Milan Worked from Barcelona Italian cloud regions in the city Same hour as us

Milan.
Residency in the city.

We are based in Barcelona, an hour and a half away and on the same clock. Milan is one of the few European cities where the local regulator has intervened directly against consumer AI products, and where the compute you would use to build one sits inside the city limits. Both of those change how a project is planned.

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

WHAT IS ACTUALLY DIFFERENT HERE

The Garante is not a theoretical regulator

Italy's data-protection authority has been among the most willing in Europe to act against AI services directly, including the widely reported temporary limitation placed on ChatGPT in Italy in 2023. The practical consequence for a Milan build is not fear, it is sequencing: the lawful basis, the transparency notice and the retention answer have to exist before launch rather than being drafted in response to a letter. We would rather design against that than around it.

Residency is a local option here

Italy has its own cloud regions, in and around Milan, from more than one major provider. That means keeping inference and storage physically in Italy is an ordinary configuration choice for a Milanese client rather than a project of its own — which for a bank or an insurer removes the transfer argument from the AI layer entirely, and it is the cheapest way to win that argument.

Italian is a quality requirement, not a checkbox

Most models write serviceable Italian and a few write good Italian, and the difference is visible to a customer in a way it is not in a benchmark. For anything customer-facing we evaluate candidate models on your own text, in Italian, with your people scoring it, before anything is chosen. Model choice is covered under choosing a model; the point here is that the choice is made with Italian output on the table, not English.

Lombardy is a supply chain, not a set of companies

The machinery and fashion businesses around Milan run on documents moving between hundreds of small suppliers — specifications, certificates of conformity, delivery notes, invoices, often as scans. That is a document-pipeline problem with an unusually messy input distribution, and the engineering that decides whether it works is the handling of the bad twenty per cent, not the good eighty.

THE WORK THAT COMES OUT OF THIS CITY

Milan holds two economies that rarely appear on the same page anywhere else: Italian finance, and Italian design. The AI problems they produce are genuinely unlike each other.

Fashion and luxury groups arrive with catalogues. A brand with tens of thousands of products across seasons, sizes, materials and six languages has a search problem that keyword matching has never solved and a merchandising problem that nobody wants an opaque model deciding. The work is search and merchandising where the ranking is explainable enough that a buyer will accept it, because if the merchandising team does not trust it they will simply override it and you have bought nothing.

Banks, insurers and asset managers bring the opposite temperament: a person reads a long document, makes a judgement, and has to defend it later. The interesting engineering is the trace rather than the model, and it is the same pattern we describe under financial services.

Machinery and industrial manufacturers bring supplier paperwork at volume, and increasingly quality inspection on the line. Those are different systems with different failure modes, and we would rather do one properly than both partially — the inspection side is covered under quality inspection.

Pharmaceutical companies in the region bring regulatory and research documents where a plausible summary is worse than no summary, because it will be believed. Retrieval with a citation back to the source paragraph is the whole requirement.

HOW A MILAN PROJECT ACTUALLY RUNS

We come up for discovery

Barcelona to Milan is under two hours with a dense timetable, so the first workshop happens in a room and we do not charge it as a line item on a normal engagement. For a catalogue or document project we want to see the real data on the day, including the scans nobody is proud of.

Then remote, no timezone arithmetic

Italy and Spain keep the same hour, including the same daylight-saving switch, so nothing needs converting. In practice that means a question raised in a Milan standup is answered before lunch rather than the next day.

And back for handover

We come back for the session where your engineers take ownership, and for a customer-facing Italian system we run the acceptance review with the people who will answer for the output — the merchandising or service lead, not only the technical team. After that it is yours, in your repositories and your cloud accounts.

QUESTIONS

QUESTIONS — 5

No. We have one office and it is in Barcelona. We travel to Milan for discovery, for handover and whenever being in the room genuinely changes the outcome. There is no Italian entity, no local partner and no representative — a team our size claiming one would be describing a mailbox.

Usually yes, and Milan is one of the easier cities in Europe to answer this in, because there are Italian cloud regions from more than one major provider. Open-weight models can also be self-hosted inside your own perimeter. Where a hosted model is used we document its retention and training terms at the time of writing, and you should confirm them against the provider's current terms before you rely on them.

Worried enough to do the work in the right order, which is the only sensible amount. The interventions that have made news involved consumer services launched without a clear lawful basis, without meaningful transparency and without an answer on retention or on how children were handled. A business system built for identified users, on a documented basis, with retention decided in advance and records that reconstruct a decision, is a different object. We build for that; your DPO or counsel owns whether it is sufficient.

That is decided by evaluation, not by assurance. We run candidate models against your own material in Italian and have your people score the output blind before a model is chosen, and the evaluation stays in place afterwards so a model update cannot quietly degrade the register. Where the answer comes back as not good enough, the honest options are a narrower scope or a human in the loop, and we will say which.

The same as anywhere else — the ranges are on the pricing page, in euros and exclusive of VAT. We do not price by postcode.

Keep reading

Solution

Merchandising & Search

→

Industry

AI for Financial Services

30-minute technical call · no deck