Jeroen Spitaels Details EPSO AI Rollout at Drupal AI Summit NYC

Public-sector AI adoption was the focus of a session at Drupal AI Summit New York City 2026, where Jeroen Spitaels, chief revenue officer at Dropsolid, described an AI search implementation for the European Personnel Selection Office online portal.

The session framed AI less as a standalone chatbot and more as an answer layer built on structured content, source citation, observability, and staged deployment. For Drupal teams working in regulated environments, the EPSO case is useful because it connects natural-language search with governance requirements rather than treating AI as a separate interface.

Jeroen said EPSO faced three recurring pressures: high question volume, repeated queries in different forms, and a multilingual operating environment. He described the solution as AI search on two surfaces, one for internal employees and one for external candidates using the EU Careers site.

The example shown in the session handled questions such as what EPSO tests involve and how candidates can contact the office. Jeroen said responses include concise explanations, links back to official documentation, and a feedback mechanism that helps internal teams identify gaps in source content.

He also shared early operational figures from EPSO webmasters, saying that out of 2,000 incoming questions, about 500 reach the contact centre and about 1,500 are resolved through AI search. He cautioned that the rollout remains deliberate and that steady-state numbers are still settling.

Under the hood, Jeroen described a retrieval-augmented generation workflow. Structured content from documents, articles, and legislation is split into chunks, converted into vectors, stored in a vector database, and updated when the underlying CMS content changes. The system then retrieves relevant chunks and passes them to a large language model under defined restrictions.

The transparency layer was presented as central to the deployment. Jeroen said Dropsolid tracks traces, model usage, cost scores, and individual interactions so teams can replay, debug, and correct problematic answers. He also described a staged rollout pattern: internal evaluation, contact-centre use, and silent launch with controlled exposure.

For public-sector adoption, the session also highlighted data safeguards. Jeroen said the AI does not retain the organisation’s data, does not train models on it, and has a broader compliance posture. Those claims were presented as prerequisites for European institutions considering AI in production, not as independently verified audit findings.

The implementation may extend beyond web search. Jeroen said an email response workflow is being scoped for EPSO using the same content and logic, but he noted that the email agent is not yet live.