Where do AI agents and automation genuinely help SMEs?

AI agents, automation, chatbots: what's the difference?

The terms are often mixed up but mean different things. Classic automation follows fixed if-then rules: a trigger sets off a defined chain of steps, e.g. "new order → create invoice → send email". It is reliable and predictable, but rigid.

An AI agent additionally uses a language model (LLM) to interpret a goal, plan intermediate steps itself and call tools such as a calendar, CRM or search. It handles unstructured text and exceptions better, but is less deterministic. A chatbot or voicebot is essentially the conversational surface in front. In practice most SMEs combine both: reliable automation for the standard case, an agent for the cases that need language and judgement.

Where AI agents genuinely help SMEs

The payoff is greatest where work is repetitive, text-based and frequent, yet still needs a little interpretation. These illustrative examples show typical areas of use for a Swiss SME:

• Customer communication: pre-sort emails, answer standard enquiries, prepare multilingual reply drafts (DE/FR/IT/EN) that a staff member only has to approve.
• Quotes and proposals: turn an enquiry into a structured draft quote with line items and prices drawn from your own catalogue.
• Scheduling and dispatch: coordinate appointment proposals, send confirmations and reminders, rebook cancellations.
• Documents and knowledge: extract data from invoices and delivery notes, summarise contracts, answer internal questions from manuals (retrieval).
• Marketing and sales: draft product copy and social posts in the four national languages, enrich and qualify leads.
• Accounting and back office: pre-code receipts, check expenses, compile recurring reports (with a human final check).

Realistic starting points: how an SME begins

Don't start with the biggest, most visible process, but with a clearly bounded one that occurs often and whose mistakes are cheap to correct. A proven path in five steps:

• 1. Pick the process: high volume, lots of text, clear rules, low risk (e.g. first-line support answers rather than contract approvals).
• 2. Define success: set one measurable metric up front, such as handling time, follow-up rate or share of approved drafts.
• 3. Tidy the data: the agent is only as good as its sources. An up-to-date catalogue, clean templates and a well-kept CRM are the precondition.
• 4. Human in the loop: let the agent only produce drafts at first (copilot mode). Autonomy is widened step by step once quality holds up.
• 5. Pilot small, then roll out: test for one or two weeks with one team, measure results against the metric from step 2, then decide.

Common mistakes and how to avoid them

• Starting too big: an agent meant to do "everything" will do nothing reliably. A narrow, clearly defined use case wins.
• No human control on risky steps: payments, contract approvals or binding customer commitments should never be fully automatic, but behind an approval.
• Ignoring hallucinations: language models occasionally invent plausible-sounding but false statements. Facts should be retrieved from reliable sources rather than freely generated.
• Considering data protection too late: feeding customer data into a tool without clarifying the legal basis, server location and data processing agreement risks breaches. That belongs at the start, not the end.
• Not measuring success: without a metric the benefit stays a gut feeling. A before-and-after comparison creates clarity and buy-in in the team.
• Not bringing staff along: automation introduced without explanation breeds resistance. Showing the benefit for daily work wins allies.

The Swiss angle: data protection and governance

In Switzerland the revised Federal Act on Data Protection (revFADP) applies. As soon as an agent processes personal data of customers, staff or suppliers, you need a valid legal basis, must inform transparently and observe data minimisation. The FDPIC is the supervisory authority.

Two points matter especially for SMEs. First: many AI services run on servers abroad. Clarify server location, the data processing agreement and whether data is used for training; the latter can often be excluded contractually. Second: the EU AI Act has extraterritorial effect and can also affect Swiss firms that operate in or supply to the EU. It classifies AI applications by risk and, for certain systems, requires transparency, such as telling users they are talking to an AI and not a human.

In practice this means: use sensitive data sparingly, log access and approvals, always have binding or legally relevant decisions confirmed by a human, and choose a provider partly on its data-protection commitments.