When a technical support system
learns to answer better
than the operators.

Vertical products, articulated versioning, users speaking different languages: technical support was a structural pressure point, not an occasional issue. First- and second-level customer care teams were constantly overloaded.

Riding the growing interest around AI, the client launched a procedure to assign a PoC: design and build a technical support system capable of reducing that pressure. With a precise constraint: no external service, no data outside the corporate perimeter. The system had to run in a fully isolated environment. Generazione AI's experience on private AI architectures did the rest.

The team was led by the CTO and COO of Generazione AI, technical lead and project manager respectively, supported by a software documentation specialist, a process analyst, an ML engineer and a front-end developer.

In the first phase the team studied the context and available material to build a shared approach. The client defined goals and KPIs; Generazione AI designed the system architecture on Digital Tech Support and set the standards to build the RAG, i.e. the private knowledge base on which the AI would operate. The client team was briefed on how the system works, the risks tied to AI adoption and the data needed to feed it.

The sheer volume and variety of material to handle, manuals, versioning, ticket history, required a partial content review. With the client team a shared format was defined to revisit tickets and reorganise information: a technical step that, in hindsight, turned out to be one of the most useful of the entire project.

With the data foundation ready and training completed, the team compared several models. The choice fell on a multimodal model with multilingual support, able to read not just text but also images like the screenshots users attach to their support requests.

The system worked well, but testing it rigorously was difficult: the client team had little time available and the content was highly specialised. We needed a way to generate a large number of realistic conversations without depending on people's availability.

Generazione AI proposed building a virtual tester: an automated system programmed to ask the technical support questions systematically, even in different languages, until all relevant cases were covered. A kind of simulated, very demanding customer. The result was a broad set of conversations that made it possible to measure performance precisely and validate the client's KPIs.

The system answered adequately to 90% of requests classifiable as first level, with no human intervention. For more complex, second-level ones, it directly handled many cases and, when needed, passed to the operator a pre-compiled card with the context of the request, cutting times and improving the quality of the handover.

The document review work brought an unexpected benefit: clearer documentation management processes, more precise change tracking, better-classified tickets. An improvement that remained even after the project closed.

The results convinced the client to evaluate an extension of the system: not only to external support, but also to internal teams, operators in training and other figures such as installers and hardware maintainers. Everything is already included in an integration plan currently being scheduled.

For Generazione AI, the virtual tester has become a stable tool: an idea born to solve a practical problem that has earned its place in our validation process for future projects.