Starting from March 2025, the Novomatic network, which includes over 900 gaming venues, has implemented the new version of SmartAML, an Android tablet application developed by Gruppo Fastal for managing anti-money laundering controls.
Two requirements pulling in opposite directions
The overhaul did not come from an appetite for new technology but from a specific problem. Extending the network toward indirectly managed outlets — bars, tobacconists, third-party venues — puts the platform in the hands of staff who do not report to the concessionaire and who must still be trained on its procedures.
Hence two requirements that pull against each other. On one side, the operator’s room for discretion during customer identification and the tracking of takings and winnings has to narrow, because that is where errors and evasive behaviour take hold. On the other, filling in a record has to become intuitive enough not to require weeks of training.
Tightening the controls slows onboarding down; simplifying the interface widens the room for manoeuvre. Version 2.0 is the attempt to escape that contradiction.
Key Features of SmartAML 2.0
Version 2.0 represents a significant technological upgrade from the previous version 1.3.3. Among the main improvements is a new development framework that ensures native Android software compilation and compliance with Google’s Material Design 3, resulting in overall usability benefits.
The architecture, by eliminating the local persistent database, increases application security and enhances operational performance.
Artificial Intelligence Integration
The most innovative element is the pervasive implementation of AI to simplify daily operations and reduce operator workload.
A new back-end server has been developed that provides computer vision services to verify documentation quality and OCR technologies to automatically populate data from document images and gaming tickets.
The system, called “AI/OCR services layer”, uses an agentic application that leverages Document Intelligence services and advanced language models (LLM) from the Microsoft Azure platform to automate data analysis and extraction. The models were trained specifically on the various ticket formats issued by the Betting, Bingo and VLT systems present in the venues.
From correction to prevention
The deepest change, however, does not lie in the models. It lies in a reversal of how the application treats error.
In the previous version the check came after the act. If the operator entered the personal details of a minor, the system blocked progress with a warning, and no one could move to the next screen without first correcting the entry. Every mandatory field had to be filled in the order the sequence prescribed.
In version 2.0 most of those warnings are gone, replaced by adaptive behaviours that act earlier. The interface simply makes it impossible to enter data that is wrong or inconsistent: calendar-style date pickers open already set to eighteen years before the current date, tax codes belonging to minors are rejected at the source, and submission of the record is enabled only once every mandatory field has been completed — while navigation back and forth between screens stays free.
It is a difference that looks slight and is not. A constraint that blocks forces the operator to work out what they got wrong; a constraint that prevents lets them work without ever facing the problem. The first has to be taught in a classroom, the second is learned by using the application — which is exactly what is needed when the staff to be trained report to someone else.
The same logic governs operational flexibility. If the player realises they have left their ID behind, the operator is not stuck: they carry on entering the gaming ticket data and resume identification when the player comes back from the cloakroom. And because photographing the documents is one of the first steps in the procedure, OCR recognition pre-fills most of the fields that follow, reducing them to a verification read.
What changes in the back office
The same services layer works downstream too. AI analyses records as they arrive and returns a summary to first-level back-office staff, who find their lists already reordered by estimated level of attention.
The practical effect is to bring the check forward. Many verifications close while the player is still on the premises, and an anomaly can be corrected on the spot instead of producing a rejected record and a later round of rework.
The result
The AI functionalities were activated progressively, once testing and model tuning were complete. On the Novomatic network, the production rollout halved the time needed to fill in a record and eliminated the most common errors — the ones that caused a record to be rejected by the concessionaire’s first-level control.
The number, though, is not the interesting part. The interesting part is that those errors are no longer caught: they are made impossible.