Made for Swiss self-employed & SMEs
AI demos are impressive. Then you run the same tool on your own invoices, contracts or client emails — and it quietly gets something wrong. These failures are not random. They are predictable, and knowing them is the difference between a useful assistant and an expensive source of mistakes.
The failures you will actually run into
Numbers. Ask an AI to total a long expense table, split an invoice or check a VAT rate, and it can hand you a confident, wrong figure. Language models produce plausible arithmetic, not verified arithmetic. Anything that touches money stays a draft until a human or a spreadsheet confirms it.
Sources. Models sometimes invent citations: a court ruling, an article number of the Code of Obligations, a study that does not exist. The sentence looks perfect; the reference is fiction. If a claim matters — legally, financially, medically — open the source yourself before you repeat it.
Context. Every model has a hard limit on how much text it genuinely holds in mind. A 120-page contract, a year of correspondence, a full payroll file: the AI may read the opening well and miss the clause on page 40. Splitting a document into chunks helps, but cross-references between the chunks get lost. Long documents need targeted questions, not “summarise everything”.
Chains of steps. An AI agent that plans and executes drifts: it takes a wrong turn early and keeps going confidently. That is tolerable for research; it is not tolerable for booking payments or sending mail on its own. Every chain of AI steps needs a human checkpoint before anything leaves the building.
Responsibility. AI cannot sign anything, cannot be held liable, and cannot exercise professional discretion. It can draft the letter; you decide whether it is sent. This is not a limit that next year’s model will fix — it is the nature of the tool.
Where AI is genuinely reliable today
Drafting and restructuring text. Summarising documents you already know. Translating between German, French and English. Turning messy notes into a clean structure, or a technical explanation into something a client actually understands. Here AI saves real hours and rarely embarrasses you — because you review the result anyway, and the mistakes are easy to catch.
“AI does not fail at random. It fails in predictable places: numbers, sources, long documents, chains of steps and anything that carries responsibility. Put a human between the AI and those five things, and the rest is leverage.”
A simple trust test before you hand off a task
- Take five real cases where you already know the correct answer.
- Run the AI on them exactly as you would in daily work.
- Count the errors — and estimate how long each one would take to catch.
- Hand off the task only if checking the output is faster than doing the work yourself.
Keep a short list of the tasks where AI failed, and why. Models improve quickly, and a limit that held in spring may be gone by winter — but it disappears for the people who retest, not for the people who gave up.
What this means for your business
Use AI where it is strong: drafting, summarising, translating, structuring. Keep a human on the five weak spots — or at least a verification step in between. That is the whole strategy, and it is an honest one: not “AI can do everything”, and not “AI is useless”, but “AI is strong exactly here, and weak exactly there”.
