The situation in Catalonia
Catalonia’s public administration is in a full-scale race towards AI. The Catalonia AI 2030 Strategy mobilises €1 billion, CatSalut is tendering consultation transcription with weight given to the AINA project’s models, and the Barcelona Supercomputing Center is tendering storage and private inference for language models. The European Union, with its technological sovereignty package of May 2026, is pushing the public sector towards its own or European infrastructure.
But between the big tenders and the daily work of a mid-sized town council, a consell comarcal (county council) or a consortium, there is a gap: small teams, lots of case files and no authorised AI tool. Meanwhile, staff use cloud chat tools on their own.
What it looks like day to day
A planning officer asks for a draft report on a planning application, based on the case file and the current bylaw, with every reference cited. The secretary’s office receives the transcript of last night’s plenary session, showing who spoke on each item, and draft minutes. At the public information desk, an assistant that introduces itself as an AI answers in Catalan about deadlines and documentation, and passes the query to a person when needed. In every case, a person reviews and decides.
How we deploy it
- Assessment. We identify the procedures with the heaviest document load, the data involved and your ENS requirements. You come away with use cases, an architecture and a fixed proposal.
- Pilot of 2 to 4 weeks. One department and one specific case, for example regulation search or plenary transcription, within a minor contract. We measure time saved and quality.
- Production. Extension to more departments, technical documentation for security and data protection, and staff training.
Catalan, first class
In Catalonia, the working language of the administration is Catalan, and an AI that only half writes it creates more work than it saves. We start from open models that master it and from the AINA project’s corpora, and we have gone further: in our Barcelona lab we built a Catalan vocabulary pipeline that raised the coverage of the ‘draft’ model that speeds up responses (speculative decoding) from 75.2% to 99.6%, so that drafting in Catalan does not lag behind Spanish.
What we are and what we are not
file.cat is a small, technical outfit, not a large consultancy. We do not hold ENS or ISO 27001 certifications, and we do not hide it: that is why we install inside your perimeter and document every measure. For large projects, we can provide the local inference part to whoever wins the contract.
When you don’t need it
If the use is drafting public texts with no personal data, or your organisation already has an authorised corporate AI platform, you may not need another one. We will tell you so clearly.