AI proof of concept for companies in Germany

What an AI proof of concept answers

An AI proof of concept answers one question: does an AI do a given task on your data well enough to be worth a budget? It is a small, working test with a number at the end.

Management has seen AI demos, and staff use ChatGPT on their own. Nobody in the company knows yet whether the AI reads your delivery notes or your support tickets as reliably as the text in the demo. A proof of concept finds out on real cases from your company, before anyone plans a project around the idea.

How I build and measure the test

The test starts with the answers. Your department picks the cases, and a person who knows the work writes the right answer for each one before the AI sees them. Then I build the AI step on a development system, with Claude or Codex as coding agents, and run it on the same cases. A script compares each answer of the AI with the answer your department wrote and counts the hits.

I measure with a script because the report of an AI can be wrong while its work is right. In June 2026 an agent in a client’s project assigned 1,831 of 1,832 records correctly and then reported 10,481 records, because it had counted a wider area than the task named. Details: Running a company of AI agents with Paperclip.

The AI in the test works under the rules I use for coding agents. It gets the rights of a new colleague on the first day: it reads copies of data and cannot change anything in your live systems. Details: AI agents under the same rules as humans.

The test sheet: seven decisions before the build

Fill in the sheet with your team before our first call, or we fill it in together in the assessment. The examples show one possible task.

Decision	Example	Who decides
The task	Read the order number and the delivery date from supplier delivery notes	The department that does the work today
The cases	100 delivery notes from the last quarter, bad scans included	The department
The right answers	Written by a clerk for each case before the AI runs	The department
The pass mark	95 of 100 correct, and no wrong delivery date passed on without a warning	Management
The data rules	Copies without personal data, on a development system	Your data protection officer
The model	A cloud model through its API, or an open-weight model on your own server	IT
The next step after a pass	A production version with user rights and logging, or a second task	Management

A test without the third row measures nothing: without the right answers written down in advance, the result is an impression.

Who I am

I am Michael Wutzke from Frankfurt. I have worked in IT and media for more than twenty years, among other roles as CIO at the Frankfurt-based company Blocksize Capital. Today I work as an AI engineer on a client’s platform with Claude and Codex, and at Claude Hacker House I teach practical work with Claude. I am also interested in open-source AI models that a company runs on its own servers. Details: Career stages.