AI automation in Aschaffenburg. Usually without the AI.

Most of what you do by hand every week follows fixed rules. A program that knows those rules is enough. I build AI in where whatever comes in looks different every time.

Book a 45-minute call

What decides it.

Whatever comes in. If it has the same shape every time, a rule is enough. If it looks different every time, AI can help read it. I figure that out for each process on its own, before I build anything.

What comes in

Fixed layout, fixed rule. No AI needed, and every invoice can be traced later.

Rules are enough here.

A fixed format: The twelve almost identical invoices at the end of the month: same line items, same customers, only the date and the amount change. An appointment that moves in the calendar and needs to move in the second place too. There’s nothing to understand there, just something to pass along.
Mistakes are expensive: Where a wrong amount or a wrong item costs money, the program should do the same thing every time. A rule does that. AI gets it right most of the time, and here most of the time isn’t enough.
Someone has to be able to check: If someone asks later why an invoice went out the way it did, you can look it up in the rule. With an answer from AI, you can’t to the same degree.

AI helps here.

Free text in emails: The order is somewhere in the body of the email, and every customer writes it differently. A rule can’t find anything reliable in there. AI can pull out the quantity and the item and drop them into the other program as a suggestion.
Photos and scans: The delivery note as a photo, the note from the job, snapped on a phone. AI turns that into text a program can work with. It reads print more reliably than handwriting.
Summarizing and sorting: Turning notes and photos from a job into a draft of the report for the customer. Sorting incoming emails by whether they’re an order, a question or a complaint.

How I build AI in.

AI suggests, a person confirms where it matters. For orders from the inbox, that looks like this:

AI reads

The email comes in, and the AI pulls out the quantity, the item and the customer and fills in the order.
Someone confirms

Someone in your business takes a look and sends it off. The typing goes away, the second look stays.
The rest follows rules

What happens next, booking, delivery note, invoice, runs on fixed rules. Nothing gets guessed after that.

Which model, and where it runs, are two questions.

The same model is often available in several places: from the maker, through Microsoft, Amazon or Google, on a server of your own, or right on the device. Which place fits depends on the data the process touches.

	From the maker	Microsoft Azure	Amazon Bedrock	Google Cloud	Ollama Cloud	Your own server	On the device
Claude Anthropic	available, no promise on where it’s processed	available, no promise on where it’s processed	available, processed in the EU or on your own hardware	available, processed in the EU or on your own hardware	not available there	not available there	not available there
GPT OpenAI	available, processed in the EU or on your own hardware	available, processed in the EU or on your own hardware	available, no promise on where it’s processed	not available there	not available there	not available there	not available there
Gemini Google	available, no promise on where it’s processed	not available there	not available there	available, processed in the EU or on your own hardware	not available there	not available there	available, processed in the EU or on your own hardware
Mistral Mistral AI	available, processed in the EU or on your own hardware	available, processed in the EU or on your own hardware	available, processed in the EU or on your own hardware	available, processed in the EU or on your own hardware	available, no promise on where it’s processed	available, processed in the EU or on your own hardware	available, processed in the EU or on your own hardware
Open models Llama, Qwen, Gemma, gpt‑oss	not available there	available, processed in the EU or on your own hardware	available, processed in the EU or on your own hardware	available, processed in the EU or on your own hardware	available, no promise on where it’s processed	available, processed in the EU or on your own hardware	available, processed in the EU or on your own hardware

available, processed in the EU or on your own hardware
available, no promise on where it’s processed
not available there

As of October 2026, per the providers’ documentation. It changes all the time and differs from model to model. The proposal says which cell it’ll be for your process.

Where a mistake costs little, say an email lands in the wrong folder, you don’t need the second look. Where you do, we decide that together before we start, one process at a time. Which AI service runs behind it and what happens to your data along the way is spelled out in the proposal.

Whether your process needs AI is what the first call is for.

Usually the answer is no. That makes it simpler.

Book a 45-minute call