Where AI pays off in an SME, and where it does not.
AI pays off in a small or mid-sized company where the work is mostly text, comes up every week and can be checked before a mistake costs anything: drafts, long documents, questions about your own material, incoming messages. It does not pay off for exact sums, binding documents sent unread, rare tasks, or decisions nobody can check.
Four questions for any task
Test every task with four questions before you spend money on it.
- Is the work mostly text: reading, writing, sorting, summarising?
- Does it come up every week in a similar form?
- Can someone who knows the work see within a minute or two whether the result is right?
- Is a mistake cheap, or caught before it leaves the company?
Where it pays off
Start with a task that gets four yes answers. These four kinds of work pass most often. In each, a person reads the result before it is used.
- Drafting from existing documents: a proposal, a reply or a report written from what is already on file, then edited by the person who sends it
- Reading long documents: finding deadlines, amounts and obligations in contracts, tenders or regulations, each with the passage it came from
- Answering from your own material: questions about products, processes or past projects, answered only from your documents and with the source shown
- Sorting and summarising incoming messages: by topic and urgency, with two lines on what the sender wants
What Swiss SMEs use it for
In March 2026 the research institute Sotomo surveyed 336 SMEs in German- and French-speaking Switzerland for the insurer AXA. 35% said they deliberately integrate AI into their processes, 39% were trying it out and 26% did not use it.
The most common uses were language work: translation (47%), correspondence (42%) and advertising copy (35%). Customer service and support stood at 17%.
The EU's statistics office counts differently. Among EU enterprises with 10 or more employees, 19.95% used at least one AI technology in 2025: 17% of the small ones and 55.03% of the large ones.
Where it does not pay off
If the answer to the third or fourth question is no, the task needs code and a person who checks as well as the model. If the answer to the second is no, leave the task as it is.
- Exact arithmetic: totals, discounts, VAT and payment plans belong in code. Anthropic's documentation for its code execution tool says the model runs code for calculations with large numbers, many steps or results where precision matters. OpenAI and Google document the same kind of tool
- Anything binding sent without review: a language model can write something false in the same fluent sentences as something true. A proposal, a contract or a reply to a complaint needs a reader before it goes out
- Tasks done twice a year: setting up, writing instructions and checking the result take longer than doing the task by hand
- Decisions nobody can check: whom to hire, which supplier to drop. If nobody can say afterwards why the answer is right, do not hand it to a model
AI plus checks: two examples
Both of my products do work that fails the fourth question: a mistake would be expensive. So the model reads, and code checks.
InkDraft turns a sales call into a proposal and a contract. The model reads the amounts from the call, and code adds them up and checks the totals.
ThesisCheck tests an investment thesis against a company's filings. Every fact needs a word-for-word quote, and code checks that the quote is in the source.
In both products a person still reads the result.
How to start small
Start with one task and one person.
- Pick one task that passes all four questions and that one person does every week
- Collect three real examples of a good result before you set anything up
- Run it for a few weeks next to the old way, and have that person check every result
- Write down every mistake and what it would have cost
- Only then decide whether the task needs more: code for the figures, a check against the source, or nothing
Questions
Do we need custom software, or is an AI assistant enough?
For drafts, summaries and questions about documents that a person reads before use, start with an assistant your team already has. Custom software is worth it where a mistake costs money: figures, contracts, and anything that goes out with no person in between.
How do we tell whether it saves time?
Measure first how long the task takes by hand and how often it comes up. Then measure the new way, including the time to check and correct. If checking takes as long as the work itself, the task fails the third question, so stop.
Will AI replace staff in a company our size?
Most SMEs in the Sotomo survey do not expect that. Of those with AI experience, 47% expected time savings with no effect on the number of jobs, and 12% expected to be able to do without staff. In the work described here, I expect the person to move from writing to checking.
Can we stop the AI from making things up?
Not completely. Anthropic's documentation lists measures that reduce it: allow the model to say it does not know, have it quote word for word from the document before it answers, and restrict it to the documents you provide. The same page says these measures do not eliminate false statements, and that critical information should always be validated.