AI agents in business: examples, limits and getting started

An AI agent completes a task in several steps and chooses the tools for this itself. For example, it reads an enquiry, looks up the customer in the ERP system used to manage the business, checks stock levels and creates a draft quote. In SMEs this works reliably today if the agent has a narrowly defined task, is given only the permissions it needs, and a person approves every step with an external effect in advance. According to Bitkom, 11 per cent of companies with 20 or more employees already use AI agents (September 2026).

Assistant, automation, agent

The boundary is fluid. Many solutions sold as agents are an automation with one AI step. Such a fixed sequence is easier to test and explain. An agent is worthwhile when the inputs vary so much that the path cannot be defined in advance.

Five examples from SMEs

1. From enquiry to draft quote

• Task: read customer enquiries by email, identify missing details, find similar past quotes and create a draft.
• Tools: read the mailbox, search past quotes, item master data in the ERP, create a draft.
• Approval: sales checks the draft and price before anything is sent.

There is more on this in the use case quote and enquiry processing.

2. From order email to ERP

• Task: extract orders from emails and PDFs, match customer and items, flag discrepancies.
• Tools: read documents, customer and item search, create an ERP draft.
• Approval: the sales back office confirms every draft and checks discrepancies separately.

3. Questions about delivery and status

• Task: answer customer questions such as "Where is my delivery?".
• Tools: look up the order, query the shipment status, write a draft reply.
• Approval: a person checks every reply. For standard cases, one click is enough.

4. Requesting missing receipts

• Task: identify open items and write a list of queries for each customer or client.
• Tools: read the export from accounting, reconcile against incoming receipts, create the list.
• Approval: the person responsible checks the list and sends it.

5. Knowledge questions in the team

• Task: answer questions from the team about manuals, policies or contracts, with the reference.
• Tools: search in approved documents.
• Approval: none, as long as the agent only answers and names the source. Legal questions go to a responsible person.

How an agent is built

An agent in a business consists of five parts:

1. Brief: a clear description of the task and its limits, including what the agent must not do.
2. Language model: the model that reads, plans and writes. Where it runs determines data protection.
3. Tools: interfaces to the mailbox, ERP, customer management (CRM) or documents. Increasingly, they are connected via the Model Context Protocol (MCP), a standard for connecting tools and data sources.
4. Approval points: the points at which a person confirms before anything is sent out or booked. How to define these is covered in the article on human-in-the-loop.
5. Log: it records every decision and every tool call so that errors can be traced.

The typical risks

• Too many permissions. An agent that is meant to read emails does not need write access to the ERP. Grant permissions for individual tools rather than blanket access to whole systems.
• Hidden instructions. An incoming email can contain text intended to make the agent do something else. Specialists call this prompt injection. It is number 1 on the OWASP list of risks for language model applications. Approvals before actions with an external effect are the most effective protection.
• Endless loops and costs. An agent that keeps planning new steps uses up tokens, the units of text that the model processes or generates. Limit the number of steps and the budget per task.
• Unclear responsibility. Who is liable if the agent offers a wrong price must be settled before the start. It is almost always the person who approves.

How to get started

1. Choose a task that occurs often, currently takes time, and where an error is noticed before approval.
2. Measure for a week how long the task takes today.
3. Initially give the agent read permissions and the ability to create drafts, rather than more extensive write permissions.
4. Test on real cases, including the messy ones.
5. Only extend permissions once the rate of correct drafts has been stable for several weeks.

What such a first step costs is covered in the article What does AI cost?. In our industry sprints, we build an agent like this in two weeks at a fixed price, for example for freight forwarders or mechanical engineering firms.