An AI agent completes a task over several steps on its own: it reads incoming items, makes decisions, uses your software to do so and reports the result. According to Bitkom, 57 percent of companies in Germany used AI in autumn 2026, the first time it was a majority. Of those using or planning AI, 11 percent already work with agents and another 29 percent plan to. This article shows where agents work reliably today and where caution still makes sense.
What is an AI agent?
A chatbot answers questions. An agent acts. It receives a goal, for example "check new incoming invoices against the purchase orders", and works through it: it opens the inbox, reads the invoice, finds the order in the ERP, compares amounts and submits deviations for approval.
To do this, the agent needs access to tools, meaning your software. That works through APIs or through the open standard MCP, which we describe in MCP explained simply.
Agent or workflow: what do you really need?
Not every automation needs an agent. A workflow, for example in n8n, runs fixed steps the same way every time. An agent pays off when inputs look different each time and decisions are needed along the way.
A simple rule of thumb: if you can write the process down as a list that always applies, a workflow is enough. In practice the mix is most common, a fixed workflow with one AI step where something has to be read, classified or written.
Where do agents work reliably today?
- Sorting incoming items: reading emails and inquiries, routing them to the right team and preparing a draft reply.
- Checking documents: reading invoices, matching them with orders and flagging deviations.
- Research and market monitoring: collecting sources, summarising them and reporting what matters.
- Knowledge questions in the team: an internal AI chat that knows your own documents, such as contracts, manuals or quotes.
- Data maintenance: completing CRM entries, finding duplicates and asking for missing details.
These tasks have something in common: they repeat, the results can be checked, and a mistake does little damage before someone notices it.
Where is caution needed?
Agents can be wrong. It becomes critical when they decide about money without review, touch contracts or talk to customers directly. Three proven safeguards help:
- Approval step: the agent suggests, a person confirms.
- Narrow permissions: the agent may only read and change what its task requires.
- Log: every step is stored and can be traced.
According to Bitkom, companies already using AI mainly name data protection (66 percent) and legal uncertainty (56 percent) as obstacles. Both can be planned for: models can run on your own server, more on that in Local AI instead of the cloud.
How do you get started with an agent?
- Pick a process that regularly costs time and whose result is easy to check.
- Write it down: what comes in, which steps follow, which exceptions exist?
- Start small and with an approval step.
- Measure after a few weeks: how much time does the agent save, and how often does someone have to correct it?
- Only then expand.
Which tools do we work with?
We build workflows with n8n and agents with open-source tools such as OpenClaw and Hermes Agent. If you like, the language models run locally via Ollama, self-hosted on servers in Germany. There is an overview on our AI & Automation page.
Sources
- Bitkom: For the first time, most companies use AI (14 September 2026, German)
