Five ways to work together.

Where to start

There are two honest entry points, and which one fits depends on a single question: do you already know what should be built?

If yes — you have a workflow in mind and want it automated — start at S1b Clarity, which turns it into a build plan, and continue into a pilot and production.

If you only need to know where you stand — is this system in scope, at what tier, and what does that oblige you to do — S1a Classify answers exactly that in writing and nothing more. It covers one named system for a flat €2,900. If the intake shows there is more than one system in play, or the risk tier turns out to be genuinely borderline, it converts to Clarity and the fee comes off in full — you are never billed twice for the same ground.

If not — and this is the more common case — you know AI belongs somewhere in the business but not where, or you have pilots that never reached production. That is a leadership gap, not an engineering one, and it is what S4 AI Delivery Lead exists for.

And if the room isn’t aligned yet, the workshop is an entry point rather than a final step: a team leaves it agreeing on what to automate and what to leave alone, which is usually the precondition for either of the other two.

The ladder

1. S0

   AI-Act Quick Check

   Three steps: does your project fall under the AI Act, at what risk tier, and what follows from that.

   Free
2. S1a

   Classify

   One named system, flat price. A 60-minute structured intake, then your risk tier in writing: the classification, the duties that attach to it, the GDPR/DPIA trigger flag, and what changes for you at December 2027 — delivered within five working days. If it turns out to be more than one system, or a genuinely borderline tier, it converts to Clarity and this fee comes off in full. Also the fastest way to re-date a classification the Digital Omnibus made stale.

   €2,900
3. S1b

   Clarity

   Every system, not just one. A paid assessment that ends with a build plan, not a deck: the full use-case inventory, the one workflow actually worth automating, what it takes, and what the regulation requires of it. The Classify fee is credited in full if you continue within 60 days.

   from €5,900
4. S2

   Build an agent

   A working pilot on your own data and your own workflow, in weeks. Real output you can judge, not a demo.

   from €18,000
5. S3

   Into production

   Hardened, monitored, and documented for audit — including the evaluation evidence the AI Act expects of a high-risk system.

   from €45,000
6. S4

   AI Delivery Lead

   A fractional lead who owns your AI initiatives and the compliance lifecycle that rides with them.

   from €4,900/mo
7. S5

   Workshop

   A team day: which use cases are worth building, which are not, and how the AI Act classifies each one. Often the first step rather than the last. Staff AI-literacy training (Art. 4) rides along as an attendance record — it is part of the governance work, not a product we sell separately.

   from €3,900/day

The numbering is a sequence of depth, not an order you must follow. Most engagements begin at S1 or S4, and plenty begin with the workshop; nobody is expected to climb every rung.

S4 — the tier most companies need first

Most AI budgets are not lost on the wrong build. They are lost because nobody owns the programme: pilots stall between departments, the compliance question arrives after the fact, and a year passes with nothing in production.

The AI Delivery Lead tier fills exactly that gap — the senior person who decides what is worth building, holds the delivery line, and runs the AI-Act and GDPR lifecycle in parallel. Fractional, because a company at this stage needs the judgement far more often than it needs the headcount.

The qualification behind it: twenty years in enterprise IT, including BMW and Oracle, a certified Project Management Professional (PMP), and four production SaaS products designed, built and operated solo — one of them an EU-AI-Act governance platform. Delivery discipline from enterprise scale, applied at the size where one person can still own the whole thing.

What fixed scope means

Every tier is scoped and priced before work starts. An open-ended day rate points the supplier's incentive away from the client's, and it makes the total unknowable at the moment you have to decide. Prices are shown as “from” because the floor is real and the scope moves; the number you agree does not.

For context, a mid-level AI engineer runs €70,000–90,000 a year, and an agency build is quoted in six figures. Every tier here starts with the free check.

Who this is for

A good fit if you:

• have a real workflow eating real hours, and data that already exists;
• need the result to survive an audit — regulated industry, professional secrecy, or a board that will ask;
• want one accountable person rather than a team you have to manage;
• can decide. A capacity-bound studio cannot carry a twelve-month procurement.

Probably not a fit if you want a proof-of-concept to show an investor, the lowest bid, staff augmentation by the day, or a model fine-tuned on data you do not have yet.

What every engagement includes

• Documentation that outlasts the engagement — written so your team, or your next supplier, can take it over.
• The compliance evidence as a by-product — AI-Act classification and technical documentation, and the technical input your DPIA and data-protection officer ask for. Engineering, not legal advice.
• Your infrastructure, your ownership — the agent runs where your data policy requires, from an EU-resident managed endpoint to fully on-premise. Source and infrastructure are yours.
• Direct access to the principal — no account manager in between.

How we contract

The practical answers your legal and procurement people ask before a first engagement — published here so they don’t have to ask.

• Direct, no intermediary. You contract agent.berlin (Inhaber: Dieter Stölzel, Berlin) directly. No agency margin, no vendor-management layer, no one between you and the person doing the work.
• Fixed scope, deliverable-based. Every engagement is a Werkvertrag against a written scope agreed before work starts — not a day rate, not an open-ended retainer. You buy a defined result, and the risk of it taking longer is mine.
• Standard paperwork, ready to send. Framework agreement, statement of work, an Art. 28 AVV for anything touching personal data, and an NDA. Usually same-day, so contracting doesn’t add weeks to a project.
• You own the output. Source, infrastructure and documentation are in your accounts and your name — from the start, not at handover. Nothing runs on a credential only I hold. See where it runs.
• Handover is a deliverable, not a favour. Architecture documentation, an operational runbook and the compliance artifacts are written to the same standard as the code, for someone who has never spoken to me. More on that under continuity.
• Larger scopes bring vetted senior partners under the same framework agreement, so “can one person carry this?” has a real answer rather than a reassuring one.
• Engineering, not legal advice. AI-Act and GDPR work here means classification, technical documentation and evidence. Your counsel or data-protection officer takes the legal position; the artifacts it rests on are mine.

Questions

What does building an AI agent cost?

A working pilot on your own data starts at €18,000 and a hardened production deployment at €45,000. A written AI-Act classification for one named system is a flat €2,900, a full assessment ending in a build plan starts at €5,900, and the AI-Act check is free. Apart from the flat-priced classification, these are entry prices for a fixed scope agreed before work starts.

We have several AI systems. Do we pay €2,900 for each?

No — that would be the wrong product for you. Classify is deliberately scoped to one named system, because a single clear answer is what most people actually need first. If the intake shows you have an estate rather than a system, Classify converts to Clarity, which inventories all of them, and the €2,900 comes off that fee in full. You will not pay twice for the same ground, and you will not be quietly billed per system without being told first.

What exactly do we get for €2,900?

One 60-minute structured intake call, and then a written classification for that system within five working days: the risk tier and the reasoning behind it, whether you are provider or deployer, the duties that attach, a GDPR/DPIA trigger flag, and what changes for that system at December 2027. It is a document you can hand to your lawyer or data-protection officer. It is engineering, not legal advice, and it states its assumptions on the page so anyone can check them.

What happens to the plan if we don’t hire you to build it?

You keep it. An assessment you paid for is yours to implement in-house or hand to any AI engineer you trust — it is written to be usable by someone else, and saying so up front is the only way the advice in it can be trusted. Plenty of engagements end there, and that is a fine outcome. Hiring the studio to build it is the other option, not the catch.

Why publish prices at all?

Because the studio is capacity-bound, so price is a filter. Published numbers let you decide whether this fits before either of us spends time on a call.

How long does a pilot take?

Weeks, not quarters — because your agent starts on a reusable engineering core shared with four production products, not on an empty repository. You pay for your problem, not for boilerplate.

What if we do not know what to build yet?

Then start at S4, not S2. Building the wrong thing well is the more expensive mistake.

Does the EU AI Act apply to us?

Often yes, and the risk tier decides how much follows. The Act has applied since 2 August 2026; the Digital Omnibus deferred only the high-risk duties, to December 2027. The free check returns an applicability answer, a classification and the duties that come with it — start there rather than guessing.

Where does our data go?

Where your policy requires. The default is an EU-resident, zero-retention endpoint that never trains on your data; from there the same agent can run in your own cloud tenant, in a network-isolated private endpoint, or entirely on your own hardware. See data sovereignty.