Made for Swiss self-employed & SMEs
Every few weeks a new AI model tops a leaderboard, and every few weeks a client asks us the same question: which one should our business actually use? The honest answer for a Swiss SME is that for most office work the choice matters less than the marketing suggests. But it does matter — in three specific situations. Here is a calm way to decide between GPT, Claude and a local model, without re-running the comparison every month.
What actually differs between models
For everyday office work — drafting a client letter, summarising a ten-page contract, answering routine email — the leading models are close enough that you will rarely notice a difference. The gaps appear in specific corners: very long documents, complex tables, unusual formatting, and questions where the model should admit it does not know. Claude has a reputation for careful work on long documents, GPT for breadth and integrations, and local models sit a step behind on peak quality but handle routine text reliably.
One warning: any concrete ranking you read — including this one — is outdated within a few months. Treat model names as interchangeable candidates, not as long-term commitments.
Test with your language, not with English benchmarks
Benchmarks run in English. Your clients write in German and French, with Swiss conventions: formal salutations, amounts in CHF, and “Sie” and “vous” used correctly. Before you commit, run ten of your real tasks through a candidate model — an actual offer, a real reply to a complaint, a genuine French letter. Most top models handle both languages well today, but formality levels and formatting conventions are still where they trip. The model that wins an English benchmark is not automatically the model that writes your letters well.
When local is the right answer
If your work involves data you are not allowed to send to a cloud provider — patient files, legal cases, client financials protected by professional secrecy and the nDSG — a local model on your own machine removes the question entirely: nothing leaves the building. The trade-off is real: lower peak quality, some setup effort, and hardware that has to be paid for. If your documents are routine letters and summaries, a good local model is enough. If you need the strongest analysis, use a cloud model on anonymised text and keep the sensitive original local.
A three-question rule
- How sensitive is the data? If it must not leave your machine, go local — or anonymise before using a cloud model.
- How demanding is the task? Routine text: any current model will do. Complex analysis or delicate formatting: use the strongest model you can access.
- What does it cost? For a one-to-five-person business, a flat-rate subscription of roughly CHF 20–30 per seat per month is almost always cheaper than paying per request. Start there, and upgrade only when you hit a wall.
Do not marry a model
Whichever you choose, keep the exit cheap. Store your prompts in a document, not only inside one chat app. Describe your workflows in plain language so that any model can run them. Then re-test every quarter with the same ten real tasks: models improve, prices change, and yesterday’s clear winner may sit in the middle of the pack tomorrow. Switching should take thirty minutes, not a migration project.
The model is the least stable part of your AI setup. Your workflows, your document structure and your routines are the durable parts. Build those so they run on whatever model is best next quarter — and stop worrying about the leaderboard.
