The Labor Department Wants Big Tech’s AI Data—but the Findings Stay Public

An Axios report published August 26, 2026 says acting Labor Secretary Keith Sonderling disclosed data-sharing agreements intended to help the U.S. Department of Labor track AI’s effects on jobs and hiring, named OpenAI, Google, Meta and Amazon among the participating companies, and said the resulting findings would be public. The private inputs are intended to supplement conventional government statistics rather than replace them.

The agreements build on an initiative that was already underway: Associated Press coverage of Sonderling’s July 16 confirmation hearing says the department was collecting information about AI use from companies and unions for Bureau of Labor Statistics review. The disclosed public commitment concerns findings derived from the additional information, not publication of the companies’ underlying records.

Private signals could narrow the information lag

The potential gain is timeliness. Federal labor surveys follow scheduled collection, processing and release cycles, while technology providers may be able to observe changes in customer activity sooner. If the records are delivered consistently, they could alert government analysts to shifts in business adoption that merit comparison with employment, vacancies, hours, wages or occupational data.

The public disclosure does not identify the fields each company will provide. Possible measures could include business subscriptions, active accounts, use within particular functions or employer expectations, but those are distinct concepts and should not be treated as interchangeable without documentation.

Nor does product activity directly measure a labor-market outcome. Increased use of an AI service could reflect automation, assistance for existing employees, experimentation or expansion at a company that is also hiring. Establishing whether jobs changed because of AI would require linking the private signals to independent employment measures and accounting for other economic forces.

Coverage presents another constraint. OpenAI, Google, Meta and Amazon have different products, customers and relationships with employers, so each company observes a different slice of the economy. Combining their records without preserving those distinctions could produce a fast indicator that appears comprehensive but systematically misses businesses outside the participating platforms.

Census data offer an established adoption benchmark

The government already operates a public survey against which the private signals can be evaluated. A Census Bureau analysis published May 26, 2026 describes the Business Trends and Outlook Survey as biweekly and nationally representative, and shows overall business AI use ranging from 17% to 20% in data collected from December 14, 2025, through May 3, 2026; expected use over the following six months ranged from 20% to 23%.

BTOS asks businesses about recent AI use and near-term expectations under a common questionnaire. Its core wording covers AI used in any business function, while supplemental questions examine functions such as human resources, customer service, marketing, finance, information technology, and research and development.

That survey design supplies elements that company records do not automatically provide: a defined target population, consistent questions, survey weights and a national denominator. Private records could contribute behavioral detail or earlier movement; BTOS can help determine whether that movement extends beyond the customers visible to a particular provider.

The two sources therefore answer different questions. A survey estimates how prevalent reported adoption is across businesses, while provider records may reveal activity within a commercial ecosystem. Agreement between them would strengthen an interpretation, but divergence would require investigation of coverage, definitions and timing before it was treated as an economic signal.

Five methodological tests will determine credibility

Publishing findings is an important transparency commitment, but public availability alone does not make an indicator statistically reliable. Any recurring product based on the agreements would need documentation that allows researchers to understand how the measure was constructed and where it may be biased.

• Sampling: The department would need to identify which employers, workers, industries and regions appear in each provider’s records, which groups are absent and whether weighting adjusts for those gaps.
• Definitions: Holding an account, paying for access, actively using a tool, deploying it in a business function and changing how work is performed represent different stages of adoption. The published methodology should state exactly which one is measured.
• Conflicts and validation: Documentation should explain who chooses the submitted fields, how government statisticians test company-supplied measures and how product changes or altered internal measurement rules affect continuity.
• Release cadence: Users need to know how often results will appear, whether initial values are preliminary and how revisions or breaks in a series will be marked.
• Reproducibility: Commercial confidentiality may prevent release of raw records, but the department can still document transformations, weights, missing-data treatment, uncertainty and validation procedures.

Harmonization cuts across all five tests. Similar-looking fields from different providers may describe different customer populations or behavior. A combined estimate would have little stable meaning if those differences were concealed behind a single number.

The promised public product remains undefined

The confirmed development is narrower than the launch of a new federal statistical series. The Labor Department has disclosed data-sharing arrangements, identified several participants and committed to public findings, but no common data definition, publication schedule or recurring indicator has yet been publicly specified.

The next consequential disclosure will be methodological: what each company supplies, how the department reconciles the inputs, how the measures are validated against established surveys and what documentation accompanies the results. Private data may shorten the delay between changes in AI use and their appearance in federal analysis, but the indicators will earn trust only if their speed is matched by transparent coverage, definitions, uncertainty and revision practices.

Also read:

• Encore Leads the Public Debt-Buyer Comparison—but Private Rivals Stay Opaque
• Companies Are Automating Decisions, but Data and Controls Set the Pace
• Big Data Powers More Finance Automation—But Lineage Still Blocks Trust