PwC's Jobs Barometer reads a billion job ads, not opinions

The setup

PwC's AI Jobs Barometer takes a different evidentiary approach than an opinion survey: its 2026 edition examines more than one billion online job advertisements from six continents, alongside firm-level financial records, to describe how wages, hiring and output differ between more- and less-AI-exposed work. The report's own methodology appendix, retrieved 16 September 2026, lays out how it scores occupational exposure before touching wages or productivity at all.

What the documents show

The appendix describes a refreshed version of economist Edward Felten's AI Occupational Exposure Index, built from O*NET ability profiles scored against ten AI applications' capability to perform each ability, then aggregated into an occupation-level exposure score between 0 and 1. For productivity, the report pairs that exposure score with Orbis firm financial data, calculating turnover per employee and wage per employee for firms in the top and bottom exposure quartiles and comparing 2018 against 2024/25, after filtering out firms with missing, negative or extreme values.

The friction

The appendix itself flags the interpretive limit plainly, stating a higher exposure score 'does not imply job loss or automation,' only that a sector's work sits more heavily in occupations where AI capabilities are relevant. The productivity comparison is a correlation between an exposure quartile and a financial outcome across surviving firms in a commercial database, filtered by size and revenue thresholds the appendix documents in detail; it is not a controlled study of AI's causal effect on any one firm's output.

What changed in the work

What this approach adds over a sentiment survey is a labor-market-data method: instead of asking employers what they believe about AI, it reads what employers actually posted and what surviving firms actually reported financially. That is a meaningfully different evidence type, though it inherits its own limits, including survivorship bias in a database that only reflects firms with financial data available in both 2018 and 2024/25.

• Does the cited figure describe an occupation's AI exposure score, or an observed wage or employment outcome?
• Is the comparison correlational, matching exposure quartiles to outcomes, or does the source claim to isolate AI's causal effect?
• Which years and geographies does the underlying job-ad and financial data actually cover?

Read against a survey like McKinsey's or Gallup's, the Jobs Barometer's job-ad-and-financial-data method offers a different kind of evidence about the same broad question, and the two approaches disagreeing on a number is itself informative about how unsettled the underlying measurement problem still is.