Transparency · data & method

The data, and what we do with it

Three things that are often confused: what you enter in the tool, what we measure about visits to the site, and what serves as the reference population for placing a score. This page separates them, and points to the privacy policy for the legal detail rather than copying it here.

1. What you enter, what we measure

Your answers to the questionnaire stay in your browser's local storage: we have no user database, and we keep no contact list. Three functions send data to our server, and only when you click their button: the AI drafting of the report and the pre-analysis of a website, which go through the Claude API (Anthropic), and sending the report by email, which goes through our email provider. The detail, provider by provider and country by country, including what each one keeps, is in our privacy policy.

For visits to the site, a single audience measurement: cookie-free statistics, of which we only see totals (pages viewed, referring page, country, region or city, device and browser type). They count a single click, the one on the “Book a debrief” button, and never the content of your assessments. Like any host, Vercel also keeps a technical log of each request, IP address included, which we only consult in the event of an incident.

2. The reference panel: 36 SMEs, and they are fictitious

To place a score (sector average, percentile), the tool compares it with a panel of 36 companies. This panel is currently entirely synthetic: 36 fictitious profiles across four sector families (a company from another sector is compared with the whole panel), with dimension scores chosen by hand to represent a realistic range of maturity. They are not clients, surveys or observed data, and these SMEs have no name: a plausible name next to a score would read as a real company. We say so wherever the panel appears, because a comparison that hid its nature would be misleading. It will be replaced step by step by the results of real pilot companies, each with the company's written consent, as publishing a score on our page of published scores already requires.

Source: 36 rows of fixed values in the site's code, not a database. They do not change from one visit to the next.

3. Demonstration: classifying the panel with k-means

An automatic classification (k-means algorithm, three groups) applied to the four dimension scores of the synthetic panel above, and not to real client data, for the same reasons as in point 2. The algorithm receives no labels: it only groups the 36 profiles that are most alike on the four axes T, C, E, S; we then sort and name the three groups from the lowest average to the highest.

Method: Euclidean distance on the four dimension scores (0 to 100 each), three starting centres chosen by the “farthest point” rule, then reassignments until stable. The calculation involves no randomness: it gives the same groups every time the page is displayed.

The figure shown for each SME is the simple average of its four dimension scores (0 to 100). It is not the KarmaScore used elsewhere on the site, which is weighted by sector with a cap of 40 % per axis (see the methodology).

4. A Swiss directory, not a joint laboratory

AI-Karma has an entry on SAIROP (sairop.swiss), the platform that maps Swiss research in artificial intelligence, coordinated by the SATW (Swiss Academy of Engineering Sciences). To be precise about what this says and does not say: it is an organisation entry describing one service, not a membership, not a label, and not research carried out jointly. It places the tool in a neutral directory, next to universities, research institutes and companies, with the same reservations as everywhere else on this site (an information tool, a self-declared score that has not been audited).

The same directory also lists Swiss research projects in AI. We claim no collaboration with this work to date, only the same directory.