AI claims in Swiss commercial register purpose statements, 2018–2026
Computed results and hand labels from a full-population analysis of AI claims in the Swiss commercial register. 65 output tables, 3 hand-label files, keyword regex definitions and SQL provenance. Covers the SHAB/FOSC publication feed from August 2018 to August 2026.
Download
The dataset is distributed as a zip archive containing 65 output tables, 3 hand-label files, keyword definitions and SQL provenance.
| File | Size |
|---|---|
| incorporating-intelligence-data.zip | 351 kB |
| README.md | 5.1 kB |
| LICENSE.txt | 878 B |
Citing this dataset
Sidorenko, Semion (2026). AI claims in Swiss commercial register purpose statements, 2018–2026 [dataset]. Zenodo. https://doi.org/10.5281/zenodo.22308758
Attribution is a condition of the licence. A link to this page is the attribution we ask for.
What this covers
Computed results and hand labels from a full-population analysis of AI claims in the Swiss commercial register. As of 31 August 2026, 3,310 companies carry an AI term in their registered legal purpose.
Contents
Output tables (65 CSV files): one per computed result, covering quarterly AI share, canton rate ratios, legal form mix, survival analysis, rename events, trademark overlap and keyword net sensitivity checks.
Hand labels (3 CSV files): judgements that a rerun cannot reproduce. Company-name verdicts (234 entries), embedding-threshold labels (90 entries) and compound-net exclusions (4 entries).
The keyword definitions in nets.py
are the regex patterns that define an AI claim in German, French, Italian and English.
Provenance (export.sql):
the SQL that produced the source data from the Prospex production database.
Three keyword definitions
The headline count (3,310) uses the extended net, which combines three layers:
- Loose net (2,860): spelled-out "artificial intelligence" in four languages, plus machine learning, deep learning and generative AI terms.
- Compound net (+434): hyphenated German AI/KI compounds, including l/I confusions the register contains, minus four hand-excluded false positives.
- Misspelling net (+16): misspelled variants the register actually contains.
Where it comes from
The source data is the Swiss Official Gazette of Commerce (SHAB/FOSC), the official publication of record for the Swiss commercial register. Prospex holds every publication since August 2018, including all incorporations, 137,637 dated purpose amendments and 88,872 company renames.
Key numbers
| Measure | Value |
|---|---|
| AI claimants (extended net) | 3,310 |
| AI claimants (strict net) | 2,806 |
| AI share of 2026 incorporations | 2.80% |
| Pre/post-ChatGPT ratio (founding clock) | 4.4× |
| Zug rate ratio vs national | 2.32× |
| Companies that added AI later | 603 (18.2%) |
| Median amendment lag | 2.2 years |
| Three-year survival, AI claimants | 94.6% |
| Hazard ratio (AI vs matched controls) | 0.98 (0.79–1.21) |
Mirrors
The same files are on Zenodo and Hugging Face. This page is the canonical version.
- Zenodo (DOI: 10.5281/zenodo.22308758)
- Hugging Face