Introduce AI agents without flying blind.
Three stages that build on each other and can be booked separately. After each one you decide again — on numbers rather than on gut feeling.
Stage one takes a week. After that you know whether stage two is worth it.
An AI pipeline is not a tool you buy. It is the way work moves through your engineering organisation — except that parts of it are no longer carried out by people but by agents: bounded assignments a model completes on its own, with access to exactly the tools that job needs, and with checks that record whether the result is good enough.
The difference from what happens in most companies today is not the technology. It is the commitment: a scoped assignment instead of a chat history, a trail instead of a screenshot, a measurement instead of an impression.
Who this is built for.
A fit if …
- you run an engineering department of roughly five people or more
- AI tooling is already in the house, but nobody can evidence what it delivers
- you need a result that holds up in front of leadership or a board
- data protection, audit and works council have a say — and should have one
Not a fit if …
- you want a tool recommendation: you get that free of charge on the first call
- nobody on your team has time — without your people it stays a foreign object
- the goal is "train once, then quiet": training alone changes nothing
- the codebase may not be touched — that would leave advice without evidence
Three stages. You can stop after each one.
Every stage ends with an artefact that is worth something even if nothing follows.
-
Assessment
We go through repositories, pipelines, reviews and tickets, and talk to the people who work in them every day. The result is a report: baseline numbers, three candidates where effort and impact are in a defensible ratio, and a sober read on data protection and auditability.
-
Pilot with one team
One team, one real path, agents running. We build next to your developers rather than in a side project: task scoping, checks, permissions, audit trail, kill switch. At the end we measure against the stage-one numbers — including when the result is uncomfortable.
-
Rollout and operations
What held in the pilot goes wide: more teams, a playbook, an operating model, two internal multipliers who build new agents themselves. We become unnecessary on purpose — that is the success case, not the exception.
What stays in the house.
Four things that remain, whether or not we keep working together.
One path in production
At least one way of working where agents contribute productively, with checks and an audit trail. Not an experiment that disappears at the next reorganisation.
Numbers from before and after
The same metrics, taken twice, with the same method. Enough to justify an investment decision — including against internal resistance.
A playbook that is yours
Task definitions, checks and configuration sit versioned in your repository. No black box, no licence on our method, no dependency on us.
Two people trained
Two of your developers can build, judge and switch off new agents themselves. That is the actual handover; everything else is documentation.
Agents without guardrails do not survive an audit.
A common reason AI initiatives get stopped inside companies is not poor quality. It is the question somebody eventually asks: who approved this, and what can we show for it?
So we do not add guardrails afterwards. They cost about a day in stage two — and they are the difference between a pilot that goes into operation and one that dies in the security review.
- Scoped permissions per task, never a master key
- A complete trail: who, what, when, with what result
- Human approval wherever money or customers are touched
- Processing inside the EU wherever the task allows
- A kill switch your leadership can find without us
A price that is fixed before we start.
Every stage has a fixed price. We name it after the first call and put it in writing before anything begins — after that it does not move. The assessment fee is credited against a following pilot.
There is no number on this website because scope and system landscape determine it. Anything else would be guesswork, and you have had enough guesswork already.
No commissions from tool vendors. We recommend what holds, not what pays us.
What took a team of five two years ago, I now build alone with AI agents, in a fraction of the time. Companies that ignore that end up paying for their competitors.
What decision makers ask us
Our developers already use Copilot. Is that not enough?
That is a typical starting point, and a good basis. The question is how much faster you have become, and how you measure it. If nobody can answer that, this is exactly where the work begins. Autocomplete in the editor is a different thing from an agent that carries a task from start to finish.
Does our source code end up at an AI vendor?
We settle that in stage one, before anything is built: which data goes where, what is contractually covered, which providers commit to EU processing and no training access. Anything that cannot be modelled cleanly does not get automated, however tempting it is technically.
What happens when an agent produces nonsense?
That is what the checks and approvals are for. An agent produces a proposal that runs through the same tests, reviews and rules as a human contribution — with explicit approval on top at critical points. Agentic does not mean unsupervised.
Will our developers lose their jobs over this?
Not in any engagement we have seen. What disappears is routine: scaffolding, trailing tests, dragging one change through several layers. What remains is design, judgement and responsibility — and there is more of that, not less. We say the same thing when the works council is in the room.
Why not just try it ourselves?
You can, and many do. The difference is rarely the tooling; it is the order: measure, build, measure again. Starting without a baseline gets you a debate after three months instead of a decision.
How much of our own time does this cost?
Stage one: roughly one day spread across several people. Stage two: two to four hours a week from the pilot team, plus the review time you spend anyway. If you cannot free that up, we say so beforehand — then the timing is wrong, not the idea.
Start with the assessment.
One week, one report, three concrete candidates. Then you decide on numbers whether a pilot follows — or whether something else should change first.
Thirty-minute first call, free of charge and without a deck.