An agent is not just a better workflow
AI agents have been the talk of the town for months. The term gets used for anything that touches AI, so a clear distinction helps. A classic automation follows fixed rules: when an invoice arrives by email, file it in the right folder and add it to the table. The sequence is defined, every step is known.
An AI agent works differently. It receives a goal, not a fixed sequence, and decides for itself which steps are needed. It can handle incomplete information and react to the unexpected. That is exactly where the appeal lies, and exactly where the risk lies. Where a system decides on its own, it can also decide wrongly.
The hype is bigger than the reality
The numbers show two things at once. AI has arrived in everyday work: according to Stanford University's AI Index Report, 78 percent of organisations were using AI in at least one function in 2024, up from 55 percent a year earlier. But real agents are still the exception. Stanford calls them the next step, while noting in the same breath that the technology leaps ahead faster than people and processes can follow.
On the vendor side, caution is warranted. The research firm Gartner expects that more than 40 percent of agent projects will be cancelled by the end of 2027, due to high costs, unclear value and inadequate control. Gartner also speaks of "agent washing": many products are sold as agents but are rebranded chatbots. Of thousands of vendors, Gartner estimates only around 130 offer genuine agent capabilities.
Reliability is the real issue
The most important point often gets lost in the excitement: an agent has to be reliable across many steps, not just one. And errors add up. If an agent were 85 percent reliable at each single step, a ten-step sequence would end up succeeding only about 20 percent of the time. The longer the chain, the greater the chance that something goes wrong somewhere.
A study by Stanford and Carnegie Mellon University puts it plainly: teams of humans and AI deliver clearly better results than fully autonomous agents. The researchers recommend step-by-step collaboration, not an unsupervised handoff. The human stays in the loop, not as a brake, but as a checkpoint at the points that matter.
When an agent is worth it for an SME
A useful rule of thumb comes from Gartner itself: an agent where real decisions are needed. A fixed automation for recurring routine. A simple assistant for quick lookups. Most tasks in an SME fall into the middle two categories. They need a reliable sequence, not an agent that decides on its own.
That is not a weakness, it is the honest answer. An agent is worth it when a task genuinely requires judgement and cannot be captured in clear rules: for example sorting requests that are phrased differently every time, or pulling together a proposal from several sources. Where the path is always the same, a fixed workflow is faster, cheaper and easier to maintain.
What a sensible start looks like
If you want to start, start small. A single, clearly bounded task with manageable risk. An agent that proposes results and a person approves them, rather than a solution that acts unchecked. That way you gain experience without putting a critical process at risk. This matches the picture in Switzerland: according to a BDO analysis, the gap between large companies and SMEs is widening, often because SMEs lack the time and the data base to experiment at scale. A small, concrete step beats a large project that stalls in the pilot stage.
AI agents are not hype without substance, but they are not magic either. For most SMEs, the right question is not how autonomous a solution can be, but where a bit more independence brings real value and where a reliable, fixed sequence remains the better choice.
If you would like to find out which of your processes are suited to automation and where an agent would genuinely make sense, we are happy to look at it together.
