- Agents pay off when they save real hours - below that, operating costs eat the gain.
- Precisely describable processes belong in classic automation, not in a language model.
- Without a clean data foundation and a process owner every AI project fails - order matters.
What an agent actually is
An AI agent is not a chatbot with a new name. It understands a task, plans steps, uses your systems - CRM, ERP, email - and delivers a result. The difference to classic automation: it copes with cases nobody defined precisely beforehand.
That is also exactly where the risk lives: what decides flexibly can decide wrongly. Which is why every agent needs clear tool boundaries and a human with the final word. In our projects this is not a footnote but architecture: the AI prepares, humans approve.
Terminology: "agent" here means an LLM system with tool access and a planning loop - not every chatbot that calls itself one.
The maths: three examples
Whether an agent pays off is not a matter of faith but of arithmetic: hours saved times hourly rate, minus operating costs. Three patterns from our consulting practice:
| Use case | Before | With an agent |
|---|---|---|
| Capturing and checking incoming invoices | Document-by-document manual work | AI reads, checks, pre-books - a human approves |
| Standard support requests | Answered ticket by ticket | An assistant with order and product data handles the routine |
| Internal knowledge search | Search, ask, wait | Answers from your own documents in seconds |
The first case is not a slide for us but production: for entsorgo, a document workflow reads and classifies different business documents, validates them against approved business rules, and prepares the case for accounting. Humans review and approve - nothing is posted automatically. The results are described in the AI accounting case study.
Agent or automation?
Rule of thumb: if the process can be described exactly, use classic automation - cheaper, faster, deterministic. If it needs judgement, language or shifting context, the agent earns its keep. Most good systems are a mix: deterministic rules where possible, AI where necessary.
Practice shows that strong systems combine deterministic checks with AI. Clear assignments and validations remain rule-based; the language model helps with ambiguous content and reading errors. Rules first, AI second.
Learning instead of administering
The underrated part of an agent project is not the model but the controlled feedback loop. Human corrections reveal where rules, data preparation or instructions need improvement.
We turn confirmed patterns into reusable guidance for future cases only after domain review. How that knowledge is stored and applied depends on the process and its protection needs. The system improves in operation without automatically adopting unreviewed individual corrections.
When to walk away
No clean data foundation, no clear process owner, or savings that cannot carry the operating costs: then the agent does not pay off. Clean up the process first, automate second - in that order. Reverse the order and you automate your chaos.
