01 — Approach
Production beats demo
We have been building with LLMs since the first GPT-3 APIs, and before that we spent twenty years taking software projects into production. That second half is the more important one for AI implementations. Because the same pattern is repeating everywhere right now: the prototype impresses, then comes the real data, the first edge case, the first cost report, and the project dies quietly.
So our implementations start at the other end. First we define how success is measured and what happens when things fail. Then we build the smallest system that meets that bar. Where a ready-made tool is enough, we recommend the ready-made tool, for example Langdock for broad team usage. We only build what has to be built. The overview of all the routes is on our Automation & AI page.
02 — Current
What's changing in AI systems right now
In 2026 the Model Context Protocol went from experiment to standard. Since Anthropic handed MCP over to a foundation under the Linux Foundation at the end of 2025, with OpenAI, Google and Microsoft as founding members, the question is no longer whether to connect your systems, but how. According to surveys from summer 2026, around three quarters of enterprise AI teams already have MCP-backed agents in production.
The second number belongs right next to it: Gartner warns that by 2027, over 40 percent of agentic AI projects will be scrapped, due to unclear value, rising costs and missing controls. Both numbers together tell the real story. The technology is mature; projects fail at the craft around it: evaluation, governance, monitoring.
Our position: we don't build an agent that has no success metric and no owner. That sounds strict, but it is the reason our systems are still running after a year.
03 — Discipline
Agents and agentic workflows
What separates an agent from a chatbot is that it works through multi-step tasks on its own: it reads the incoming request, gathers the relevant data, drafts the response, files the case in the right system and raises its hand when it gets stuck. Typical cases at our clients are proposal preparation, invoice checking, candidate pre-screening and support triage.
The crucial part is the scoping: which steps the agent may take alone, where a human has to approve, what happens under uncertainty. We work out that scope before the first line of code. If you have a process in mind and want to know whether it is worth turning into an agent, get in touch for a no-strings use case check. We will also tell you honestly when a simple workflow without AI is the better answer, just as we regularly do for AI in mid-sized companies.
04 — Discipline
RAG: answers from your documents
Retrieval Augmented Generation means the model answers not from its training knowledge but from your documents, with source references. That sounds simple and fails in practice almost always at the same spots: poor chunking, outdated document versions, no permission checks on retrieval. We build RAG pipelines with clean source preparation, document-level access control and a test set of real questions that every change has to pass before it goes live.
05 — Discipline
MCP connections to your existing systems
The value of an agent stands or falls with its access to your systems. Via MCP servers we connect CRM, ERP, ticketing system, databases and internal APIs so that every AI tool can use them in a standardized way, with cleanly defined permissions per connection. The advantage over one-off integrations: you build the connection once and use it from every client, whether that is your own agent, Claude or ChatGPT.
06 — Discipline
Direct API integrations and cost monitoring
When AI is meant to be built firmly into your software or your processes, we integrate the OpenAI and Claude APIs directly: model choice by task instead of by habit, prompt versioning, caching where possible and fallbacks for when a provider stalls. From day one this includes cost monitoring per use case. Token costs look harmless individually and quickly become a line item in the monthly close at volume. We want you to know that number before it knows you. If the integration grows into a product of its own or an internal system, our custom development takes over seamlessly.
07 — Operations
Evaluation, monitoring, error handling
Whether an AI system is good is decided not at the acceptance meeting but in week six. That is why every implementation of ours ships with three things: an evaluation set of real cases that makes quality measurable and runs again on every model or prompt change. Monitoring that makes answer quality, latency, error rates and costs visible. And defined error handling: what happens when the model produces nonsense, the API goes down or a case falls outside the pattern. Most of the time the answer is: hand over cleanly to a human instead of failing silently.
An AI implementation is finished when it notices its own bad days.
Which process should run first at your company?
Tell us briefly which process you have in mind and which systems are involved. We will tell you honestly whether an implementation pays off, what it costs and where you should start. First call free of obligation, 30 minutes.