AI agents for SMEs: what actually works in 2026

Two years ago the conversation was about chatbots that answer questions. In 2026 everyone talks about AI agents: systems that don't just reply but complete tasks on their own. They read an email, check stock levels, draft a quote and file it for approval. For small and medium-sized businesses that sounds attractive. The real question is which parts already work reliably day to day.

What is an AI agent?

An AI agent is a language model that is allowed to use tools. Instead of only producing text, it calls functions: a CRM lookup, a calendar entry, a query in your accounting system. It plans several steps and checks intermediate results. The difference from classic automation is that an agent can handle unstructured input, such as an informal customer request by email.

Where agents help reliably today

• Triaging requests: reading incoming emails and forms, sorting them by topic and urgency, and routing them to the right person.
• Preparing quotes: deriving line items from a request, pulling prices from the price list and creating a draft that a person reviews.
• Capturing documents: reading invoices and delivery notes and transferring the data into Bexio, Abacus or another system.
• Internal knowledge questions: staff ask in plain language about processes, templates or product data and get answers with sources.

All of these share one pattern: the agent prepares, a person decides. That pattern works best today.

Where it still falls short

Agents that spend money, send contracts or change customer data without oversight are too risky for most SMEs. Language models still make mistakes, and an error in a multi-step workflow carries forward. Data protection matters too: as soon as an agent touches customer data, the Swiss revised Data Protection Act (revFADP) applies. You need to know where data is processed and who has access.

How to start sensibly

1. Pick a process that happens often and is easy to measure. For example: how long does it take today from request to quote?
2. Let the agent only read and suggest. Grant write access only once its suggestions have been right for weeks.
3. Keep logs. Every action the agent takes should be traceable.
4. Review after four to six weeks. Time saved, error rate and how the team accepts it.

Bottom line

AI agents are past the hype stage, but they don't run themselves. Start small, keep a person in the decision and measure properly, and you will save noticeable time today. That is exactly how our AI check works: half a day, one concrete process and an honest assessment of whether an agent is worth it.