice, augmented intelligence combines human/collective and artificial intelligence, keeping humans at the center of decision-making, with technology serving only as a valuable aid (for detection, repetitive tasks, flow analysis, identification of suspicious transactions, etc.). The division of roles is therefore clear: AI handles detection, classification, analysis, and prioritization, while the compliance expert handles interpretation, justification, arbitration, and decision-making. This is what enables the shift from a compliance model where teams search for information and handle alerts to one where technology intelligently prepares the analyst’s work, with a particular emphasis on traceability, explainability, and the justification of decisions—elements that are particularly important in AML-CFT.
Companies companies are beginning to promote this distinction: according to a global survey of chief financial officers published in August 2026 by Censuswide, 90% have established a decision or compliance threshold above which human validation is always required for decisions made by AI.
Toward the “Enhanced Compliance Officer”!
Augmented Intelligence does not turn the compliance officer into a mere supervisor of technology; rather, it enhances their ability to apply their expertise where it creates the most value.
With its “Glass Box” approach, AP Solutions IO offers more efficient and controlled regulatory and legal compliance: technology helps manage and analyze complexity, while experts retain responsibility for judgment and decision-making.
The result: less time spent on processing, and greater capacity to understand, assess, and manage risks, while keeping human responsibility at the heart of the system. This is the principle behind the “augmented compliance officer”: technology that enhances human expertise and ensures sound decision-making, without ever replacing it.
Aurélien Zachayus
Co-Founder – CEO at AP Solutions IO
Reduce your false positives today
