Berlin

Berlin AI development.
Data that stays put.

Berlin is the German city where most of the inbound is about a product rather than a plant — an AI feature that has to work inside something people already pay for. The local complication is not the technology. It is that Germany supervises data protection through sixteen Land authorities, so the guidance that binds you is not necessarily the guidance your competitor in Munich is reading.

Worked from Barcelona
Supervised by the BlnBDI
Same hour as us

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

All Where We Work →

What is actually different here

Your supervisor is decided by geography

Germany does not have one data protection regulator. Each Land runs its own, and a company established in Berlin answers to the Berliner Beauftragte für Datenschutz und Informationsfreiheit, with the federal BfDI covering federal bodies and the telecommunications and postal sectors. The practical effect on a build is small but real: when we cite regulator guidance in a design document, we cite the one that will actually be read to you, and where the Land authorities differ in emphasis we say so rather than averaging them into something nobody published.

The same office also handles freedom of information

Berlin's authority supervises both data protection and Informationsfreiheit. That combination matters if you build for the city, for a Land body or for a public-sector customer, because the same system can be simultaneously subject to a disclosure request and to a duty not to disclose. It is a design question, not a policy question: what the system logs, how long it holds it, and whether a log written to help you debug a model becomes a record someone can ask for.

German compute is a configuration, not a project

Frankfurt is the workhorse region for German workloads and Microsoft also runs a Germany North region, so keeping storage and inference physically inside Germany is normally a matter of choosing where things run rather than building anything special. Where a hosted model has no German endpoint, the honest answer is either an EU endpoint with the transfer written down or a self-hosted open-weight model — not a claim that the region list will have changed by the time you need it. Region availability moves; confirm it against the provider's current documentation when you commit.

Product AI fails on economics before it fails on law

Most Berlin engagements are a feature inside a product with a per-seat price attached, which means the constraint is cost per request and quality at the ninetieth percentile, not compliance. A feature that costs more to serve than the plan it sits inside is a business problem wearing a model's clothes. We treat evaluation and cost modelling as part of the build rather than as something you commission afterwards when finance asks.

The work that comes out of this city

Berlin's economy is software, mobility, public-sector technology and media, and the shape of the work follows that almost mechanically. Very little of it is a factory problem. Most of it is somebody's roadmap.

Software and SaaS companies arrive with a feature that already half exists: a summariser, an assistant, a search box that is meant to answer rather than list. The technical question is rarely whether a model can do it — it usually can, on the ten examples in the demo. The question is what happens on the ten thousand it has not seen, and whether you can tell the difference before a customer does. That is an evaluation problem, and we build the harness with the feature rather than after it. The pattern is described under in-product AI and the measurement side under evaluations.

Mobility companies bring operational data with a geography attached — vehicles, journeys, incidents, support tickets that reference all three. The interesting systems join those together, and the interesting constraint is that a journey is personal data about a person even when your dashboard treats it as a row.

Public-sector and govtech work brings a different bar entirely. The output has to be explainable to someone who did not commission it, the record has to survive a change of supplier, and the transparency duties are not optional extras. That work usually starts with governance rather than a model, which we cover under AI governance.

Earlier-stage companies want something in front of users before the next raise, and that is a genuinely different engagement — smaller, faster, and honest about what has been deferred. It is described under startups.

How a berlin project actually runs

We fly up for discovery

Barcelona to Berlin is a direct flight of a little over two hours, and the first workshop is worth doing in a room with your engineers and whoever owns the roadmap. On a normal engagement we do not bill it as a line item.

Then remote, on the same clock

Berlin and Barcelona are in the same timezone, so there is no offset to manage at all — no morning window, no waiting for someone to wake up. A question asked in your standup is answered during it.

And back for handover

The session where your team takes ownership works better face to face. Afterwards the system lives in your repositories and your cloud accounts, and you are not dependent on us to change it.

Questions

Questions about Berlin AI Development.

Do you have a Berlin office?

No. We have one office and it is in Barcelona. We come to Berlin for discovery, for handover and for anything where being in the room changes the outcome, and the rest is remote on the same clock. A firm our size claiming a Berlin presence would be describing a coworking desk.

Which German authority supervises us?

Normally the one for the Land in which you are established, so for a Berlin company that is the Berliner Beauftragte für Datenschutz und Informationsfreiheit; federal bodies and the telecommunications and postal sectors sit with the BfDI instead. It matters less for the architecture than people expect and more for the documentation than they hope, because the guidance you will be measured against is that authority's. Your data protection officer owns that determination, not us.

Can everything stay in Germany?

Usually yes. The major providers run German regions, and open-weight models can be self-hosted inside your own infrastructure, so a system where no personal data leaves the country is an ordinary configuration rather than a bespoke build. Where you want a specific hosted model that has no German endpoint, we will say so plainly and describe the alternatives instead of quietly routing it elsewhere.

Do we have to work in German?

Our working language is English, and technical documentation and code are delivered in English by industry convention. That is separate from what the system speaks: German-language input and output is routine, and we build and test for it. Evaluating whether German output is good enough for your customers is a job for your German-speaking staff, and we build the review tooling so they can do it without reading logs.

What does a Berlin engagement cost?

The same as anywhere else. Our ranges are on the pricing page, in euros and exclusive of VAT, and we do not price by postcode. What does move the number is whether the feature ships to your own users or to a regulated third party, because the second one carries more evidence work.