Brussels Worked from Barcelona Multilingual by requirement Same hour as us

Brussels.
Multilingual by law.

Everywhere else we spend the first meeting explaining what the AI Act says. In Brussels the client often has a better sense of it than we do, because someone down the corridor worked on it. That changes the conversation for the better: it starts at classification, obligations and dates rather than at whether any of this is allowed, and it leaves more time for the two things that are genuinely hard here — languages and record-keeping.

Last reviewed 5 September 2026 · Engineering guidance, not legal advice.

WHAT IS ACTUALLY DIFFERENT HERE

Which supervisor applies depends on what you are

This is the one city where the answer is genuinely not obvious. A Belgian company or association answers to the Belgian data protection authority, the Autorité de protection des données or Gegevensbeschermingsautoriteit. EU institutions, bodies and agencies do not: they are covered by their own regulation and supervised by the European Data Protection Supervisor. If your organisation sits somewhere between the two — a body funded by an institution, or a contractor processing on its behalf — that distinction decides which rulebook the documentation has to be written against, and it is worth settling before the design work rather than during it.

Three official languages, plus the one everyone works in

Belgium runs in French, Dutch and German, the institutions add a further two dozen, and the working language in the room is often English. A document system here is multilingual as a starting condition, not as a phase two. The engineering consequences are concrete: retrieval quality varies by language, extraction accuracy varies by language, and an evaluation set that is ninety per cent English will tell you almost nothing about how the system behaves in Dutch. We build the test set per language or we do not claim the number.

Proximity to the rulemaking is a real advantage

Clients here tend to have direct access to how a rule was meant to be read, which shortens arguments considerably. It also raises the bar on what we are expected to produce: the technical documentation, the record of what the system does and the description of its limits are read by people who read that sort of thing professionally. That is a reasonable standard, and it is the standard we would want anyway.

In-country compute is available

Belgium hosts a long-established hyperscale region, with the main Irish and German regions a short hop away, so keeping storage and inference inside Belgium or inside the EU is a configuration decision rather than a special build. As always, region and service availability changes over time — confirm the specifics against the provider's current documentation when you commit to a design.

THE WORK THAT COMES OUT OF THIS CITY

Brussels produces a kind of work that barely exists elsewhere: organisations whose core output is understanding, positioning and responding to text. Associations, federations, public affairs practices and policy teams all live in consultations, position papers, transpositions and correspondence, in several languages, on deadlines set by someone else.

The useful system for that is not a chatbot. It is retrieval done properly over a corpus that the organisation already owns — with citations back to the source paragraph, with version awareness so that a superseded draft is not quoted as current, and with an honest signal when the answer is not in the corpus at all. The failure mode that matters here is a confident synthesis of two documents that contradict each other, and the defence against it is architectural rather than a matter of prompt wording. That is covered under retrieval systems, and the regulatory frame under the EU AI Act.

Logistics is the other Brussels constant. Belgium is a transit country with major ports and air freight on either side of the capital, and the resulting document flows have the same shape as anywhere else with an extra language or two attached. That work is described in our logistics document pipeline.

Chemicals and pharmaceutical organisations here bring regulatory dossiers, safety documentation and literature review, where the requirement is traceability rather than fluency: every assertion has to point somewhere. The life sciences version of that work is under life sciences research.

HOW A BRUSSELS PROJECT ACTUALLY RUNS

Discovery in person, on your premises

Barcelona to Brussels is under two hours. We come for the first workshop, and where the material cannot leave your building we do the corpus review on site rather than asking for a copy of it.

Then remote, with no offset

Same clock as Barcelona, so the working day is shared end to end. For organisations working to consultation deadlines, a same-day answer is worth more than a faster model.

And back for handover

Handover includes the technical description of what the system does and where its limits are, in a form your own people can put in front of a reviewer. After that it runs in your infrastructure.

QUESTIONS

QUESTIONS — 5

No. We have one office and it is in Barcelona. We travel to Brussels for discovery and handover and for anything that has to happen in a room. In a city where a great many organisations maintain an address purely to say they have one, we would rather not add to it.

Not directly through the large framework contracts — we are a small firm and we are not on them, and pretending otherwise would waste your time. What does work in practice is direct engagement with associations, federations and private organisations, or subcontracting to a prime who holds the framework. If the procurement route requires something we cannot meet, we will say so in the first call rather than after four weeks of paperwork.

No. We build software. We can build the system that finds, extracts, compares and tracks the documents your policy people work from, and we will not offer an opinion on what the policy should be. The division of labour is deliberate: you would not want engineering advice from a lobbyist either.

By treating each language as its own quality problem. Retrieval, extraction and generation quality all differ by language, and averaging them into a single accuracy figure hides exactly the case that will embarrass you. We build a labelled evaluation set per language, report per language, and where one is materially worse we say so and either fix it or scope it out rather than shipping a number that flatters the system.

Our ranges are on the pricing page, in euros and exclusive of VAT. We do not price by postcode. Multilingual work sits higher in a range than a single-language equivalent, because the evaluation effort multiplies even when the build does not.

Keep reading

Retrieval Systems (RAG)

30-minute technical call · no deck