Artificial intelligence arrived in SMEs from the wrong end: the website chatbot first, because it is visible, when it is almost always the lowest-yield use. Here is what we observe in the field, after integrating these tools into several Swiss companies.
What produces a measurable return
Processing incoming documents
By far the best effort-to-gain ratio. Supplier invoices, delivery notes, handwritten forms: a model extracts the fields, a human validates. You go from three minutes of data entry per document to fifteen seconds of checking.
The key point: keep the human on validation. A 95% extraction rate is excellent, and catastrophic if the remaining 5% goes straight into the accounts.
Assisted writing on your own data
Tender responses, product descriptions, spec sheets, translations. The gain is real as soon as the model works on your documents rather than its general knowledge: otherwise it produces text that is plausible and wrong.
Internal search
A thirty-person company accumulates thousands of documents nobody can find. A semantic search engine connected to your folders, able to answer "where is the customer returns procedure?", saves time nobody was measuring because it was spread across everyone.
What almost always disappoints
The general-purpose chatbot on a brochure site
On an SME site with a few hundred visitors a day, the chatbot answers questions whose answer was already on the page. It adds a heavy script, degrades performance, and frustrates visitors who just wanted a phone number.
It becomes relevant above a certain volume, or when it does something specific: check availability, track an order, qualify a request. In other words, when it stops being a chatbot and becomes a tool.
"Autonomous agents" acting without supervision
The technology works in a demo. In production, on processes that commit the company, the error rate remains too high to remove human validation. An agent that sends a wrong quote to a customer costs more than everything it saved.
Bulk SEO content generation
Publishing forty generated articles improves your output statistics, not your rankings. Search engines have become effective at detecting content with no added value, and a volume of weak pages can drag down how the whole domain is perceived.
How to approach it
Three questions, in this order:
- Which repetitive task takes the most time? Not the most visible: the most frequent.
- Does that task have a clear input and output? If so, it can be automated. If not, no AI will rescue a vague process.
- What happens if the result is wrong? That answer sets the level of supervision, and therefore the real gain.
What it costs
A targeted integration (one process, with human validation) lands between CHF 8,000 and 25,000, plus tens to hundreds of francs a month in API usage depending on volume.
The question is not "should we use AI?" It is "which specific task, measured in hours, do you want back?" Without that sentence, the project has no success criterion.
