Saphan

RESEARCH / 6 SEPTEMBER 2026

Sometimes a link is missing.
Not one more model.

Business surveys talk about knowledge, data, compatibility and use cases. They invite us to look for where the work gets lost before choosing a technology.

SAPHAN’S READING AND INTERPRETATION

The need is not always visible.

The OECD survey of more than 5.000 SMEs in seven countries, carried out in late 2024, shows that how well AI fits the work, skills and what happens to the information are among the reported obstacles. It does not cover Switzerland and measures statements, not a causal effect on productivity. OECD, 2025 publication.

Our reading: before proposing an AI, look at real work and a first useful result. The trial may conclude that a rule or an existing tool is enough.

The last mile matters.

In the 2025 Eurostat data, lack of expertise, data quality and incompatibility with existing systems are among the obstacles reported by businesses that do not use AI but have considered it. This group represents neither all businesses nor micro-enterprises: the survey covers businesses with at least ten people in the sectors selected. Eurostat, 2026 edition, pages 34–36.

Our reading: producing an answer is only part of the work. The destination format, the access, the exceptions and the recovery can decide the real usefulness.

The Swiss picture is not just a single rate.

A regional survey in the Biel/Bienne-Seeland area collected 207 valid responses out of 879 mailings in 2025. Respondents cite in particular data security, understanding and technical complexity. This is a regional signal, with possible self-selection, not a representative measure of all Swiss SMEs. Survey by the Biel/Bienne-Seeland associations, December 2025.

A KOF/ETH article also distinguishes the individual use of tools from the obstacles to using data in a company surveyed. Its historical statistics do not describe current adoption. KOF/ETH, February 2025.

What this changes for Saphan.

An entry point through the situation, not through a catalog of models. An example you can test, not just an explanation. And a result that can reach an existing tool, with what is needed to take it over.

These choices are our interpretation of the sources. They are neither a study of your company nor proof of a Saphan result.