Work Impact Study + Augmentation Trials

In one large study, an AI assistant helped newer customer-support agents most.¹ Across the economy, young workers in AI-exposed jobs are being hired less than their peers.² Both can be true, and Washington cannot yet tell which effect is winning, or where. Congress should put better questions in federal surveys, follow workers' outcomes over time, and fund tests of AI that makes people more capable.

The Problem

About one in five American businesses now uses AI; among firms with 250 or more employees, 37% do.³ The same tool can make a beginner better at the job or remove the work through which beginners learn.

The evidence points both ways. In a study of 5,172 customer-support agents, an AI assistant raised issues resolved per hour by 15% on average, and less experienced agents improved in both speed and quality.¹ Yet payroll records show employment of 22- to 25-year-olds in AI-exposed occupations 19% below where it would be had it kept pace with less-exposed peers, with declines concentrated where AI substitutes for workers' tasks.² Yale's Budget Lab, working from the Current Population Survey, still finds no clear evidence of AI-driven disruption in the labor market.⁴

Federal statistics can't referee the dispute. Census asks AI-using firms whether AI changed their employment, and only 2% report a decrease.⁵ The Stanford data show the shift running mainly through fewer hires,² and Census's own researchers note that their survey provides no information on AI-driven changes in hiring.⁵ Washington asks whether AI cut jobs; the early warning signs are in hiring. Measurement fails in three places:

1. The questions keep moving. When Census reworded its AI question in November 2025, measured use jumped from about 10% of firms to 18%, partly because of the new wording.⁵
2. Workers drop out of view. The early warning on young workers came from a private payroll company's records; the study's authors note that federal sources "either do not collect this information or lack statistical precision."²
3. Good practice stays local. Nearly two-thirds of AI-using businesses report no organizational changes to put it to work, and only about 15% report training their staff.⁵

Why legislation: Census has already shown it can ask about AI.⁵ The White House's 2025 AI Action Plan calls for a Labor Department hub to lead "a sustained Federal effort" on AI's labor-market impact,⁶ and senators from both parties have proposed writing that hub and new survey questions into law.⁷ What the agencies lack is a durable mandate, dedicated funding, and a program that tests what works; no statute provides them today. A country asking its workers to adapt owes them sturdy evidence about what they are adapting to.

The Solution

A four-step staircase: each step stands alone, and each step up adds evidence and public commitment. Scope: employees, independent workers, and businesses of every size, led by the Labor Department with Census, BLS, the Bureau of Economic Analysis, and NSF; a companion proposal for a federal lab covers AI's harms to health and safety, while this program covers work, pay, and productivity.

Step 1 — Ask the right questions. Add consistent, task-level AI questions to existing business and household surveys: what the AI does, who uses it, when it arrived, whether it assists or replaces a task, and what happened to hiring. Publish definitions and bridge estimates whenever wording changes. A bipartisan Senate bill would put AI questions into five federal surveys, including the Current Population Survey.⁷

Step 2 — Follow the workers. Track pay, hours, job changes, training, autonomy, errors, and job quality alongside output and costs, linking records inside secure statistical enclaves under existing confidentiality law. Publish only aggregates, with stated uncertainty. Linked records are how a quiet hiring freeze becomes visible.

Step 3 — Test what works. Offer time-limited grants to businesses, community colleges, and worker partnerships that test tools meant to make people better at their jobs. Require credible comparison groups, worker consent for experimental data, and publication of unfavorable results. Such tests are practical: the customer-support study tracked agents through the tool's staggered rollout.¹

Step 4 — Spread what works. Publish practical guides on training costs, failure conditions, and which workers benefited, plus annual regional and occupational reports so schools and workforce programs can respond. NIST's Manufacturing Extension Partnership already advises small and midsize manufacturers through centers across the country;⁸ AI at work needs the same kind of help.

Where to start: Step 1 is the floor; Census is already partway there. Step 2 is the heart of the proposal.

Administration and enforcement: Labor issues a common research plan within 12 months and first results within 24, under a five-year authorization with appropriated funds. Statistical confidentiality rules govern all linked records; grant agreements allow audits, recovery for misuse, and suspension for concealed results. The program may not collect workers' private messages.

Risks and Mitigations

• Correlation isn't causation: Firms that adopt AI differ from firms that don't, and the Stanford authors call their findings "early, descriptive indicators" rather than causal estimates.² Trials and long panels narrow the gap, and findings should inform policy without triggering it automatically, but uncertainty will remain.
• Vendor influence: A public grant should not buy a favorable case study. Require independent evaluators, preregistered plans, and the right to publish; some bias in who applies remains.
• Burden and surveillance: Reuse existing surveys, sample small firms, and bar any use of research data in decisions about individual workers. These safeguards cost precision for small occupations.

Similar Bills

Fit measures similarity to this proposal's mechanisms: High = direct precedent; Partial = useful component with material differences; Related = adjacent approach.

Federal

State

What this adds: Pending bipartisan bills would add AI questions to federal surveys and require employers to report AI-related layoffs. This proposal adds what neither provides: records linked over time that can see hiring as well as layoffs; funded, independently evaluated trials of AI that makes workers more capable; and a standing channel to spread what works.