Process Optimization

Before I automate anything, I look at where the time is actually going. The problem is often not where people assume it is.

Then I connect what belongs together. Your systems usually don't talk to each other, not because it is impossible but because nobody ever set it up. When data flows automatically, the manual steps disappear and so do the errors they produce.

Most of what people want automated needs no AI at all. A workflow that does the same three things every Tuesday should do exactly those three things every Tuesday: deterministic, traceable, no surprises.

For that I use n8n, an automation platform that I run on your own infrastructure. What separates it from the well known cloud services is not the feature set but the location: your data never leaves your building. With customer data, quotes or HR records, that is often the point at which automation becomes approvable in the first place.

Where n8n runs out of road, I write that particular step as a service of its own in .NET and hook it into the same workflow. You end up with one chain in which standard building blocks and custom parts run side by side, rather than another isolated system sitting next to the ones you already have.

AI can do a great deal here, though not everywhere and rarely in the way vendors promise. I use it where it has a real, measurable effect: classification, text processing, pre-sorting large volumes. Where a simple script is more reliable and cheaper, I write the script. You are paying for results, not for a buzzword.

Where AI does come in, the legal framework belongs in the plan from the start. I hold the AI competence certification from the AI Training Institute and work in line with the EU AI Act. For you that means AI solutions you won't later have to explain, either to regulators or to your own customers.

An AI agent is a different thing from a fixed workflow. It decides for itself what to do next, and it can trigger real actions: sending mail, changing records, placing orders. That is exactly where it becomes useful and risky at the same time.

An agent is the first kind of software that can be talked into things. Anything it reads may contain instructions: an incoming email, an uploaded PDF, a web page. Build without planning for that and you have a system an outsider can steer.

So every agent I build gets a fixed harness along **zero trust** lines. Nothing is treated as trustworthy, neither the input nor the agent itself:

• Only the tools it genuinely needs, each with its own tightly scoped credentials. No general purpose key, no direct write access to your database.
• Permitted actions are listed explicitly, rather than forbidden ones. Anything not expressly allowed does not happen.
• Anything it reads counts as data, never as an instruction. Content from emails, documents or the web cannot redirect the agent.
• Irreversible steps need a human. Payments, deletions and anything that leaves your building get prepared and then released by you.
• Hard limits on steps, runtime and cost, so a loop cannot consume your budget overnight.
• Every step is logged: what was read, what was decided, what was triggered. Every action stays traceable.

That is a lot of caution for a tool that looks so easy in a demo video. It is also the difference between an agent you can run in production and one you switch off after the first incident.

Not every process should be automated. Some run too rarely, others change too often, and some are simply cheaper by hand. I will tell you where the effort pays for itself and where you are better off keeping your money. That is part of the advice, not the small print.