Where Swiss SMEs actually stand with AI, October 2024
A short status report: what changed in twelve months, what did not, and the three questions worth answering before you buy anything.
A year ago the AI question in most Swiss SMEs was simple: are we allowed to use ChatGPT? In October 2024 the question is harder, and more useful. Which of our daily tasks is a language model genuinely good at, and what happens to our data on the way there?
This blog will follow that second question. Short posts, dated, no hype, written for companies with ten to two hundred people rather than for research labs.
What changed in twelve months
Models started thinking before answering. On 12 September 2024 OpenAI released o1, the first widely available model that spends extra time working through a problem before it replies. It is slower and dearer, and on multi-step tasks it is markedly better.
The cost per answer fell. At its developer conference on 1 October 2024, OpenAI introduced prompt caching: a 50 per cent discount on input text the model has recently seen. For an application that resends the same context all day, that is not a rounding error.
Small models became usable. Meta’s Llama 3.2, released on 25 September 2024, includes versions small enough to run on a laptop. Running a model on your own hardware stopped being a research project.
What did not change
The unglamorous parts of a Swiss SME are exactly where they were. Master data still lives in five systems. Quotations are still assembled by hand from an old Word file. Supplier invoices are still typed into the accounting package one line at a time.
A model that writes elegant German does not know your prices, your delivery terms, or which customer is three reminders deep. Capability alone does not close that gap.
Three questions before you buy anything
- Which task, precisely? Name one. « AI for the company » is not a project, it is a budget line with no owner.
- Where does the data go? Which provider, in which country, kept for how long, and used for training or not.
- Who checks the output? An answer nobody verifies is not time saved. It is risk that changed desk.
If you cannot answer all three in two minutes, the pilot is not ready, whatever the vendor demo looked like.
Next post: why 1 October mattered more for your budget than any model launch this year.