Methodology

How this report was built.

One data provider, one primary segment, one time window, and a five-step taxonomy pipeline validated across four peer-reviewed studies. Read it once; trust the rest of the report accordingly.

A · Data foundation

Data source, segment, and coverage

The entire report draws on individual-level workforce data from Revelio Labs, accessed via Wharton Research Data Services (WRDS). Revelio assembles professional profiles from LinkedIn, XING and similar networks, and enriches them with machine-imputed salaries, predicted gender, education records, and standardized occupational classifications.

The report uses the austria_located segment — all workers physically located in Austria, regardless of nationality. It is comparable to Eurostat's "place of work" employment definition. The Austria-vs-abroad / diaspora lens in chapters 3 and 5 uses the total segment (all Austrian-educated workers regardless of current location).

Time period: 2018–2025 (2025 preliminary — Revelio's collection lag means some 2025 hires and moves are not yet observed; anomalies in the 2025 column are artefacts of this lag, not true labour-market events). Austrian coverage: roughly 22.5% of Eurostat official employment, reflecting XING and LinkedIn prevalence in the DACH region — within-country trends are robust; absolute cross-country headcount comparisons should carry the coverage caveat.

Salary metric: median machine-imputed annual EUR salary (ECB annual averages). Imputation carries individual-level uncertainty but produces aggregate medians that align closely with published Austrian salary surveys. Gender classification: machine-predicted by Revelio at ~95% aggregate accuracy; binary; all figures are machine classifications, not self-reported identities.

Data use

All figures published in this report are aggregated statistics derived from Revelio Labs individual-level employment data, accessed under academic licence via Wharton Research Data Services (WRDS). The underlying user-level records are not redistributed; this site publishes derived headcounts, shares, medians, and longitudinal trends only. Aggregations are based on the Vienna University of Economics and Business (WU Wien) WRDS subscription; reuse of the figures should cite this report and Revelio Labs as the upstream data provider.

B · Taxonomy construction

The five-step pipeline — identical across four co-authored studies

The Core AI taxonomy used in this report is not ad hoc. It is the same role universe that the authors developed and validated in four co-authored research papers — specialised here to the Austrian labour market. The pipeline is identical across all four studies:

STEP 01
Revelio v3 base taxonomy
Start from Revelio's standardized occupational classification system: ~17,000 fine-grained role categories mapped consistently across firms and over time.
17,000 → candidate set
STEP 02
Skill- and keyword-based filtering
Identify the candidate universe of AI-related roles using a combination of skill endorsements and title keywords drawn from prior AI-workforce research (Alekseeva et al. 2021; Acemoglu et al. 2022; Babina et al. 2024).
STEP 03
Ensemble coding
Combine rule-based classifiers with LLM-assisted coding to produce an initial tier assignment (Build / Enable / Integrate) for every candidate role.
STEP 04
Independent expert validation
Two to three independent academic experts in AI and data science (external to the author team) manually review every role. Disagreements are resolved through discussion.
STEP 05
Inter-rater reliability
Agreement is quantified via Cohen's κ = 0.82 (two-coder protocol) and Fleiss' κ = 0.84 (three-coder protocol). Both indicate "near-perfect agreement" (Landis & Koch 1977).
→ 315–369 AI roles · 3 functional tiers

How the three research taxonomies map to this report

The three co-authored papers each group the same role universe slightly differently for their specific research question. All three are internally consistent with the Build / Enable / Integrate structure used throughout this report.

This report (Austria)	Build, Apply, Operate (2026)	ICIS 2026	PF & AI Talent (2026)
Build	Build (25 roles)	AI Invention (42)	Deep AI
Enable	Operate (60)	AI Scaling / Governance (96)	AI-support
Integrate	Apply (257)	AI Application / Design (231)	AI-infused
Adjacent	— (out of scope)	—	—

"Core AI" = ~120 roles with include_in_core=1 across Build / Enable / Integrate. "Full AI" adds the outer Adjacent ring (BI, domain analytics, decision support) and covers ~370 roles in total.

C · Validation evidence

Why we trust the numbers

κ = 0.82–0.84

Inter-rater reliability

Expert agreement on role classifications reaches κ = 0.82–0.84 across replications — the threshold Landis & Koch (1977) label as "near-perfect".

External validation

Cai et al. (2024) · Liang et al. (2025)

Revelio's firm-level aggregates correlate strongly with firm-reported SEC disclosures (Cai, Chen, Rajgopal & Azinovic-Yang 2024) and track official labor statistics closely across industries and over time (Liang, Lourie, Nekrasov & Shevlin 2025).

Peer uptake

Strategy · IS · Management

The same workforce-intelligence source has been adopted by peer-reviewed strategy, IS, and management research (Babina et al. 2024; Marchetti & Puranam 2026; Tambe 2025), reflecting methodological maturity.

Caveats

Revelio is not a census — coverage varies across countries and over time.
2025 is preliminary; year-over-year comparisons involving 2025 should be read with Revelio's collection lag in mind.
Machine-predicted gender is binary — non-binary identities are not separately represented.
Salary data are imputed medians; reliable at the cohort level, not for individual comparisons.
Brain drain is defined as the next observed position being outside Austria; it is not a permanent emigration measure.

D · Related research

Part of an ongoing research programme

This report is part of an ongoing co-authored research programme on AI workforce composition and firm performance.

BMVIT — Austrian Federal Ministry for Transport, Innovation and Technology · 2016
Industrie 4.0 und ihre Auswirkungen auf die Transportwirtschaft und Logistik
Federal-ministry report on the implications of Industrie 4.0 for Austria's transport and logistics sector. Co-authored at the request of the BMVIT, this is the structural baseline against which the present report's 97-percent zero-AI-firm finding is read: the same Mittelstand readiness gap, translated from sensors and ERP into models and data pipelines a decade later.
Working paper · 2026
Build, Apply, Operate: How AI Workforce Portfolio Composition Shapes Firm Value
Using a panel of 1,927 U.S. firms (8,645 firm-years, 2019–2025) and a purpose-built taxonomy of 342 AI roles, the paper shows that firms with complementary Build and Apply capabilities — conditional on Operate capacity — capture higher market valuations. Validated by independent expert coders (Cohen's κ = 0.82).
Under review at ICIS 2026 · 2026
AI Workforce Architecture and Digital Innovation: Complementarity, Governance, and the Value of Balance
Using a sample of 3,236 U.S. firms (18,010 firm-years, 2012–2021) and a taxonomy of 369 AI roles (Invention / Application-Design / Scaling-Governance), the paper shows that balanced AI workforce architectures generate more novel and broader digital innovation than concentrated ones, and that Scaling/Governance staffing amplifies the Invention × Application/Design complementarity. Three-coder validation (Fleiss' κ = 0.84).
Working paper · 2026
Search-Build Decoupling: Performance Feedback, Frontier AI Recruiting, and Internal Redeployment after the ChatGPT Moment
S&P 500 panel of 472 firms, 30.5M individuals, 65.6M job postings (2018–2025). Documents that post-ChatGPT, performance shortfalls are associated with reduced frontier AI recruiting and elevated deep-AI departures — a decoupling absent before November 2022. Taxonomy of 315 AI roles validated by three independent domain experts.
SMS (Working paper) · 2026
AI as a Ruthless Contrarian: How LLM-Powered Devil's Advocacy Enhances Strategic Decision Making
Two pre-registered laboratory experiments (N = 194, N = 228) show that embedding an LLM as an intense devil's advocate — not as an alternative generator — produces the largest gains in strategic plan quality, while the same ruthlessness from a human source backfires. LLMs remove the social cost of dissent.

Meet the authors →