Questions & answersWhat are AI agents and when do they pay off?
An agent is not a chat. It is a chain with handovers. And it only pays off if the chain costs time today.
An AI agent receives a task and works through it in several steps: it reads, pulls data from systems, checks, prepares a case and hands it over. Unlike a chatbot it acts, unlike a fixed workflow it decides the steps itself. It pays off when a case occurs often, runs across several systems and the steps depend on the case. You decide how many permissions it gets, and most agents start with reading and preparing.
When do AI agents make sense?
When a case occurs often, runs across several systems and people, and the steps depend on the case. Complaints: read the message, find the order in the ERP, check warranty, assemble the case. Incoming invoices: extract data, match with order, flag deviation, present for approval. Supplier notices: collect, compare with requirements, rate urgency. In all three cases the effort is in the chain, not in one step.
When do they not make sense?
For rare cases, because the setup only pays off with volume. For cases with a high share of judgement, where a person has to check every step anyway. For workflows a fixed process solves just as well, because that is easier to operate. And where nobody is willing to define permissions and limits in writing.
Prerequisites
- A mapped workflow with all systems, roles, handovers and exceptions.
- Interfaces to the systems the agent should read or write, or interim paths such as exports.
- A written permissions concept: which level, what waits for approval, where it stops.
- A language model, local for customer and contract data, and an orchestration such as n8n.
Options: the four permission levels
- Read. View data, evaluate messages and documents, recognise connections.
- Prepare. Write drafts, assemble cases, formulate recommendations. A person decides.
- Change. Create or modify records, by defined rules, with a log.
- Trigger. Postings, replies to customers, orders. Only after proven reliability, always revocable.
Benefits
Shorter lead times because the chain runs in minutes instead of days. Complete cases at first contact because the agent has already pulled order data and history. Early visibility of deviations. And a log of every step that allows control and improvement.
Limits and risks
An agent with too many permissions can cause damage before anyone looks. An agent without stop rules can run in loops. And an agent whose decisions are not logged can neither be checked nor improved. Hence: permissions narrow, approval before effect, log always. If the agent touches HR decisions or safety, the EU AI Act applies more strictly and the classification should be checked beforehand.
Example
The customer service of a supplier took complaints by email, transferred them into a form, matched them with orders and passed them on to quality assurance. Every stage waited for the previous one, a week of lead time. An agent with read and prepare permissions pulled order and delivery records from the ERP, checked deadlines and presented a complete case with a recommendation. Quality assurance decided the same day. After three months of parallel operation the agent was allowed to create standard cases itself, special cases stayed with people.
Frequently asked
What distinguishes an AI agent from a chatbot?
A chatbot answers a question. An agent receives a task, uses tools such as databases and interfaces and works through several steps until the task is done or prepared for decision.
What distinguishes an agent from process automation?
A workflow follows a fixed sequence you define beforehand. An agent decides itself which steps to take in which order. For most processes the fixed workflow is the better entry. Agents come in when the workflows are too variable for fixed rules.
How much may an agent decide?
As much as you give it. Four permission levels have proven themselves: read, prepare, change, trigger. Most agents in mid-sized companies start with the first two. Everything binding waits for approval.
What happens with errors?
At read and prepare an error becomes visible before it takes effect because a person approves. At higher levels, confidence thresholds and stop rules apply. Every step is logged and traceable.
How do I start?
With a case that occurs often and runs across several systems. Map the workflow, define permissions, test in a parallel run. The whitepaper with thirteen questions is the checklist for that.
Conclusion
AI agents pay off for frequent cases across several systems with case-dependent steps. They are not a replacement for people but for the chain of collecting, matching and passing on. Anyone who starts with narrow permissions and extends them only with proven reliability keeps control and gains the time.
Are your processes ready for AI agents?
Thirteen questions for a self-check, five build-out levels and the four permission levels in detail. Direct download, no form.
To the whitepaperRelated questions: What are AI agents and when do they pay off?, How can sensitive data be processed with AI?, Which open-source AI is suitable for companies?
AI agents for business
Agents that prepare cases and trigger nothing without approval. Four permission levels, a log of every step.
To the solution: AI agents