An AI agent is a system where a language model does not just answer once, but plans steps, uses tools (email, calendar, databases, web) and checks its own results until a task is done. In 2026, businesses use them for bounded jobs like research, inbox triage, document processing and customer support - not for unsupervised running of a company.
A definition you can use
Anthropic draws the cleanest line. A workflow is AI following a fixed, pre-programmed path. An agent is different: the model itself decides what to do next, which tool to call, and when the task is finished. That flexibility is what makes agents useful for tasks where you cannot write down every step in advance - and also what makes them harder to control.
What agents do well in 2026
Research and summarisation: gather material from many sources, compare it, and return a sourced brief.
Inbox and message triage: read, classify, draft replies, and escalate what needs a human.
Document processing: extract data from invoices, contracts and forms, and route it into your systems.
Customer support: resolve standard requests end-to-end, hand over the rest with context.
Operations glue: watch a calendar, a mailbox or a folder and act when something arrives.
What they still cannot do
Agents make mistakes a junior employee would not, and they make them confidently. They misjudge ambiguous instructions, take wrong turns in multi-step tasks, and cannot be accountable for a decision. OpenAI's own building guide recommends starting with a single, well-defined task, clear tools, and a human checkpoint before anything irreversible - then widening scope only when that works.
The honest market picture
Gartner predicts over 40% of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls. The projects that survive tend to share three traits: a narrow, measurable task; clean access to the data the agent needs; and a human approval step for anything that spends money, sends messages, or changes records.
What this means for a Swiss SME
You do not need an "agent strategy". You need one painful, repetitive process - quote drafting, invoice matching, support triage - where a supervised agent can save real hours in the next quarter. Start there, measure it, then expand. Swiss data protection law (FADP) applies to whatever data the agent touches, so the same rules as for any AI tool apply.
Sources
Last updated: 20 October 2026.
