AI workforce policy needs attributable labour data, not exposure proxies

The Indiana Business Research Center's labour-market analysis examines what available data can reveal about AI and work. The U.S. Census Bureau's Business Trends and Outlook Survey provides recurring business-reported indicators, including AI use. Together they illustrate an important evidence distinction: exposure, reported adoption and attributable labour outcomes are different measurements.

An occupation exposure score estimates how much of a role's task mix could be affected by AI. It does not observe whether an employer deployed a system, whether employees used it, whether tasks changed, or whether headcount moved because of that deployment. Surveyed AI use is closer to adoption, but still may combine experimentation with production use and cannot by itself identify effects on a particular worker.

Build an evidence ladder

The first rung is exposure: a task or occupation has characteristics that make AI technically relevant. The second is adoption: an organisation reports or logs actual use. The third is observed change: task allocation, cycle time, quality, hiring, hours or pay changes after deployment. The fourth is attribution: evidence supports the conclusion that a specified intervention contributed to that change rather than demand, restructuring, seasonality or another technology.

Each rung needs a denominator and time window. “Jobs affected” is meaningless without defining the population, observation period and type of effect. A useful organisational record identifies the workflow, tool, deployment date, eligible roles, participating units, comparison group where possible and pre-defined outcomes. It should also record concurrent reorganisations, hiring freezes and demand shocks that could explain the same result.

Use proxies for targeting, not verdicts

Exposure scores remain useful. They can identify roles for interviews, task mapping, training and risk review. Business surveys can reveal where adoption is accelerating and where support may be needed. The mistake is to turn those proxies into a count of jobs “lost to AI” or “saved by AI” without an attribution design.

Attribution does not always require a randomised trial. Staged rollouts, matched comparison units, interrupted time series and detailed before-and-after workflow measurement can improve confidence. Qualitative evidence also matters: managers and employees can identify which handoffs changed and where effort moved. The method should be proportionate to the decision. A training pilot needs less certainty than a redundancy programme or public claim about regional job loss.

Distribution must remain visible. An average cycle-time improvement can coexist with increased monitoring, reduced entry-level learning or a transfer of exception work to a smaller group. Segment outcomes by role, tenure, location and employment arrangement where lawful and appropriate. Record whether workers had access to training, whether use was mandatory and whether performance measures changed during the observation period.

A workforce evidence ledger can connect these layers without pretending they are equivalent. It should label every metric as exposure, adoption, observed change or attributed outcome; link it to source and method; state limitations; and name the decision it supports. The Skills Atlas can help define the task and capability vocabulary. Policy should move from broad proxies to intervention-specific evidence as consequences become more material. Confidence should rise before consequences do.