Automation · AI agents AI agents for business: what they do, where they fail and how to start

What an AI agent is

An AI agent is a program with a language model at its core. Unlike a simple chat, it is given a goal and access to tools: it may read emails, look something up in a database, create a document or call an interface. The model decides which steps it needs and in what order, checks intermediate results and reports back when it gets stuck.

Under the label agentic AI this is being promoted heavily at the moment. For businesses a sober view is more useful: an agent makes sense where a person today jumps between several programs, gathers information and turns it into a proposal. That groundwork is exactly what the agent can take over.

AI agent, chatbot and classic automation

In practice these are often combined: fixed rules handle what is always the same, the agent takes the unclear cases, and a person decides what goes out.

Six typical uses in businesses

• Sorting enquiries and preparing quotes: the agent reads incoming enquiries, assigns them, looks up prices and terms and presents a draft quote.
• Incoming invoices: it reads invoices, matches them against purchase orders and delivery notes and flags deviations for accounts.
• Second-line customer service: it answers questions about order status by checking the system and hands anything sensitive to your team.
• Appointments and jobs: it arranges appointments with customers, takes calendars and travel times into account and proposes a schedule.
• Reports: it collects figures from several sources and writes a draft of the weekly or monthly report.
• Research: it compares suppliers, standards or tenders and summarises the results with sources.

These examples describe typical processes, not client projects. Which of them brings you the most depends on where your team loses the most time today.

Where the limits are

• Errors: language models can produce plausible but wrong information. That is why a person approves anything that goes out or moves money.
• Cost per run: every step uses computing power. At very high volumes, a fixed rule is often cheaper.
• Access rights and security: an agent may only do what its task requires. External input such as emails can contain hidden instructions. The security organisation OWASP puts these attacks at the top of its list of risks.
• Data protection: personal data needs a legal basis, a data processing agreement and ideally servers in the EU. More under AI and data protection.
• EU AI Act: organisations that use AI must ensure sufficient AI literacy in their teams. If an agent talks to customers directly, they must be able to tell that they are communicating with an AI.

How to start with your first agent

1. Choose one process: frequent, time-consuming, with a clear result. For example, preparing quotes from email enquiries.
2. Set rules and a measure of success: what should the agent do, and what never? How will you measure that it works well, for instance time per case or error rate?
3. Human in the loop: the agent prepares, your team approves. Only when results have been right for weeks is approval relaxed.
4. Test in parallel: the agent first runs alongside the existing work. That shows you without risk where it gets things right and where it does not.
5. Expand: only when the first process works does the next one follow.

Build it yourself or have it built?

Tools such as n8n, Make or Microsoft Copilot Studio make first agents possible even without programming. For experiments that is ideal. As soon as an agent works with customer data, writes into your systems or processes many cases, further questions arise: permissions, logging, error handling, data protection, operation and maintenance. That is where support pays off. As specialists in AI implementation we develop AI agents for small and mid-sized businesses, connect them to your existing systems and accompany the test phase.