AI agents for small and midsize companies in Germany

What an AI agent does in a small company

An AI agent is a language model that works through a task with tools until the task is done. It can read an order confirmation from an email and enter the delivery date in the customer record of your CRM.

The agent chooses its next step itself. That makes it useful where the input varies, such as emails in free text or PDFs in different layouts. It is also the reason an agent needs limits that a fixed program does not need.

Script, workflow tool or AI agent

Many routine tasks need no AI. Before I build an agent, I compare it with the two simpler options, and the task review uses this table.

	Script	Workflow tool	AI agent
Input	One fixed format	Fixed formats from several systems	Varies: free text, scans, different layouts
Rule	Fixed and simple	Fixed, with branches	Written in words, needs judgment
Who changes it	A developer edits the code	Staff edit the flow	Staff edit the written rule
Cost per run	Server time	Server time and the tool’s license	Server time and model usage
Typical failure	Stops at an unknown format	Stops at an unknown format	Goes on with a wrong reading, so a check follows each run
Pick it when	The format stays the same	A fixed path runs across several systems	A person reads and decides today

A task can use all three. A script fetches the files, the agent reads them, and a second script checks the result before anything reaches your live data.

How I keep an agent within bounds

An agent in your company gets the limits of a new colleague on the first day. These are the layers I use in client work:

• Rights at the system: a database role that only reads, and API keys for the test system.
• A test system first. The agent runs its jobs there until the trial run shows its results match your staff’s.
• A confirmation outside the chat before any change to live data. An approval that the agent quotes from the chat does not count.
• A read-only script that measures each result. When the script and the agent disagree, the script’s number stands and the agent runs again.

A written rule alone does not bind an agent. It follows the rule most of the time, and a machine check catches the cases where it does not. Details: AI agents under the same rules as humans.

Who I am

I am Michael Wutzke, an AI engineer in Frankfurt with more than twenty years in IT and media. At Policen Direkt I designed the interfaces of a contract management portal and connected it to an existing CRM system. Today I help a client rebuild its platform with Claude and Codex, and I teach building MCP servers and virtual organizations of agents at Claude Hacker House. I am also interested in open-source AI models that a company runs on its own servers. Details: Career stages.