Short answer: AI agents take a goal, plan the intermediate steps themselves and operate tools such as email, CRM or databases along the way. In mid-sized companies they reliably carry narrow tasks with clear stop rules. Without approvals and a full log, they do not belong in production.
Few terms are used as loosely in 2026 as AI agent. Vendors now apply it to any chatbot with database access. That is annoying, because there is a real shift underneath: the difference between an assistant that produces text and an agent that performs actions is, for a business, the difference between a suggestion and an invoice that has already been sent. This article sorts out what works today, what does not, and where the line runs.
What separates an agent from an assistant
A language model on its own answers questions. An assistant adds context, for example your documents. An agent brings three further capabilities:
- Planning. It breaks a goal into steps instead of executing a single instruction.
- Tools. It is allowed to trigger actions: send a message, create a record, call an API.
- Loops. It checks the result of a step and decides whether to correct, continue or stop.
That third property is what makes agents useful and risky at the same time. A workflow in n8n or Make always follows the same path. An agent can improvise when the input is unusual. If you are unsure which of the two worlds you need, the comparison in n8n, Make or Zapier helps.
Rule-based workflow, assistant and agent side by side
| Criterion | Rule-based workflow | AI assistant | AI agent |
|---|---|---|---|
| Typical job | File a document, create a record | Draft text, summarise | Handle a case end to end |
| Behaviour when surprised | Stops | Answers anyway | Tries another route |
| Traceability | Very high | Medium | Only with a complete log |
| Running cost | Low, usually per run | Low to medium, per user | Medium to high, per case and token |
| Setup effort | Days | Hours | Weeks, including testing |
| Makes sense when | The process never varies | People decide at the end | Many variants, clear stop rules |
Three uses that hold up today
1. Triaging and enriching the inbox
The agent reads incoming messages, assigns them to a case, pulls customer data from the CRM and drafts a reply. Nothing goes out without approval. This is the most honest entry point because a human stays the final authority.
2. Following up on quotes
After seven days without a response, the agent writes a reminder that refers to the actual content of the quote and offers two appointment slots. What the same process looks like without AI is described in Automating quotes. The agent only replaces the generic paragraph with a real reference.
3. Research with a citation requirement
The agent searches internal documents and the web, returns an answer and the sources behind it. No source, no answer. That single rule filters out most hallucinations.
What does not hold up: open-ended accounting decisions, price approvals, employment-law assessments and anything that leaves the company unchecked. A broader list of use cases with fast payback is in AI automation for SMEs.
A worked example: an agent on the inbox
A distributor receives 60 enquiries per working day. Two colleagues spend an average of six minutes each on assignment, lookup and drafting.
- Effort today: 60 times 6 minutes is 360 minutes, or 6 hours per day. Across 21 working days that is 126 hours per month.
- At 42 euros fully loaded per hour, that is 5,292 euros a month.
- With an agent, handling drops to 2 minutes per enquiry, so 42 hours or 1,764 euros. Saving: 3,528 euros per month.
- Against that: model cost of roughly 0.04 euros per case, so 1,260 cases times 0.04 equals 50 euros, plus 150 euros for platform and tools and a 600 euro support retainer. Total 800 euros.
- Net saving 2,728 euros per month. A one-off implementation of 12,000 euros pays back in about 4.4 months.
The whole number rests on those two remaining minutes. Measure the figure again after four weeks. In practice it starts closer to three minutes, because exceptions surface that nobody had described before.
What has to be settled before you start
Bitkom surveyed 604 German companies with 20 or more employees. The largest barriers to AI adoption are legal uncertainty and missing technical know-how at 53 percent each, followed by a lack of staff capacity at 51 percent. At the same time, a quarter of the companies that use AI run a single application only (Bitkom press release, 15 September 2025). Jumping from that starting point straight to autonomous agents skips two learning stages.
Four things belong in place before the first agent:
- Stop rules. Which amount, which recipient group, which action forces a human approval?
- Logging. Every action with timestamp, input, tool and result. Without it, later review is impossible.
- Least privilege. The agent gets its own account with minimal rights, not a manager login.
- Competence. Article 4 of the AI Act requires providers and deployers to take measures supporting the AI literacy of their staff; no specific level has to be guaranteed (Article 4 EU AI Act) [check legally].
Which further obligations apply is summarised in EU AI Act: what SMEs actually have to do. And if the real question is which building blocks fit together at all, start with Choosing the right AI stack.
Why agent projects fail in mid-sized companies
The three most common reasons have little to do with technology.
The process was never written down. An agent can only automate what someone can explain. If the answer to how the process works is that Mrs Berger has done it by instinct for twelve years, then this is not an AI project yet, it is a documentation project. That effort is the real price of automation, and it is a one-off.
The data does not hold. An agent that stitches customer data together from three systems with different spellings and customer numbers will produce wrong matches. Clean the master data before you start, not afterwards.
Nobody owns it. An agent needs one person who spends ten minutes a week reading logs and flagging anomalies. Without that role, nobody notices when the agent has been getting one case type wrong for three weeks.
What the agent is allowed to see
Decide before the first line of configuration which folders, mailboxes and tables the agent may reach. The rule of thumb: as little as possible, and a filtered copy in preference to full access to the original. Personnel files, health data and contracts under negotiation stay outside the scope unless a separate assessment says otherwise [check legally]. How such a review works in practice is described in our GDPR practice guide.
How to start in four steps
First, pick a case that occurs often, is clearly bounded and whose errors would be visible. Second, write the process down without AI, exceptions included. Third, run the agent for four weeks with mandatory approval and note every correction. Fourth, only then decide which steps may run unattended.
The tools we currently consider are listed on our tools page. For a structured starting point, use the AI check, or reach us through the contact form.
Frequently asked questions
Do AI agents require our own servers?
No. Most companies start on a hosted platform. Own infrastructure only pays off with particularly sensitive data or very high case volumes.
How do I stop an agent from sending something wrong?
With a hard approval gate on every outbound action and separate accounts with minimal rights. You can never fully prevent a mistake technically, but you can organisationally.
What does an agent cost to run?
Model costs for simple cases sit in the range of a few cents, but rise with every additional loop. Add platform fees and support. In the example above the total was 800 euros a month for 1,260 cases.
Are AI agents high-risk systems under the AI Act?
That depends on the use case, not the technology. An inbox agent normally is not, an agent that pre-sorts job applications is [check legally].
How do I spot a vendor using the label loosely?
Ask for the log. Anyone who cannot show which tools the agent called and with what result is selling a chatbot.
