Building an AI-Ready America: Adopting AI at Work

US House Committee on Education and Workforce

Subcommittee on Health, Employment, Labor, and Pensions

Hearing on “Building an AI-Ready America: Adopting AI at Work”

Good morning, Subcommittee Chairman Allen, Subcommittee Ranking Member DeSaulnier, and members of the Subcommittee on Health, Employment, Labor, and Pensions. It is an honor to testify before you.

My name is Revana Sharfuddin, and I am a labor economist and research fellow in labor policy at the Mercatus Center at George Mason University. I study how new technologies diffuse through labor markets and how policy can help workers share in the gains from innovation.

Today, my testimony focuses on a simple proposition: The United States needs stronger federal measurement of AI use at work—by both firms and workers—before we can evaluate how AI is affecting employment, wages, and job quality. Without better measurement, policymakers will be asked to make high-stakes decisions using fragments: small convenience samples, inconsistent survey wording, and headline-grabbing anecdotes.

I also want to begin by acknowledging the human side of this discussion. Many workers are hearing constant predictions, often without context or evidence, that AI will “replace” them. That anxiety is not irrational: It reflects real uncertainty about how employers will implement the technology and whether workers will have a voice in that process. Recent nationally representative survey evidence finds that American workers are more worried than hopeful about AI’s future use at work.¹

My testimony makes three points:

1. What we know: Credible US data show that AI adoption and use are rising, but remain unevenly concentrated in certain sectors, firm sizes, and tasks.
2. What we do not know: Current federal data cannot yet distinguish exposure to AI from labor market impact, nor can it reliably measure task redesign within occupations, which is where many near-term changes are likely to occur.
3. What Congress can do: With modest investments and careful design, federal statistical agencies can build a measurement system that is timely, comparable over time, and worker centered.

Why This Issue Matters Now

AI is not new, but the pace of diffusion has accelerated. Since the release of widely accessible generative AI tools in late 2022, both employers and workers have experimented with using AI for writing, coding, customer support, and internal knowledge management—often adopting and testing AI tools faster than standard federal surveys were designed to detect. At the same time, the policy conversation has shifted from whether AI is “coming” to whether the United States will adopt AI in ways that raise productivity while protecting worker opportunity.

In this environment, measurement is not an academic luxury; it is a prerequisite for accountable policy. When Congress debates training investments, unemployment insurance modernization, or labor standards for AI-mediated work, it needs to know: Where is AI actually being used? For what tasks? By which workers? And with what observable labor market correlates?

What We Know from Credible US Data About AI Diffusion

AI adoption by firms is real, but still concentrated. Using the US Census Bureau’s Annual Business Survey (ABS), recent research finds that less than 6 percent of US firms used AI-related technologies in 2018, but adoption rises sharply with firm size and is more common in manufacturing and information.² Because large firms employ many workers, the employment-weighted adoption rate is much higher than the firm-level rate; the same research estimates that just over 18 percent of workers were employed at AI-using firms in 2018.³

High-frequency federal data show rising reported AI use—along with measurement sensitivity. The Census Bureau’s Business Trends and Outlook Survey (BTOS) provides a near-real-time view of business conditions and technology use. Beginning with data released in December 2025—after cognitive testing and expert feedback—the BTOS broadened its core AI questions from AI used “in producing goods or services” to AI used “in any of its business functions.” Because the wording change produced a level shift, the Census Bureau created a new time series.⁴

Under this broader, business-function framing, in November and December 2025 about 17–18 percent of firms reported using AI in the prior two weeks, and roughly one in five expected to be using AI within the next six months.⁵

Sectoral patterns reinforce this uneven diffusion. In late 2025, adoption was highest in information (14–18 percent) and professional services (9–13 percent), compared with single-digit shares in accommodation and food services and transportation.⁶

Smaller firms are adopting AI more slowly, though the gap is narrowing. By mid-2025, less than 9 percent of small businesses reported using AI, compared with low double-digit rates among large firms.⁷

Worker use and perceptions do not mirror firm adoption one-for-one. On the worker side, nationally representative surveys show both rising AI use and substantial anxiety. In early 2025, the Pew Research Center found that only a minority of workers reported using AI on the job, even as a majority expressed concern about how employers would deploy AI.⁸ By fall 2025, Pew estimated that about one in five US workers was using AI in some capacity at work, with use concentrated in writing, research, and analytical tasks.⁹

Exposure is not the same as impact. Much of the existing research suggests that AI is more likely to reshape tasks within occupations than to eliminate entire jobs in the near term. Task-based analyses show that while many occupations contain some tasks that AI can assist with, relatively few occupations are fully automatable with current technology.¹⁰ This distinction matters because focusing on job counts alone can obscure substantial reallocations of time, skills, and responsibility within existing roles.

What We Do Not Know—and Why These Gaps Matter for Policy

Despite recent progress, federal labor market statistics were largely designed to count jobs rather than to measure how work is done. As a result, several critical gaps remain.

First, task-level change is largely invisible in household surveys. The Current Population Survey (CPS) can tell us how many Americans are employed and in what occupations, but it does not systematically ask whether AI has changed the tasks workers perform or the tools they use on the job.¹¹

Second, firm surveys typically ask whether AI is used, not how it affects workers. Surveys such as the ABS and BTOS capture information on adoption and expectations but do not directly measure AI links to redesigning tasks, changing skill demand, or shifting hiring and wages within firms.

Third, measurement is sensitive to definitions and question wording. The BTOS experience in late 2025 illustrates how small changes in survey language can produce large level shifts, thereby complicating trend analysis and policy interpretation.¹²

These gaps matter because policy debates often move faster than the data. Without task-level and worker-level evidence, policymakers risk over-attributing short-term labor market fluctuations to AI—or missing early warning signs of concentrated disruption.

Federal Measurement Efforts Since 2015: Progress and Limits

Congress and federal agencies have recognized these measurement challenges for nearly a decade.

In 2019, responding to direction from the Senate Appropriations Committee, the Bureau of Labor Statistics (BLS) submitted a report outlining a strategy to measure the effects of automation, digitization, and AI on the US workforce.¹³ The Government Accountability Office reinforced this need, concluding that improved data were necessary to assess and plan for the workforce impacts of advanced technologies.¹⁴

BLS subsequently commissioned a comprehensive review of key measurement constructs and data gaps, emphasizing the importance of capturing task change and technology adoption using existing surveys where possible.¹⁵

The Census Bureau has made the most visible progress on firm-side measurement through the ABS and the BTOS. However, worker-side measurement and the systematic linking of adoption data to labor outcomes remain limited.

Interpreting Today’s Labor Market: Why Caution Is Warranted

It is tempting to attribute any signs of labor market softening to AI. The evidence does not support that conclusion—at least not yet.

As of late 2025, the US labor market continued to exhibit relatively low unemployment by historical standards, with the unemployment rate fluctuating narrowly around 4 percent through much of the year.¹⁶ This does not rule out AI-driven disruption, but it does suggest that aggregate employment statistics are a blunt diagnostic tool for a technology that primarily operates at the task level.

More informative signals would include sustained within-occupation task redesign, differential hiring patterns within AI-adopting firms, and changes in training demand—none of which are currently well-captured in federal data.

Actionable Steps Congress Can Take

Congress does not need to create an entirely new statistical system to improve AI measurement. Three targeted steps would materially strengthen the evidence base:

1. Direct BLS to pilot an AI-at-work supplement to the CPS. A short, periodic supplement could ask workers whether they use AI tools, which tasks are affected, and whether technology has eliminated or expanded parts of their job. This approach would mirror earlier CPS supplements on computer and internet use and could be fielded at modest cost.
2. Make AI measurement a permanent feature of Census Bureau business surveys—and link it to worker outcomes. Congress should support continued AI modules in the ABS and BTOS, while enabling secure linkage to the Longitudinal Employer-Household Dynamics (LEHD) data to examine hiring, separations, and wages at AI-adopting firms.
3. Require coordinated annual reporting across agencies. BLS, Census Bureau, National Science Foundation, and National Institute of Standards and Technology each hold part of the picture. A joint annual AI-and-workforce indicators report would provide Congress with a consistent, transparent snapshot of adoption, use, and emerging labor market correlations.

These steps are feasible within existing infrastructure, impose limited respondent burden, and would dramatically improve policymakers’ ability to separate hype from harm.

Conclusion

AI is already changing how work is done in the United States—but in uneven, task-specific ways that our current data struggle to capture. Workers sense this uncertainty acutely. The appropriate response is neither complacency nor panic, but measurement-first policymaking.

By investing in better data now, Congress can ensure that debates over training, labor standards, and social insurance are grounded in evidence—and that AI adoption at work proceeds in a way that is both innovative and worker centered.

Thank you again for the opportunity to testify. I look forward to your questions.