AI agents in SMEs: Where they are already worthwhile today – and where they are not

AI agents are the buzzword of the year – hardly a software announcement can do without them. In SMEs, the situation looks more sober: A recent survey on behalf of Alibaba.com and the bevh from September 2026 shows that only about one in four companies uses AI agents. This is not a cause for panic or serenity. This article ranks what the numbers say, where agents already provide reliable benefits today and how you set up a first assignment so that it does not get stuck in the pilot stage.

What an AI Agent Actually Is

A chatbot answers questions. An AI agent performs tasks: He breaks down a goal into steps, calls up tools – such as an ERP search, an email template or an interface to the ticket system – and decides what to do next based on the results. The difference is not the language ability, but the Capability to act. This is precisely where the benefits and risks arise at the same time.

Technically, a lot has happened here in the past year and a half. Standards like this Model Context Protocol (MCP) it makes it much easier to connect agents to existing systems in a controlled manner than to program each connection individually.

What the numbers say

Three surveys from this year paint a fairly uniform picture – even if they measure differently:

• Loud KfW Mittelstandspanel between 2022 and 2024, around 20 percent of small and medium-sized enterprises used AI, down from just 4 percent between 2016 and 2018. For companies with more than 50 employees, the last share was 36 percent.
• The AI index Mittelstand 2026 salesforce and the German SME Confederation (around 700 companies surveyed) report that 51.2 percent use or test AI. The share of AI agents has nearly doubled from 2024 to 16.6 percent.
• The YouGov survey on behalf of Alibaba.com and bevh (438 decision-makers from e-commerce and procurement, September 2026), there is a total of 24 percent agent use and around 45 percent for companies with ten or more employees.

The differences are mainly explained by sample and definition – what counts as an “agent” is interpreted differently by each study. Nevertheless, the trend is clear: usage is growing fast, and the gap between companies that use AI systematically and those who wait is widening.

What it's about

More interesting than the usage rates are the reasons why companies hesitate. In the AI index Mittelstand, 39.9 percent cite a lack of knowledge about meaningful applications as the biggest hurdle, followed by the protection of data and trade secrets (32 percent) and a lack of legal clarity (26.6 percent). The YouGov survey shows a lack of trust in technology (32%) and an unclear economic benefit (31%).

Remarkable: Less than 5 percent of companies in the AI index have abandoned AI projects. So the problem is rarely that AI does not work – but that it is not even clear where to start.

Where Agents Reliably Benefit Today

Agents play their strength where tasks are rule-based enough to automate, but too unstructured for classic software. Four samples have proven themselves in practice:

1. Entry processing

Inquiries, orders, complaints and invoices come by e-mail, PDF or form – in any conceivable form. An agent reads it, assigns it, adds missing data from the ERP and creates a prepared process, which a person only checks. In the YouGov survey, around 40 percent of respondents cited order processing and customer service as a central burden.

2. Knowledge work with internal documents

An agent who can access manuals, contracts, specifications, and service histories not only answers questions, but gathers information from multiple sources and creates a draft—such as an offer, disruption analysis, or executive summary. The basis for this is a clean RAG system.

3. Data maintenance and data quality

Finding duplicates, comparing master data, marking incorrect entries: This is laborious, recurring and prone to error for humans. As agents help, we have contributed to AI agents for data quality described.

4. Routine tasks in IT

Pre-sort tickets, evaluate log files, prepare standard requests such as password resets or authorization requests. Here, the benefits are quickly measurable, and the processes are usually well documented.

Where you should be careful

An agent who is allowed to act can also act wrongly. Three principles prevent most problems:

• People in the process at external impact. Everything that leaves the company – an email to customers, an order, a payment – should be released by a human at the beginning.
• Minimum rights. An agent gets exactly the accesses he needs for his task, and nothing more. Reading access is a good place to start.
• Traceability. Each step is logged: which data is read, which tools are called, which decision has been made. This is not only important for troubleshooting, but also with regard to the EU AI Act and that new product liability law.

How to Start Right

1. Choose a process, not a technology. Find a task that happens frequently, takes time and the result is easy to check.
2. Measure the actual state. How long does the process take today, how often does it occur, what is the error rate? Without these figures, the benefit cannot be proven later.
3. Start small and controlled. An agent who makes suggestions that a person accepts or rejects quickly provides insights – without risk.
4. Think about data protection from the beginning. Where sensitive data is processed, models that on-premises or in a European cloud running.
5. Only expand when the results are correct. More rights and more autonomy only exist if the agent works reliably on a small scale.

Conclusion

AI agents have arrived in the middle class, but far from widespread. This is an opportunity: If you start with a well-chosen, narrowly defined application case, you gain experience before the competition follows – and avoid the expensive mistakes that arise in hasty large-scale projects. The key is not the most powerful model, but the right process and clear guidelines.

If you want to find out where AI agents in your company bring the most value, we support you with our AI consulting in Stuttgart – from potential analysis to prototype to operation.