From 2 November we are a GbR, and we build agent offices

As of 2 November 2026, KI Maschine is a GbR, a German partnership under civil law. Two partners in equal shares, Martin Sokolowski and Miguel Marquez Gonzales, two developers with around 25 years of experience each, eight of them side by side at the same company. The legal form is new, the work is not.

What we build

We call it an agent office. Picture a Monday morning. At half past six nobody is in the building yet, but the inbox holds supplier invoices from the weekend, two enquiries from the contact form and the dictation of a Friday meeting.

By eight o'clock that has turned into a folder. Every invoice comes with a posting suggestion, with the supplier recognised, the line items listed and an account assigned. For each enquiry there is a draft quote that follows the quotes you sent for similar jobs. The dictation has become minutes with a task list, and every task has a name behind it.

None of it has been sent. None of it has been posted. It sits there and waits.

Then a person comes in, looks at the folder and decides. The first invoice goes through. On the second one the account is wrong, so they correct it. The quote for the customer there was trouble with last year, they would rather write themselves. What they approve moves on. What they do not approve does not happen.

What you get out of it

The difference is not cosmetic. A machine that writes a draft takes nobody's job away. It takes the twenty minutes off the bookkeeper that would otherwise have gone into typing out invoice numbers, and it takes the cold start out of writing a quote for sales. What is left is the work somebody was hired for, which is checking and deciding.

Approval stays with a person, and there is a sober reason for that. An agent gets things wrong. It misreads an invoice total, it assigns a supplier to the wrong cost centre, it writes a delivery date into a quote that you cannot meet. A mistake in a draft sitting on a screen costs thirty seconds. The same mistake in a payment that has already gone out costs you a morning and an uncomfortable phone call.

That is why we build it this way round. The agent prepares, the person approves. There is a pleasant side effect to that. A system where a human grants approval does not order itself the next run, and it does not send out a series of emails overnight that nobody has read. It does not protect you from every mistake, but it protects you from the expensive ones.

All of this runs on the company's own server. The invoices, the quotes, the minutes and the templates sit where the rest of your data sits, not in an account that gets more expensive at some point or changes its terms.

What runs at our own place

We operate what we build. Our cloud with files, calendars and contacts, our mail server, the team password manager, the Git server and this website run on a server we run ourselves. Backups happen on several levels, one copy is kept off site, and for a total failure there is a written recovery plan. It starts with the password safe, because without it nothing else can be unlocked, and the two credentials for it are on paper.

This blog is the second example. It is written by eleven agents working to a school, meaning a set of rules in a database. The rules say who we are writing for, how we write, which words we do not use and which claims we can back up. Every rule comes with a reason and an example. A check reports when a draft breaks one of them, before anyone reads it.

After that the post sits there as a draft and waits for a person. They read it, cut, correct and approve. This one included.

That is the same setup as in the agent office. Rules in one place, agents working to them, a check before anything goes out and a human approval at the end.

Which language model, you decide

The choice of model belongs to the company, not to us. A public model over an interface is a legitimate choice, a model on your own hardware in house is just as legitimate, and different tasks can use different ones. The decision can be changed later, because the model stays a replaceable part and the data and the system belong to you.

What we do not claim is that nothing leaves the building. If you publish through an interface or use a model in the cloud, you are sending data there. That belongs on the table before anything gets built, not afterwards. For us, sovereignty means freedom of choice and no dependency, and part of that is saying openly where the line runs.

If you want to know what this would mean for you

Write to us. Name one recurring task in your company that eats time every week, and we will tell you whether an agent office is worth it for that or whether ordinary automation comes cheaper.