Context mix of AI mentions in FY2025 reports ( mentions classified on a systematic sample, max 40 per document). The classification was checked by hand on a random sample of mentions — agreement (review result, Polish).
Shares exclude mentions about the economic environment (MACRO) and structural hits (tables of contents, biographies). “Substance” = the context contains a number, the name of an AI project or product, or a verifiable result. Note: WIG20 companies usually publish a management board report, DAX companies a full annual report; the different document scope may affect the context mix (methodology §7).
Same company, same document type, three consecutive annual reports. Each panel has its own scale — we compare a company with itself, not companies with each other. Frequency of mentions only, no context classification.
Is this just the new CSRD statements? From FY2024 the reports carry a sustainability section (ESRS) with its own AI paragraphs. With the ESRS pages removed, companies still show an increase (median ) — the new statements do not explain the rise. Test: test-csrd.md (Polish).
Full series with document URLs and checksums: seria-czasowa.md (Polish).
Each company against peers from the same sector — including companies from European countries selected for the same sectors. A company is compared within its sector only if it has at least 3 peers from at least 2 other countries.
Square-root scale (AI mentions per 1,000 words of the FY2025 annual report) — it stretches low values so differences among rarely-mentioning companies stay visible. Hover a dot for the company name and its most frequent context.
European companies, same method (companies with reports for all three years): companies write about AI more often in FY2025 than in FY2023, median increase ; with the ESRS pages removed — . The rise is not specific to Germany and Poland. Series: seria-czasowa-EU.md, CSRD test: test-csrd-EU.md (Polish).
Bar length = frequency of AI mentions (per 1,000 words of the FY2025 annual report); segments = the context in which the board writes about AI. Hover a bar for details.
Click a company row to expand its mentions (quote ≤25 words, page, term) — every number is auditable. * = context mix computed on a systematic sample.
| Company | Mentions | /1,000 words | Dominant context | Substance |
|---|
Every claim on this page goes through three tests: an estimator swap, removing any single company (leave-one-out) and a bootstrap over companies. The result is published with the data.
| Claim | Status |
|---|---|
| Revenue is a minority context — of mentions | robust below both cost and risk in of bootstrap draws; removing any company: |
| Most mentions carry no substance — | robust bootstrap 95%: |
| Within-company increase in frequency — companies write about AI more often in FY than in FY | robust, not unanimous independent of estimator, sector and document type; leave-one-out on the ratio of medians |
| DAX↔WIG20 frequency gap | fragile ranges depending on method; document type is almost perfectly confounded with country |
Full test: analiza-odpornosci.md (Polish).
What we measure. Frequency and context of AI mentions in the FY2025 annual reports of 20 WIG20 companies and the 20 largest DAX companies — the full Geschäftsbericht / integrated report (DAX) and the management board report or full annual report (WIG20). Documents retrieved only from official, public investor-relations pages; one language version per document. Multilingual EN/DE/PL dictionary (AI, KI, künstliche Intelligenz, sztuczna inteligencja, machine learning, GenAI, LLM…); word-boundary matching, overlapping hits counted once; normalised per 1,000 words.
Context classification (methodology v). Each mention (±2 sentences) → one class: revenue/growth, cost/efficiency, risk/compliance, talent, buzzword; plus a "substance" flag (a number, project name or concrete result — judged on the whole context). Customer-service bots count as cost, not revenue. Mentions of the economic environment and structural hits (tables of contents, biographies) count towards frequency but not towards the context mix. Classification by an LLM on a systematic sample (max 40 mentions per document); a human reviews a random ≥10% before publication — publication threshold: ≥80% class agreement.
Change over time. For the same companies, the reports for the two preceding fiscal years — same document type and language, same dictionary and code. The comparison is within each company, so sector and document-type confounders do not apply (methodology §5b).
Limitations. Purposive sample (the largest companies of both indices) — a signal, not a benchmark. Document type is almost perfectly confounded with country (DAX companies — full annual reports, most WIG20 companies — management board reports), so WIG20↔DAX differences cannot be separated from document differences; we report them with that caveat. "Mentions AI rarely" ≠ "is behind" — the report describes, it does not judge. Short quotes with attribution; no redistribution of full texts. Full methodology (including the 10-company pilot): methodology (Polish reference version: metodyka) · source data: dataset.json (every mention with its quote, page and document checksum).