"Agent" has become the default label for almost every AI product released this year. For a business owner, that makes it hard to tell what is new and what is a chatbot with a fresh name. This article separates the two, using recent reporting and research, and sets out where an agent is worth piloting in a small Swiss firm.

What an agent is, and what it is not

An AI agent does more than answer a question. It carries out a sequence of steps, such as reading a record, checking it against another system and preparing an action, with less step-by-step instruction from a person. At the start of the year, Euronews Next's outlook for 2026 described agents as still fairly limited: one might plan a holiday, but a person still had to click to buy the flight. That division of labour, where the system prepares and a human commits, remains the right mental model for most SMEs.

Not everything sold as an agent qualifies. In mid-2025, Gartner told Reuters that many vendors were engaged in agent washing, meaning they had rebranded assistants and chatbots, and estimated that only about 130 of the thousands of vendors in the space were offering the genuine article. The firm also forecast that more than 40% of agentic AI projects would be cancelled by the end of 2027, citing rising costs and unclear business value. Its senior director analyst, Anushree Verma, described most projects as "early stage experiments or proofs of concept".

None of this means agents are useless. It means the buyer has to do more of the sorting.

Where agents perform well

The tasks that suit agents today share a profile: repetitive, rule-based, spread across a few systems, and easy to check. At a Google Cloud event in Doha this autumn, Alex Rutter, who runs the company's AI business across Europe, the Middle East and Africa, used invoice reconciliation as his example: an invoice is checked against payment and supplier records before payment is approved. He described agents as well suited to highly repetitive database work, with people reviewing the exceptions rather than every record.

For a Swiss SME, comparable tasks include:

  • matching supplier invoices to purchase orders and delivery notes;
  • checking incoming forms or applications for missing fields before a person sees them;
  • updating CRM records from structured enquiry emails;
  • compiling a weekly status summary from project and ticketing tools.

Each has a clear correct answer, a limited set of systems involved, and a person who can spot an error quickly.

Where they still fall short

Three limitations appear consistently in this year's reporting.

Long, open-ended tasks. An Anthropic analysis of how its model is used at work, reported by Euronews Next, found that success rates fall as tasks become longer or more difficult, which reduces the time people actually save. Once the authors accounted for errors and the human effort needed to correct them, their productivity estimates became more cautious. The practical lesson is to break work into short steps, each of which can be verified.

Rules nobody wrote down. Aithos, a Dutch non-profit, tested 12 agent models against scenarios based on the EU AI Act and GDPR. According to Euronews, the best performer complied in 54% of scenarios and the weakest in 7%. The models were not told to follow the law, and that is the point: an agent will not infer your legal or contractual obligations on its own. They have to be built into its instructions and, more importantly, into what it is technically allowed to do.

Access and security. Cisco's president, Jeetu Patel, told Euronews that agents acting on someone's behalf should be vetted much as a new employee would be. For a small firm, the equivalent is simpler: an agent gets the minimum access its task requires. Under the Swiss Federal Act on Data Protection, you should also be able to say which personal data it processes and where.

A checklist before you pilot an agent

  1. Can you describe the task in fewer than ten steps? If not, it is probably too open-ended for a first pilot.
  2. Is every output checkable? A reconciliation can be verified line by line; a "market analysis" cannot.
  3. What can it read, and what can it change? List these separately. Read access is far less risky than write access.
  4. Where does it always stop? Name the actions that need a person: sending, paying, signing, deleting.
  5. Who handles exceptions, and how quickly? An agent that flags problems nobody reviews creates a backlog, not a saving.
  6. What will you measure? Hours per week, error rate or turnaround time, recorded before the pilot begins.
  7. Can you switch it off? Keep the manual process documented and working until the pilot has proved itself.

The realistic position

Agents are useful in 2026, but in a narrower band than the marketing suggests. The firms that benefit treat them as a careful junior colleague with written instructions and limited permissions, not as an autonomous department. Start with one reconciliation-style task, keep the approval step, and let measured results decide whether the scope grows.

Wondering which of your processes suits an agent and which does not? The free AI Readiness Audit maps your recurring tasks and ranks what is worth automating first.

Sources

  • Euronews Next, "From 'AI slop' to world models, bubbles and small models: What to expect from AI in 2026", 1 January 2026. link
  • Reuters (republished by The Star), "Over 40% of agentic AI projects will be scrapped by 2027, Gartner says", 25 June 2025. link
  • Euronews, "From answering questions to carrying out complex tasks: The next phase of AI", 28 September 2026. link
  • Euronews Next, "Will AI replace jobs? Anthropic report finds the answer is not so straightforward", 19 January 2026. link
  • Euronews, "AI agents actively ignore EU law to achieve goals, study finds", 2 June 2026. link
  • Euronews, "Cisco president warns AI agents need 'background checks' like human employees", 11 February 2026. link