Executive summary

This week’s updates centre on the intersection of AI measurement, governance, and workforce data modernisation. Key developments include a landmark O*NET report proposing a suite of 16 AI impact indices to systematically measure how AI reshapes occupations, a bipartisan policy brief calling for enhanced AI questions in major U.S. federal surveys, and new U.S. Census Bureau data on business AI adoption from the Business Trends and Outlook Survey. A peer-reviewed study on open-source AI in public sector agencies offers practical guidance for statistical offices weighing proprietary versus open-weight models. Together, these developments signal a maturing ecosystem in which measuring AI’s effects on work is becoming as important as deploying AI tools themselves.

What is new this week

Measurement and indexing of AI impact

O*NET proposes 16 AI impact indices for occupational analysis

Published in June 2026 by HumRRO and the National Center for O*NET Development, this report reviews 19 major studies on AI’s impact on work and identifies a critical gap: most existing research relies narrowly on task-level analysis and overlooks nearly a century of established job analysis research, including contextual and adaptive performance dimensions. The authors propose that O*NET develop a suite of up to 16 AI impact indices, organised around modern models of job performance, covering AI’s ability to augment or automate knowledge, skills, task performance, and contextual behaviours.

For statistical offices, this framework is directly relevant to workforce planning and survey design. Agencies can use these indices to anticipate which occupational categories within their own organisations are most exposed to AI-driven change, and to design longitudinal surveys that track these shifts over time. The proposed indices leverage large language models and O*NET’s existing occupational taxonomy, making them scalable and regularly updatable. The report emphasises transparency and consistency, which aligns with the quality standards expected of official statistics.

Survey modernisation and data collection policy

Bipartisan push to add AI questions to major U.S. federal surveys

Published on June 17, 2026, by the Foundation for American Innovation, this policy brief responds to a bipartisan Senate letter urging the Bureau of Labor Statistics (BLS) and the Census Bureau to improve data collection on AI’s workforce impacts. The brief proposes concrete additions to three key surveys: the National Longitudinal Survey of Youth (NLSY), the Current Population Survey (CPS), and the Job Openings and Labor Turnover Survey (JOLTS).

For the CPS, the authors recommend three short monthly questions covering whether workers use generative AI for work, how often, and whether their employer provided access. An annual supplement would capture self-reported time savings and task substitution. For the NLSY, AI questions should be incorporated before the 2027 cohort begins data collection, capturing early-career AI adoption. The brief also argues against revising JOLTS in favour of expanding Enhanced Wage Records (EWRs) through state Unemployment Insurance systems, which would provide occupation-level employment and earnings data at far greater scale than any household survey.

This policy discussion is highly relevant for statistical offices worldwide. The proposed survey modules offer a practical template for any agency seeking to measure AI adoption in its national labour force surveys. The EWR approach also illustrates how administrative data can complement survey data to achieve higher precision at lower respondent burden.

U.S. Census Bureau releases BTOS data on business AI adoption

On June 18, 2026, the U.S. Census Bureau released new data products from the Business Trends and Outlook Survey (BTOS) covering AI adoption across U.S. businesses. The supplemental questions, fielded between November 2025 and February 2026, capture how AI is being adopted across industries, geographies (by state), and firm sizes, as well as the types of tasks AI supports and how it is changing work. The data show that AI adoption rates ticked down to 19.5% of businesses as of May 2026, with approximately 80% of companies not yet adopting AI.

This release is significant for statistical offices for two reasons. First, it demonstrates a replicable model for embedding AI adoption questions into existing rapid-turnaround business surveys. Second, the data themselves provide a benchmark for understanding the pace and distribution of AI diffusion across the economy, which is essential context for any agency designing AI-related survey modules or evaluating the representativeness of AI-focused samples.

Governance and strategic adoption

Open-source AI in the public sector: navigating model choice

Published in Government Information Quarterly (Volume 43, Issue 2, June 2026), this peer-reviewed study by Nicholas Robinson is the first to examine open-source AI (OSAI) adoption specifically in public sector agencies. Drawing on 31 interviews with decision-makers in Australian, Canadian, and German agencies, the study finds that technological factors — particularly fit with existing infrastructure, control over model behaviour, and hardware availability — are more influential in OSAI adoption decisions than in traditional open-source software choices. Organisational considerations such as digital sovereignty and data protection are also more prominent for AI than for conventional software.

A key finding is that AI models are more homogenous and easier to switch between than traditional software, reducing fears of vendor lock-in. However, the choice to adopt OSAI still involves long-term commitments, particularly investment in on-premises hardware and the development of internal sovereign capabilities. The study uses the Technology-Organisation-Environment (TOE) framework to structure its analysis.

For statistical offices evaluating whether to deploy open-weight models (such as Llama or Gemma) versus proprietary APIs (such as GPT or Claude), this study provides a structured evidence base. The findings suggest that data protection and sovereignty concerns are the primary drivers of OSAI adoption in government, and that agencies should assess their hardware readiness and internal AI skills before committing to a self-hosted model strategy.

OECD “Statistics in the AI Era” podcast: making data findable for AI systems

On June 22, 2026, the OECD published the second episode of its “Statistics in the AI Era” miniseries, featuring a conversation between OECD Chief Statistician Steve MacFeely and Prem Ramaswami of Google. The episode addresses a critical challenge: as AI assistants increasingly shape how people access information, official statistics face a new usability test — whether authoritative public data is structured and labelled in ways that AI systems can reliably find, interpret, and cite.

This development highlights an emerging priority for national statistical offices: ensuring that their data portals, metadata standards, and dissemination formats are compatible with AI-driven information retrieval. Offices that do not invest in machine-readable, semantically rich metadata risk having their authoritative data bypassed in favour of less reliable sources that are more easily indexed by AI assistants. Agencies should review their data catalogues against emerging standards such as schema.org Dataset markup and DCAT-AP to ensure discoverability.

Development Category Key Finding Implication for Statistical Offices
O*NET AI Impact Indices Measurement 16 proposed indices covering task, knowledge, and contextual performance Framework for workforce planning and longitudinal survey design
FAI Federal Survey Brief Survey Design Proposed AI modules for CPS, NLSY, and EWR expansion Template for measuring AI adoption in national labour force surveys
BTOS AI Adoption Data Data Release 19.5% of U.S. businesses using AI as of May 2026 Benchmark for AI diffusion; model for embedding AI questions in business surveys
Open-Source AI in Government Governance Sovereignty and hardware readiness drive OSAI adoption Evidence base for proprietary vs. open-weight model decisions
OECD Statistics in AI Era Discoverability Official statistics must be AI-readable to remain authoritative Priority: machine-readable metadata and DCAT/schema.org compliance

Implications for statistical offices

This week’s developments collectively reinforce a theme that is becoming central to the modernisation agenda of national statistical offices: the need to measure AI, not just use it. The O*NET indices and the FAI policy brief both argue that existing measurement frameworks are inadequate for capturing AI’s effects on work, and both propose concrete, scalable improvements grounded in established survey methodology. The BTOS data release from the Census Bureau provides a timely empirical anchor, showing that AI adoption remains uneven and that the majority of businesses have yet to integrate AI into their workflows.

The governance literature, represented by the open-source AI study, adds an important practical dimension: the choice of AI model is not merely a technical decision but a long-term strategic commitment with implications for sovereignty, infrastructure, and internal capacity. Finally, the OECD podcast episode points to a less-discussed but increasingly urgent challenge — ensuring that official statistics remain the authoritative source of record in an information environment increasingly mediated by AI assistants.

Next actions

  • Review the O*NET AI impact indices framework and assess its applicability to the agency’s own occupational classification system for workforce planning purposes.
  • Examine the FAI-proposed CPS and NLSY survey modules as templates for designing or updating AI adoption questions in national labour force or household surveys.
  • Analyse the BTOS AI adoption data as a benchmark for understanding the current pace of AI diffusion and to contextualise agency-level AI adoption strategies.
  • Conduct an internal review of the agency’s data portal metadata against DCAT-AP and schema.org Dataset standards to assess AI discoverability.
  • Use the Robinson (2026) TOE framework to structure an internal assessment of the agency’s readiness to adopt open-weight AI models, focusing on hardware infrastructure, data protection requirements, and internal skills.

Sources