Process automation and AI integration.
First I look at how the work actually gets done — with the people who do it every day. Then the work that comes back daily and moves nobody forward disappears: copying data, producing documents, triaging requests. And where AI helps, it goes into the software you already run — not into another tool nobody opens.
Why analysis comes before tooling
Automation built on a poorly understood process only makes the poor process faster — afterwards the detour is cast in software.
So this starts with an afternoon at the desk of the people doing the work, not with picking a tool: which file gets copied where? Where does someone wait for someone else? The result is a sober list of where the time goes. Sometimes the best automation is deleting a step entirely.
What usually turns out to be automatable
The worthwhile candidates look much the same everywhere: frequent, clearly ruled, professionally undemanding.
- Map the workflow, name the bottlenecks, put numbers on effort and return
- Automate recurring work, from quote to invoice
- Connect the systems you already run: CRM, ERP, e-mail, file storage
- Extend your existing software with AI instead of replacing it
Connect systems instead of replacing them
The real problem is rarely one bad system but the gap between several good ones: the CRM knows something the ERP needs — and that gap is bridged by people who retype and forward.
Those gaps can nearly always be closed through the interfaces the systems already ship with, without touching any of them: far cheaper than a replacement. Where AI helps, it joins in; where conventional rules are enough, that is what you get.
AI integration: into the software you already have
An additional AI tool means another tab, another login, another place knowledge lives. Most of these tools get tried for two weeks and never opened again. Integration means the opposite: the AI arrives where the task arises — the draft reply is already in the ticket, the quote text already in the right format in the right system.
Before anything is built, we settle which data may go to a model outside the building and at which points a person looks before anything goes out. AI is sometimes wrong — an integration that ignores this produces fast text and slow corrections.
- Triage, summarise and file incoming e-mails and documents
- Generate drafts where they are needed: quotes, replies, reports
- Enrich and reconcile data somebody currently looks up by hand
- Build AI features into your own software, via API or MCP
The person stays, the typing goes
AI does not replace people, it amplifies them: whoever did the work before still makes the same decisions afterwards. What disappears is the retyping, the forwarding and the checking up.
That is not a reassurance formula, it is a condition for the thing working — automation that arrives looking like redundancies gets quietly undermined by exactly the people whose cooperation it needs.
Questions about this
How long does an analysis take?
For a bounded area, usually one to two days on site plus write-up. What you get at the end is a list with effort and expected return per candidate — including when the conclusion is that automating it is not worth it.
Do we have to change our systems?
As a rule, no. The normal case is connecting the systems you already have through their interfaces. Changing systems is a separate, far bigger project and should never be a side effect of an automation exercise.
What happens to our data if AI is involved?
We settle that before anything is built: which data may leave the building and which may not. If data cannot leave, that is a requirement like any other — not a reason to bury the project.
Does this work with our industry software?
Usually, yes. What matters is whether your software has an interface or stores its data in a reachable form — which is true of most systems by now, including older ones. Whether it is true of yours can be settled in a few minutes in the intro call.
Which AI models do you use?
The ones that fit your data protection requirements and the task. That can be a large model in a European data centre or a smaller one running on your premises. I work a lot with Claude but am tied to no vendor — and I will also say so when your task needs no AI model at all.
The other areas
Most projects touch more than one of these.
- Custom software development from HamburgCustom web applications, APIs and legacy modernisation from a freelance software engineer in Hamburg — including advice that sometimes argues against a rewrite.
- AI training for business teams and managersHands-on AI training for business teams and managers, using your real tasks — covering the EU AI Act's AI-literacy obligation along the way. Hamburg and remote.
- Claude Skills and MCP server developmentDevelopment of Claude Skills and MCP servers: hand recurring tasks and formats to an AI permanently, and make internal systems safely reachable for it.
Sound like your problem?
A 30-minute call, no obligation. By the end you know whether it is worth doing — including when the answer is "probably not".