AI MVP development for companies in Germany

What an AI MVP is

An AI MVP is the first version of a product in which an AI model does the main job for the user, such as sorting a support request or drafting an offer. Real users work with it on their own cases, and your company sees whether they keep using it.

The risk sits in the answers. A product can work on each screen and still fail because the model gets one case in ten wrong. The first version of an AI product therefore needs parts that a plain MVP does without, and the table below lists them.

What a wrong answer costs your user

Before I build, the concept settles what happens when the model is wrong. The answer decides how much a person checks in the first version.

1. The model drafts a text that the user edits before sending it. A wrong answer costs a minute of editing, so the model can write freely.
2. The model answers a customer. A wrong answer leaves your company, so your staff read the answer first and release it.
3. The model changes data, for example an order in your ERP. The first version only prepares the change, and a person confirms it outside the chat.

A first version can start at the first level and move on once the test cases show that the model holds.

The concept comes first

The concept settles who uses the product, which task the AI takes on for them, which screens lead there and what the first version leaves out. We do this work together before the first line of code exists.

I have written concepts for digital products for more than twenty years. They include more than 20 corporate websites, shopping carts and order funnels with PayPal, WorldPay and Amazon Pay, and portals that present figures of life insurers such as profit sharing and solvency ratios, which I worked out directly with the chief actuary. In 2022 I wrote the concept of a real-time market data product and took it to its first institutional customers. Because I have also programmed products like these, my concept already plans what the build needs. I challenge other people’s concepts and plans and give reasons for my view.

For your AI product the concept contains:

• the users and their task, described with real cases from your business
• the information architecture: which content and functions the product has and where a user finds them
• the screens of the main task, as a sketch or a clickable prototype
• the AI task with example answers, and the point at which a person checks them
• the data rules and the limits of the first version
• a roadmap for the versions after the first

How I build it with Claude and Codex

An AI MVP has two kinds of AI in it, and I keep them apart. The coding agents Claude and Codex write the code of the product. The model inside the product answers your users, and your data rules decide which model that is.

The coding agents work under rules written into your repository, and hooks block what a rule forbids, such as a push to the production branch. The same rules bind human developers, so your team or a later vendor continues in the same repository. Details: AI agents under the same rules as humans.

In my client work an AI agent once did its job correctly and reported a wrong count, and since then a separate script measures each result. In your product the script runs the test cases before each release. When a new version of the prompt fixes one case and breaks two others, the script shows it before your users see it.

The parts of an AI MVP

What the first version of an AI product needs, and what it leaves for later. Bring the table to our call, and we mark what applies to your product.

Part	What the first version needs	What it leaves out
The AI task	One task the model does for the user, described with example cases and their right answers	A second AI task
The model	A cloud model through its API under a business contract, or an open-weight model on your server, behind one interface in the code	A model trained on your data
Instructions	The prompt and the example cases, kept in the repository and versioned like code	An editor in which staff change the prompt
Test cases	Cases from your business with answers written by your staff, run by a script before each release	A dashboard of scores over time
Human check	A person reads or confirms where a wrong answer would reach a customer or change data	Automatic sending
Feedback	A button with which users flag a wrong answer, with the case stored for review	Prompt changes made by the product itself
Data rules	The data that goes to the model, agreed with your data protection officer	Personal data where the task works without it
Running costs	The cost of each request logged, with a monthly ceiling	Routing between several models to save cost
The product around it	A web app with a sign-in, the screens of the main job, an imprint and a privacy notice	Native apps for the App Store and Google Play

The test cases decide the most. Without right answers written in advance, nobody can tell whether a new version of the prompt is better or worse.

Who I am

I am Michael Wutzke from Frankfurt, with more than twenty years in IT and media. As an advisor I support product concepts, MVP development and technical roadmaps. At Policen Direkt I programmed the front end of a contract management portal in Angular and connected it to an existing CRM system. At the Frankfurt-based company Blocksize Capital I was Head of Decentralized Finance and Node Operations, then CIO. Today I build software with Claude and Codex for a client, and at Claude Hacker House I teach building MCP servers. I am also interested in open-source AI models that a company runs on its own servers. Details: Career stages.