AI for SMEs: six use cases that pay off.

A use case pays off when three conditions hold at once: the task comes up at least twenty times a month, a human reads the result anyway, and the information needed already exists in usable form. If one of the three is missing, the use case moves down the list, however good it sounds.

Case volume is the hardest of the three, because you cannot argue your way around it. Two hours saved a week is a noticeable amount in a ten-person business. Two hours a quarter is not. So do the arithmetic before every attempt: cases per month times minutes saved per case. If that comes to less than half a working day a month, the task is not a candidate.

The second condition — a human reads the result anyway — is your best protection at the start. As long as someone checks a draft before it goes out, a mistake by the system costs seconds of correction. It only becomes expensive once unchecked output reaches clients. The third condition is about the starting position: prices in a maintained list, text blocks in an ordered folder, contracts as searchable files rather than scans with no text layer. Where that foundation is missing, tidying up is the first step, not a second-best option.

The quickest way in is an assistant that turns an enquiry, a price list and earlier quotes into a draft, which a person then checks and releases. The point is not the finished quote. It is the first eighty per cent of the text: the line items, the phrasing, the structure.

Effort: low to medium. Without any system integration you need a business subscription, a clean price list and three to five good example quotes as templates — that is an afternoon's work. Connect it to an ERP or accounting system and it becomes a small project.

Realistic benefit: businesses typically report that pure writing time roughly halves, while checking time stays. At forty quotes a month and forty-five minutes each to begin with, a saving in the region of twelve to eighteen hours a month is plausible. The second effect is often the bigger one: response speed. Quote on the same day and you win work that otherwise goes to whoever was quicker.

Risk: wrong prices and invented line items. Language models phrase things convincingly even when a number was never in the source. The countermeasure is unglamorous: prices come from one maintained list, and every quote is read by a person before it goes out.

Email triage means incoming messages get categorised automatically, tagged with an urgency and assigned to the right person — the decision stays with a human, the sorting does not. The benefit does not come from answering faster. It comes from no longer searching.

Effort: medium. The categories have to be defined first and tested against real messages from the past months, otherwise you end up with a taxonomy nobody recognises. Reckon on a week of lead time for definition and testing, plus an integration with your mail system that needs technical support.

Realistic benefit: largest in businesses with a general catch-all address handling twenty to eighty messages a day. The gain is in throughput time: enquiries reach the right person hours rather than days after arriving. Time saved per message is small; summed across everyone involved it is substantial.

Risk: an urgent message filed in the wrong place. So every implementation needs a rule that escalates when in doubt rather than filing, plus a residual category a person reviews daily. A system that never puts anything in the residual category is suspicious, not good.

Instead of digging through folders, staff ask a question in ordinary language and get an answer with a pointer to the source document. Technically this runs on a searchable index of your own files that a language model queries at answer time — the term for it is retrieval augmented generation.

Effort: medium to high, and almost all of it sits in the preparation. Scanned PDFs with no text layer are pictures and have to be recognised. Documents spread across five locations with three versions of the same contract have to be consolidated. This work happens once, nobody enjoys it, and it decides the outcome more than any choice of technology.

Realistic benefit: high in businesses with a body of knowledge that has grown over years — manuals, contracts, technical data sheets, minutes. New staff become self-sufficient faster, and the stream of questions to the one person who knows everything thins out. That second effect is hard to measure and often the larger one day to day.

Risk: outdated sources, quoted convincingly. Without a process that keeps the material current, the system will serve neatly formatted answers from an old state of the world after a few months. The second risk is access rights: if the index does not reflect who may see which document, a search aid turns into a data protection incident.

This is the stage before a booking: taking enquiries, checking availability, proposing slots, chasing callbacks — and catching everything outside office hours that would otherwise be a backlog the next morning. The calendar itself stays the single source of truth; the system proposes, it does not decide.

Effort: low if an online calendar with clean availability already exists and enquiries arrive through a form. Considerably higher once the telephone is involved, because voice systems need setup, testing and ongoing supervision.

Realistic benefit: greatest in businesses whose core trade is appointments and who receive many enquiries outside office hours. The gain is less the minute saved than the enquiry not lost: someone who asks in the evening and gets a proposal straight away does not ring the next supplier.

Risk: double bookings and an impersonal impression. The calendar has to remain the single source of truth, and there must be a route to a human available at any time. A system that makes it hard to get out costs more in client relationships than it saves in time.

A recording or transcript of a meeting becomes a summary with decisions, tasks and owners. The benefit is not the elegance of the minutes. It is that minutes exist at all — in most SMEs that is not the normal state of affairs.

Effort: low. The function is already built into common meeting tools and assistant subscriptions. What it mainly needs is a decision about where the recordings live and for how long. For client conversations, add the consent of everyone involved, obtained and documented before recording.

Realistic benefit: between ten and twenty minutes per meeting, plus the harder-to-measure effect that tasks get captured with a name and a deadline rather than merely discussed. It works particularly well for handovers between shifts, sites or account managers where the handover used to be verbal.

Risk: confidentiality and misattribution. Recordings of HR conversations, contract negotiations or discussions involving particularly sensitive data do not belong in an arbitrary tool. And a summary that attributes a decision to the wrong person does more damage than no summary at all. So: get it approved before it gets distributed.

First-pass translation for your website, product descriptions, instructions and standard correspondence — as a draft that someone with the language reworks, not as a finished result. For a Swiss SME moving from the German-speaking part into Romandie or into an English-speaking market, this is the cheapest first step.

Effort: very low for the draft, noticeable for quality assurance. You need a terminology list with the twenty to fifty technical terms of your trade, and a person who commands the target language. Without that person you get text that is grammatically correct and still sounds wrong in the market.

Realistic benefit: translation costs drop sharply, because translation turns into editing. Reach matters more: content that stayed monolingual on cost grounds becomes bilingual at all. One caveat applies to findability — a straight translation rarely ranks, because search terms and habits differ by language area. What that takes is described under visibility in AI answers and search.

Risk: legal texts and binding commitments. Terms and conditions, privacy statements, warranty and liability wording do not belong in a machine first draft without legal review — a mistranslated obligation still binds you.

AI replaces no decision, no line of responsibility and no missing demand — and it does not tidy up your files, it assumes they are already tidy. These four limits explain most of the disappointed expectations in small businesses.

• No substitute for professional judgement. A system proposes what was usual in similar cases. Whether that is right in the case in front of you is still decided by a person who carries the responsibility.
• No fix for organisational problems. If enquiries sit unanswered because ownership is unclear, automation only speeds up the journey to the bottleneck.
• No substitute for demand. If too few enquiries arrive, handling them faster gains you nothing. Then local visibility and your offer are the subject, not efficiency.
• Not a clean-up service. Unstructured filing, contradictory price lists and scans with no text layer remain a problem. A model does not solve them, it only makes them more expensively visible.
• No headcount savings at the push of a button. In businesses of this size the gain shows up as time recovered for work that had been left undone, not as a post removed. Justify the move with redundancies and you lose the cooperation you need to make it work.

Use of AI in Switzerland falls under the revised Federal Act on Data Protection. It is written technology-neutrally and therefore applies to AI-assisted processing without any special provision. As of August 2026 there is no dedicated Swiss AI act; a consultation draft implementing the Council of Europe convention has been announced.

In practice that means three duties. First, transparency: the people concerned must be able to see that their data is processed and for what purpose. Second, documentation: AI-assisted processing of personal data belongs in the register of processing activities with purpose, data categories, recipients and retention period. Small companies are partly exempt from the register requirement, but not for high-risk processing. Third, a data protection impact assessment where a high risk can arise — extensive profiling, say, or particularly sensitive data.

Then there is the contractual side. Check whether the tier you have chosen rules out the use of your input for training, and where the processing takes place. Business tiers usually regulate this; free accounts often do not. Anyone placing output on the market in the EU also falls under the EU AI Act, whose obligations for high-risk systems were pushed back in May 2026. Even so, the most effective first step for an SME is not an expert opinion but a one-page internal policy: which data may go in, which may not, and who decides in case of doubt. Binding advice on your situation comes from your lawyer — this page places the question, it does not answer it for you.

The most reliable start takes ninety days: thirty days with one single use case tried by one person, thirty days extending it to a team with measured numbers, thirty days deciding between build-out and stop. Start broader and you measure nothing, and learn correspondingly little.

In the first thirty days one person picks a task with high volume and high tolerance for error, works with it for four weeks, and writes down what worked and what did not. In the second thirty days the team joins and two numbers get measured: cases per week going through the new route, and time per case compared with before. In the third thirty days you decide — build out with system integration, keep it as a plain subscription, or stop. Stopping after ninety days is a good outcome when it rests on numbers.

On costs: tool subscriptions sit in the double-digit franc range per person per month and are usually enough for text work. As soon as your own data, existing systems or access rights come in, you have a project. At DLM Digital the sixty-minute first conversation is free, analysis and guidance are billed by time spent with a ceiling fixed in writing beforehand, a validation prototype starts at CHF 10,000, and a web application runs between CHF 3,000 and CHF 25,000. How such a prototype gets cut is described in our guide to MVP development; a full engagement is set out at AI Consulting Switzerland.