AI's impact on the workforce: We need data to understand it

Director and Executive Editor, Indiana Business Research Center, Indiana University Kelley School of Business

On July 29, 2026, I had the honor of testifying before the U.S. Senate Subcommittee on Employment and Workplace Safety, a subcommittee of the Committee on Health, Education, Labor and Pensions. The hearing examined “The Impact of AI on the Workforce” and, specifically, the AI Workforce PREPARE Act (S. 3339).

I was invited as a subject-matter expert, not as an advocate for a predetermined conclusion about AI's effects. My role was to describe what the evidence shows so far and assess whether the data systems on which the bill depends are ready to do what it asks of them.

What the evidence shows so far

AI adoption has grown rapidly from experimentation to routine use. In the latest available national data, collected between August 10 and August 23, 2026, more than 22% of U.S. firms reported using AI in some business function during the preceding two weeks, according to the Census Bureau's Business Trends and Outlook Survey (BTOS).¹ Harvard's Project on Workforce finds that 61.8% of working-age adults used generative AI in some form as of May 2026.²

Adoption is far from even across the country. The state comparisons in Figure 1 use an earlier BTOS collection period, June 29 through July 12, 2026, because those are the data that were available at the time of my testimony. Current use among firms ranged from 11.1% in Arkansas to 28.8% in Massachusetts, with Indiana at 20.9% — close to the national rate for that collection period and near the middle of the state distribution.³ Several high-adoption states have large technology or professional-services sectors, while several lower-adoption states have economies more concentrated in agriculture or traditional manufacturing. These descriptive patterns do not, by themselves, establish why adoption differs among states.

Figure 1: Share of firms currently using AI in some business function by state, June 2026

What's happening within occupations tells an even more interesting story than the adoption numbers alone. Research from the Indiana Department of Workforce Development, published in InContext earlier this year, compared Indiana occupations with high and low exposure to generative AI. Job postings in AI-exposed occupations shrank 41% from 2022 to early 2026, compared with 35% in unexposed occupations. Advertised wages in exposed occupations grew faster — 16% versus 11% — and the historical unemployment-rate advantage held by AI-exposed workers nearly disappeared.⁴ These findings are suggestive, but they do not establish that AI caused the changes.

Federal Reserve research also points to a mixed and still-emerging picture. A survey of corporate executives by researchers affiliated with the Federal Reserve Banks of Atlanta and Richmond found limited near-term aggregate job loss alongside changes in the composition of firms' workforces.⁵ Dallas Fed research found stronger nominal wage growth in more AI-exposed industries, although the effects varied substantially by occupation and by the value placed on workers' experience.⁶ Taken together, these early findings look less like uniform job elimination and more like a labor market adjusting unevenly, with some occupations losing demand while others gain pay.

I also cautioned the Subcommittee against reading too much into layoffs announced by a handful of large technology firms. A 2025 National Federation of Independent Business survey found that 98% of the 119 small employers in its sample that were using AI reported no change in employee counts.⁷ The small sample and the survey's timing warrant caution, but the result underscores that the experience of the largest technology companies should not be treated as representative of all employers.

Why data infrastructure is the real question

The PREPARE Act is, at its core, a data bill. Its premise is that policymakers, training providers and workers lack the data and forecasts needed to anticipate AI-driven disruption. I agree with that premise. I focused the Subcommittee's attention on a second-order question the bill does not fully resolve: Are the federal and state data systems on which it relies capable of producing timely, reliable and AI-attributable labor market signals?

My answer was mixed. Some resources, particularly state administrative wage records, are more capable than is generally understood in Washington but remain underused. Other resources, such as occupational classifications and projections that update on multi-year cycles, are not yet well suited to a policy question moving this quickly. I offered section-by-section recommendations concerning the bill's research hub, its WARN Act disclosure standard, Bureau of Labor Statistics forecasting capacity and researcher access to federal microdata. For InContext readers, however, I want to highlight the recommendation closest to home.

The case for enhanced wage records

Every quarter, employers covered by Indiana's unemployment insurance system report wage data to the Indiana Department of Workforce Development. Indiana's 2025 state budget legislation — House Enrolled Act 1001, enacted as Public Law 213 — requires the department to collect additional information from covered employers, including each employee's Standard Occupational Classification code or primary job title, employment classification and hourly rate of pay. The law also directs the department to conduct employer outreach aimed at increasing accurate reporting.⁸

In my testimony, I argued that these administrative records — especially once enhanced and linked with systems connecting education and workforce outcomes — can provide one of the most comprehensive, frequently updated views of covered employment and earnings at the state and local levels. They do not cover every kind of work, and reporting practices vary, but they offer geographic and industry detail that many survey products cannot provide.

My recommendation to the Subcommittee was to direct new federal data and research provisions to connect with this existing state infrastructure and to fund the technical assistance states need to make their records more comparable. That approach would build on established reporting systems instead of creating a separate federal collection from scratch.

What this looks like on the ground in Indiana

To show the Subcommittee what county-level analysis can make possible, I included supplementary findings from an IBRC working paper by Riley Hudelson-Zipper. The paper estimates generative AI exposure across all 92 Indiana counties by combining Lightcast employment data, county-level occupation-industry staffing patterns and Felten et al. exposure measures for image-generation and language-modeling tools.⁹

Figure 2: Intensity of AI-exposed jobs across Indiana counties

The county scores cluster within a relatively narrow band, and higher-exposure counties are dispersed across the state rather than confined to a single region. Martin County has the state's highest estimated exposure, driven in part by an unusually large employment concentration in engineering services. Ohio County also receives a relatively high score despite its large hospitality sector, a result that warrants closer examination of its occupational staffing patterns.

The estimates demonstrate the value of county-level analysis, but they require careful interpretation. Exposure indicates where AI could affect job tasks; it does not show whether employers have adopted AI, whether workers have been displaced or when any effects might occur. The scores should therefore be combined with employer-reported adoption, employment trends and local knowledge before they are used to target training investments. Used in that way, county-level data could support skills inventories and more focused workforce planning without treating model estimates as observed outcomes.

Where this leaves us

I did not come to the Subcommittee arguing for a particular verdict on AI's impact. I came to describe what the evidence shows, where Indiana's data infrastructure is already ahead of what Washington may assume and where targeted investment — not another parallel system — will determine whether this bill can accomplish its goals.

Enhanced wage records are a technical change with an outsized potential payoff. They can help move us from knowing generally that Indiana's workforce is changing toward understanding where change is occurring, how quickly it is happening and which workers and employers may need support.

My full written testimony and a video of the July 29, 2026, hearing are available on the Senate Committee on Health, Education, Labor and Pensions website.¹⁰