AI implementation that runs in production — not stuck in a sandbox.
From pilot to company-wide rollout: I build tailored AI solutions for mid-sized companies, integrate them into your existing systems, and make sure they actually get used.
Why so many AI pilots fail — and how we avoid that
Most AI pilots don't fail at the model. They fail at integration into the real business process, at lack of team adoption, at unclear success metrics, or at architectures that demo well and burn in production. An AI solution nobody uses is not a solution.
My implementation approach therefore doesn't start with the model — it starts with the user. Who will use this daily? Which clicks does it save? What happens when the model hallucinates? Only after these questions are answered do we get to architecture, model selection, and data plumbing. The result: solutions still running after three months and delivering value after a year.
I work end-to-end: architecture, data pipeline, model integration, frontend, evals, monitoring, security, GDPR documentation, handover to your team or co-operations. You have a single point of contact for the entire solution — not a chain of subcontractors blaming each other.
What's standard in implementations with me
Not because it sounds modern but because it hurts in production otherwise.
- 01
Evals from day one
Before we ship a single feature we define measurable quality criteria. So we know whether model updates make it better or worse — and can avoid silent regressions.
- 02
Human-in-the-loop where it matters
On any decision that legally or financially counts, a human stays accountable. AI proposes, human confirms. GDPR-friendly and risk-reducing.
- 03
Per-use-case observability
Every call is logged, every answer is traceable. You see costs, latencies, quality trends — and can investigate bad answers specifically.
- 04
GDPR-compliant architecture
EU-hosted models, clean DPAs, documented legal bases, data minimization. On-premise or hybrid where required.
- 05
UX your team enjoys
A technically correct solution nobody uses is no success. We design interaction so it saves clicks rather than adding them.
- 06
Handover-ready or co-operations
I build so your team can take over — clean docs, tests, onboarding. Or I stay on retainer for ongoing operation.
What typical implementation projects deliver
- 4–8 weeks
from kickoff to a production pilot with real users
- 30–60%
faster handling on document-heavy processes (RAG)
- ≤ €1
typical LLM cost per business transaction in optimized setups
Typical project flow
From kickoff to production rollout. Concrete phases, clear deliverables.
- 01
Discovery & architecture — 1–2 weeks
We sharpen the use case, clarify data access and security requirements, decide on model class, define success metrics and evals. Outcome: architecture sketch and sprint plan.
- 02
Pilot build — 4–8 weeks
We build a working pilot with real data and a small pilot user group. Not a prototype but a production-grade solution — just at limited scope.
- 03
Pilot operation & tuning — 2–4 weeks
We run the pilot with real users, collect feedback, measure against the eval suite, harden weak spots. This is where we decide whether to scale or adjust.
- 04
Scale & rollout — variable
Only after a successful pilot do we scale to full user volume. Performance, cost control, training, integration with adjacent systems. The pilot becomes a production tool.
- 05
Handover or co-operations
Either your team takes over with clean docs and onboarding, or we move into a retainer for ongoing development. Both work.
What surprised us: it wasn't an AI project, it was a software project with AI inside. That's exactly why it shipped.
FAQ — AI implementation
What does a typical implementation project cost?
Depends on scope. Pilot projects with clear boundaries typically range from €25,000 to €80,000. Larger implementations with integrations and rollout sit above. You get a fixed price or a clear time-and-materials estimate before kickoff.
Which models do you use?
Whatever fits the use case and data protection requirements: GPT-class models via Azure OpenAI (EU hosting), Claude via Anthropic, open-source models like Llama or Mistral on-premise where needed. Model selection is part of architecture — not ideology.
How do you integrate with our existing IT?
I work with your IT, not against it. We use your existing APIs, auth systems, data sources — and ship in your stack (typically cloud-native on Azure / AWS / GCP, on-premise where required).
Who's accountable for GDPR and data protection?
You as data controller. I deliver GDPR-compliant architecture, vendor DPAs, documentation of the legal basis, data flow diagrams — everything needed for a clean DPIA.
What happens after the project?
Two options: your team takes over with onboarding, docs, and a transition period. Or we move into a monthly retainer for joint development and operations.
Can you take only parts of the project?
Yes. If your team builds the frontend and you only need help on AI integration, we do that. If you only want architecture review, that works too. Full-scope is the most common but not the only mode.
You might also like
Let's talk for 30 minutes.
I listen, ask questions, and tell you honestly whether and how I can help.