Most companies buy ten tools. What they need is one workspace
A chatbot here, an email automation there, an avatar video because someone saw one online. The result is five isolated tools that know nothing about each other. The real lever is elsewhere: one place where AI agents work with your actual mail, files and systems and deliver results you simply approve.
The idea you walk in with is rarely the one that makes you money
That is not a criticism, it is the rule. Ideas come from what you see, and what is visible is what stands out: talking avatars, chatbots, image generators. What actually eats hours is invisible and dull. Sorting enquiries. Collecting documents. Typing the same thing into five systems. Following up. Writing minutes.
That is why I look at a real working week before anything gets built. Sometimes that ends with me advising against your original idea. It is uncomfortable and it saves you a five-figure sum.
Four things that together are more than four tools
- An inbox that works ahead. Incoming mail is read, classified, linked to the right case and supplied with a draft reply. You read and send instead of searching and drafting.
- Documents that read themselves. Quotes, invoices, delivery notes, minutes. Figures and deadlines are extracted and land where you need them, instead of in a folder nobody opens again.
- Recurring routines as a single instruction. You describe once what should happen, then trigger it in one sentence. Build the weekly report, prepare follow-up emails, check appointment data.
- Everything in one place, with a log. Every action is traceable, every sensitive action needs your approval. No agent sends anything in your name that you have not seen.
„Our system has no interface“ stopped being a no this year
The most common reason automation projects die in mid-sized companies is a sentence from the software vendor: there is no open interface, or only in the expensive package. For industry-specific software, practice and law firm systems and older inventory management, that is the norm.
That can now be worked around. An agent can operate an application the way an employee does: open windows, fill fields, read lists, export. It is slower than an interface and it works everywhere. For the ten to three hundred transactions a day that mid-sized companies deal with, that is entirely sufficient.
What gets settled beforehand, because otherwise it becomes expensive: whether your licence agreement permits automated operation, how credentials are stored securely, and who is liable if the routine enters something wrong. That goes into the quote before anything is built.
Start
Up to five people. Mailbox and file access, three configured routines from your day-to-day work, approval paths, logging, training. Roughly four to six weeks to everyday use.
6,900 €
one-off, then 349 € per month for operation, ongoing adjustment and a model budget with a fixed ceiling
Expansion
Up to twenty people, up to eight routines, roles and permissions, reporting. For companies that want breadth after the first success.
12,900 €
one-off, then 690 € per month
Bridge to a system without an open interface
2,900 € per system, then 89 € per month for monitoring. Includes the feasibility check and the legal clarification. Available on its own or together with Start and Expansion.
Where this sits in the market
Single-process automation in German-speaking markets currently sells at between 8,000 and 15,000 €, multi-stage projects between 15,000 and 35,000 €. Pilot projects at a fixed price sit at around 6,900 €, daily rates of specialised AI consultants between 1,800 and 3,000 €, and higher at the large consultancies. Start sits at the bottom of that range, Expansion below comparable multi-process projects.
The common rule of thumb: automation pays off when it replaces roughly ten working hours a week at a fully loaded hourly rate of at least 50 €. At 60 € and fifteen hours that is about 43,000 € a year.
From 5,000 € project value the rule is: prototype first, fixed price second. In week one a real routine from your business is running. Only then do we confirm the price.
Not every automation is a data protection case
The reflex is understandable: as soon as AI is involved, data protection arrives as a conversation stopper. It is true that the rules apply as soon as personal data is processed. It is equally true that a large share of worthwhile routines never touches personal data at all: website copy, analyses of anonymous figures, internal documentation.
So I sort first instead of braking across the board. Where real names, health or contract data are involved, I work with models operated in Europe or on dedicated hardware, plus a processing agreement. Where no such data occurs, that effort would only be expensive and slow.
Frequently asked
Does this replace employees?
In the companies I work with, no. What gets done is work nobody would ever have been hired for: following up properly, documenting everything, looking at figures weekly instead of quarterly. Anyone promising otherwise is selling you a story.
What does operation cost on top?
Model usage is billed by consumption and for most companies of this size lands in the low double digits per month. You get an estimate based on your real volumes before we start, plus a spending cap so nothing runs away.
We have no IT department.
That is the normal case among my customers. Operation, monitoring and adjustment are part of the monthly fee, and your contact is me, not a ticket system.
What if we want to take it over later?
Then you get access, code and documentation. I do not build anything you cannot get out of.
Arrange a call
Reply usually within 24 hours. First call 30 minutes, free of charge.