How Lyell Data models AI adoption & usage

Executive summary

Scope & definitions

Industry classes

Industries are mapped to ISIC/NAICS families. Country pages may show local classifications with a mapping table. Example groups: IT Services; BFSI; Telecom; Manufacturing (and splits); Retail & E-commerce; Logistics; Healthcare; Education; Public Sector; Agriculture.

Out of scope

• Micro-firms (<10 employees) unless explicitly covered by a survey anchor.
• Hobby/individual usage for adoption metrics (still captured in usage/MAU where relevant).
• Proprietary vendor telemetry not available as public signals.

Data sources (public signals)

We blend multiple open signals. Each source is assessed for coverage, bias, and update cadence.

Modeling framework

4.1 Adoption score (by country × industry)

1. Normalize signals. For each signal s_k we compute a regional/sectoral normalization (z-score or min-max) to reduce cross-country scale effects: ŝ_k = norm(s_k | region, sector).
2. Weighted fusion. Combine normalized signals by learned weights w_k: A* = Σ (w_k · ŝ_k).
3. Anchors & calibration. Scale and shift using anchor jurisdictions/sectors with official stats: A = α_sector · A* + β_country.
4. Size split. Apply firm-size offset: A_SME = A − δ_size, A_Large = A + δ_size (δ depends on sector and country structure).
5. Bounds. Clamp to [0,100].

4.2 GenAI/ChatGPT usage (MAU)

1. Traffic triangulation. Web/app traffic, device shares, and session heuristics → preliminary user counts.
2. Deduplication heuristics. Cross-device, VPN/shared IP adjustments, and app/web overlap correction.
3. Panel calibration. Where available, calibrate to public panel reference points; otherwise regional analogues.
4. Professional tilt. Adjust for enterprise usage share using postings mix and enterprise signals.

Uncertainty & modeled ranges

We reflect uncertainty as a band around the point estimate.

1. Signal variance. Each signal contributes a variance σ_k² estimated from rolling windows or cross-section dispersion.
2. Composite standard error (SE). SE = √( Σ (w_k² · σ_k²) ).
3. Range width. [A − γ·SE, A + γ·SE] with γ in [1.0, 1.4] depending on coverage; widen if coverage < 60%.
4. Clamping. Clip to valid domain (e.g., 0–100%).
5. Trend bands. Charts use the same SE + a small revision-risk premium for newer months.

Temporal methods

• Smoothing. Light EMA or Holt smoothing for noisy signals; preserve turning points.
• Nowcasting. When anchors lag, extrapolate with short-term signals; replace with actuals upon release.
• YoY & revisions. YoY computed on revised history; we maintain a changelog of back-revisions.

Quality assurance & governance

Versioning & revisions

• Semantic versioning: major.minor.patch
• Back-revisions documented with reasons (new anchors, bugfix, re-weight)
• Stable URLs for CSVs; metadata includes method_version

Ethics & privacy

• No PII; only aggregated, public or derivative signals
• Respect for national data laws (e.g., GDPR/LGPD); remove or mask where required
• Vendor telemetry used only if publicly disclosed or licensed

Known limitations & biases

• Access bias. Internet connectivity, censorship, and payment rails can under/over-represent users.
• Enterprise skew. Job postings lean toward larger firms; SME adoption may be understated.
• Language/culture. GenAI prompts and product choices vary by language; cross-country comparability is approximate.
• Early adopter optics. Low base countries can show large YoY changes that overstate maturity.

How to cite & reuse

Content and CSVs are licensed under CC BY 4.0. Please attribute as follows:

Lyell Data (2025). AI Adoption in <Country> — 2025. https://example.com/ai/industry/<slug>

Academic (APA):

Lyell Data. (2025). AI Adoption in <Country> — 2025 (Version <x.y.z>) [Dataset]. CC BY 4.0. https://example.com/ai/industry/<slug>

CSV schema

Authorship & review

FAQ

Is the “modeled range” a confidence interval?

No. It’s a modeled uncertainty band that blends statistical variance with coverage and revision risk.

Why does my country show a wide range?

Likely due to sparse anchors or high disagreement among signals. As coverage improves, ranges narrow.

How often do you revise datasets?

Minor: rolling; Major: monthly/quarterly depending on anchors. See the changelog.

Can I get raw signal slices?

Yes—contact us for a license where permissible; we only ship aggregated, non-PII views by default.

Appendix

Formulas (compact)

Normalize:   ŝ_k = norm(s_k | region, sector)
Fusion:      A*  = Σ w_k · ŝ_k
Calibrate:   A   = α_sector · A* + β_country
Size split:  A_SME = A − δ_size ; A_Large = A + δ_size
Uncertainty: SE = √( Σ w_k² · σ_k² ), Range = [A − γ·SE, A + γ·SE]
        

ISIC ↔ NAICS quick mapping (examples)

Evidence table (template)

Changelog (method_version)

• 1.1.0 — Added revision-risk premium in trend bands; clarified SME threshold text. 2025-08-08
• 1.0.0 — Initial public release: adoption & usage frameworks, range model, QA checks. 2025-07-20