In this article, we will discuss AML control automation as the fight against money laundering is constantly evolving. At Revelis, we have extended our Moneying™ platform with Generative AI techniques specifically to provide a useful tool to meet the continuous new guidelines issued by relevant authorities on how to identify emerging anomalous behaviors.
What is Moneying™
An anomaly indicator, no matter how well formulated, does not exist in isolation: it must run on real data, produce prioritized alerts, and return the necessary context to the analyst to make a decision. Moneying™ is the environment where this happens. The platform ties together customer risk profiling, transaction monitoring, and relationship analysis, leveraging the Rialto™ framework for large-scale data processing.
Moneying™ leverages both core approaches of Artificial Intelligence: deductive and inductive models. Deductive models are structured using regulatory rules produced by investigative authorities and remain explicitly expressed: readable, versionable, and individually verifiable.
Inductive models learn from data and step in where rules fall short: in risk grading, recognizing relational patterns, and reducing the noise that stifles compliance teams under false positives. Deduction and induction do not replace each other; they complement each other.
As a result, the foundational requirement holding the entire architecture together is that every output produced by the platform is accompanied by the underlying evidence that determined it, in a format that is readable by the analyst and usable in audit/inspection settings.
AML control automation: a constant challenge
Anti-money laundering resembles cybersecurity: a perpetual game between attackers and defenders. On one side, bad actors seek new vulnerabilities to exploit; on the other, defenders reinforce systems to close them. The Bank of Italy and investigative authorities oversee this front by regularly updating their guidelines, describing emerging schemes to intercept in natural language; financial intermediaries are tasked with translating those indications into active controls over transaction flows.
This is where the mechanism slows down. Between the regulatory description of an anomalous behavior and the executable rule capable of recognizing it lies a manual translation process: slow, subject to interpretation, and required to be repeated at every update. The result is a safeguard relying on static, cumbersome architectures to modify, leaving defenders chasing behind, trying to seal the breach after it has already been crossed.
To narrow that lag and achieve true AML control automation, it is essential to make the dynamic phase of “translation”—namely, the implementation of inductive and deductive processes for detecting anomalies in financial transactions. Generative AI, and specifically Large Language Models (LLMs), intervenes precisely at this stage: given a sufficiently precise description of the behavior to intercept, the model produces the sequence of chained operations that recognize it in data, and a downstream system deploys it into production. Thanks to this AML control automation, the new control becomes available as a guardrail in a matter of days, rather than months.
Moneying-GenAML™
The capabilities described above are implemented in the GenAML module of Moneying™, a Moneying plugin developed under the Moneying-Plus project. It is a new plugin engineered to shorten the gap between the publication of a new recommendation by authorities and its deployment into production, revolutionizing AML control automation. The starting point is the document as-is: a Bank of Italy circular, an anomaly template, or an investigative report. The framework extracts its control logic: which data to observe, with which filters, on what thresholds, and over what time frame, and translates it into an executable indicator for the institution’s transaction flows.
In practice, GenAML is a plugin that operationalizes AML control automation starting directly from natural language descriptions.
The output is neither plain text nor a code snippet to be integrated manually: it is a series of concatenated operations called a workflow inspectable step-by-step, which the analyst can review, refine, and validate before production deployment. The model does not replace the compliance officer; rather, it provides an operational draft for effective AML control automation on which officers exercise their human judgment, shifting human labor from transcription to verification.
Furthermore, the cycle does not end with the initial generation. When guidance is updated, or when false positives indicate a threshold is poorly tuned, the same framework acts on the existing indicator and produces a new version. It is this incremental revision capability that makes enforcement dynamic: adapting AML control automation becomes a matter of minutes, not weeks.
Large Language Models and Model Context Protocol
Entrusting a language model with free-form code writing in a regulated domain is a difficult proposition to justify: the risk of hallucination is concrete, and the outcome is difficult to audit. GenAML’s approach to AML control automation is different. The LLM is not asked to generate arbitrary software, but rather to plan: given the regulatory text and relevant information on the underlying data structure, it composes a sequence of operations selected from a closed, pre-validated catalog of primitives data acquisition, SQL filter, transformation, export and defines how they chain together. Naturally, these primitives are customizable.
This is where the Model Context Protocol (MCP) comes into play the standard used to expose available tools and their function signatures to the model. The protocol acts as a contract: it defines what the model can invoke and with which parameters; anything outside that scope is simply non-expressible. Far from being a constraint, this limitation is the guarantee: every plan generated for AML control automation is syntactically valid by construction, fully traceable in its individual components, and reproducible.
This gives rise to an architecture where model creativity is confined strictly where needed (understanding regulatory intent and breaking it down into steps), while execution remains in the hands of deterministic and certified components. A balance that renders such AML control automation fully compliant with audit requirements typical of the financial sector.
AML Control Automation: How the Plugin Works
Figure 1 illustrates the pipeline used to achieve AML control automation through workflow generation. The user uploads reference documents (PDF or text files, typically official circulars and data schemas) and describes the control to be implemented in natural language. From there, the process is automated. The model, instructed as a planner, outputs the pipeline in JSON format by composing available tools. The execution plan is displayed step-by-step to the user, allowing every decision to be verified before execution.
The final step is conversion: the pipeline for AML control automation is translated into a native workflow that can be downloaded and imported directly into the Moneying™ platform, which relies on Rialto™ as its execution engine. The outcome is a native asset within the existing data management environment, rather than a parallel system requiring separate maintenance.
Alongside generation, the system provides an editing mode: starting from an existing workflow, it can be evolved using a new natural language instruction an essential feature for dynamic AML control automation that, in daily anti-money laundering practice, occurs even more frequently than creation from scratch.
The effectiveness of GenAML in AML control automation
As with any system automating a previously manual step, the effectiveness of GenAML in AML control automation depends on key baseline conditions worth highlighting. The first concerns data rather than algorithms: the quality of generated indicators mirrors the quality of the underlying sources. Knowing in advance which data streams are queryable, how frequently they update, and the reliability of customer profile fields is the single most critical factor determining project timelines.
Furthermore, human oversight remains central to the platform: AML control automation focused on generation does not eliminate human governance; it repositions it, shifting analyst workload from writing to verification. The return on investment is highest when the compliance team is involved early in defining the presentation format of indicators and accompanying evidence. It is the legibility and transparency of the output that makes the workflow not only executable, but fully governed which, in a regulated domain, is the most critical property.
Conclusions
The gap between the regulatory description of an anomalous behavior and the executable rule that detects it has always existed; what has changed is the speed at which money laundering techniques evolve, and consequently, the cost of compliance lag. As long as translation remains a manual task, defenders will continue to play catch-up, hindering true AML control automation.
GenAML does not promise to replace human oversight; on the contrary, it presupposes it. It redefines boundaries: assigning planning for AML control automation to the language model (constrained to a validated catalog of operations), deterministic execution to the infrastructure, and final judgment on recommendations to the analyst. GenAML completes this chain by outputting native workflows for Rialto™, while Moneying™ provides the operational environment where indicators run on real-time data alongside statistical models and risk profiling.
The primary value is not merely saving development effort significant as that is. It is the capability that this form of AML control automation provides to adapt to regulatory updates within days rather than months.
Would you like more information about Moneying?
Contact us at: info@revelis.eu
Authors
Simone Mungari – Enterprise Solution Engineer
Francesco Cupello – Enterprise Solution Engineer
