AI agents in the mid-market: how much autonomy you should allow
According to eco, 54 percent of companies use or are trialling AI agents, and 74 percent allow them at most to work under human control. How to define what an AI agent may do, which safeguards are common and how the PASSION4IT AI workshop sets the limits for each use case.
How much an AI agent may decide on its own is something you define per use case, not for the whole company: which task it takes on, which data and systems it can access, at which point a person approves and who in the business is responsible for it. The PASSION4IT AI workshop settles exactly that in one day at a fixed price of €3,900 net, with the split between AI and people for each use case and the rule for which data may go into which tool. Book a meeting directly at passion4it.de/en/booking/.
The occasion is a survey by eco, the German internet industry association, published on 14 October 2026 among 500 companies: 54 percent already use AI agents or are trialling them. At the same time, 74 percent would allow an AI agent at most supporting work or independent work under human control. Only 5 percent would let it decide fully autonomously.
What you take away from this article:
- What distinguishes an AI agent from an AI assistant
- Which levels of autonomy exist and which ones companies choose according to eco
- Which safeguards are used most often
- How to set the limits for a first agent
- Who supports you and what it costs
What is an AI agent and how is it different from an AI assistant?
An AI agent completes a task in several steps by itself and accesses systems to do so, such as a mailbox, a ticketing system or a database. An AI assistant answers a question or produces a draft, and a person takes the next action.
The difference matters for accountability. With an assistant, someone checks the result before anything happens. An agent acts, and every action has consequences in the system: an email goes out, a record changes, an order is triggered. That is why an agent needs clearly described limits before it starts.
How much autonomy should I give an AI agent?
At first so little that every action with an external effect is approved by a person. You grant more autonomy only once the agent has handled a task reliably over a defined period and you have measured how often a person had to step in.
Most companies in the eco survey see it the same way. 40 percent would use AI agents at most for supporting work, another 34 percent would allow independent work as long as human control remains. 13 percent would allow autonomous decisions in individual sub-processes, 5 percent fully autonomous decisions.
| Level | What the agent does | Who decides | Share according to eco |
|---|---|---|---|
| Supporting work | collects, sorts, drafts | the person, at every step | 40% |
| Independent under control | completes steps, a person checks | the person, before any external effect | 34% |
| Autonomous in sub-processes | decides within a defined process | the agent, within fixed rules | 13% |
| Fully autonomous | decides without approval | the agent | 5% |
Which safeguards do companies use for AI agents?
Most often an automated emergency shutdown (47 percent) and strict assignment of rights and roles (45 percent). These are followed by regular security audits at 37 percent, continuous monitoring of AI outputs at 32 percent and transparent action logs at 31 percent. Multiple answers were possible.
For a first agent in a mid-sized company, rights and logs matter most. An agent should only access the systems it needs for its task, and every action should be stored traceably. That way you can establish what happened if in doubt and tighten the approval at the right point.
How do I set the limits for a first AI agent?
With a description per use case that answers four questions. Which task does the agent take on, and which explicitly not? Which data and systems may it access? At which point does a person approve before anything goes out? And who in the business is responsible for the agent and checks the results?
Add success criteria that let you decide after a few weeks whether the agent gets more autonomy, stays as it is or is stopped. Which requirements should be written down before a vendor builds an agent or chatbot is covered in Requirements specification for a chatbot.
Who helps with getting started on AI agents?
Depending on whether you already have a concrete use case, a different route fits.
| Option | What it delivers | Cost |
|---|---|---|
| Internal departments | know the processes and bottlenecks an agent could take over | internal time |
| Software vendor with agent features | sets up agents within its own product | usually licence and setup costs |
| PASSION4IT AI workshop | six-hour day for five to twelve managers, use cases with the split between AI and people, data rule and pilot plan for 30 days | €3,900 net fixed price, €2,150 own share after BAFA funding in the former West German states |
In the AI workshop each use case gets a use case card with benefit, data, limits and the split between AI and people, plus an AI pilot canvas for the next 30 days with success criteria and a responsible person. Managers without IT background can take part, at your site or in Viechtach.
Frequently asked questions (FAQ)
How widespread are AI agents in German companies?
According to eco, 25 percent of companies use AI agents in individual departments, 11 percent across the company and 18 percent in a test phase or pilot projects, 54 percent in total. Another 7 percent have concrete plans, 33 percent have none.
Does an AI agent always have to be controlled by a person?
For actions with an external effect, approval by a person is advisable at the start. An agent can get more autonomy in a defined process once it has worked reliably there over a set period.
What does the PASSION4IT AI workshop cost?
€3,900 net at a fixed price, with no billing by effort. With BAFA funding the own share is €2,150 in the former West German states and €1,100 in the eastern states. If the workshop takes place at your site, travel costs are added, which we agree beforehand.
Where do the figures on AI agents come from?
From the eco survey »Unternehmen in Deutschland setzen KI-Agenten kontrolliert ein« of 14 October 2026. The polling institute Civey surveyed 500 companies online for eco between 18 September and 3 October 2026. According to eco, the statistical error is between 9.1 and 9.4 percentage points depending on the question.