An AI subscription is an input. Whether it becomes a productivity gain depends on what happens to the work around it: which tasks change, who checks the output, what data the tool touches and how the business knows the result is better. New Census Bureau research suggests that for most U.S. firms using AI, that second step has not happened yet.
That finding does not prove that redesigning work first is the best AI strategy for every employer. No study we have reviewed establishes a universal return on either approach. But it does show where most firms currently stand, and it identifies the questions an employer should be able to answer before a purchase rather than after it.
What AI-using firms report
The evidence comes from an April 2026 Census Center for Economic Studies working paper, “The Microstructure of AI Diffusion,” which analyzes a detailed AI supplement to the Bureau’s Business Trends and Outlook Survey with a November 2025 to January 2026 reference period. In that period, 18% of firms used AI in a business function, or 32% on an employment-weighted basis.
Among AI-using businesses, the paper reports:
- 64% reported no organizational adjustments to use AI. The authors say this suggests “a reliance on ‘off-the-shelf’ AI/GenAI tools or a significant lag in organizational restructuring.”
- The most common adjustments, training staff and developing new workflows, each applied to about 15% of AI-using firms.
- Deeper changes, such as new data-management and storage practices or complementary capital investments, occurred in roughly 7% to 8% of firms. Hiring staff trained in AI was the least common adjustment.
- 57% of AI-using firms integrate AI in three or fewer business functions, most often sales and marketing (52%), strategy and business development (45%) and IT (41%).
- Among firms reporting any effect of AI on worker tasks, 66% said AI only augmented those tasks rather than replacing them.
In other words, the typical AI-using firm has added a tool to a few functions without changing much else. That may be sensible at an early stage. It also means the aggregate adoption numbers say little about how deeply AI has been built into how work gets done.
What the evidence links, and what it does not
The same paper finds that broader integration is associated with better reported outcomes. The breadth of AI use across business functions, worker-task use and operational investments all have “positive and statistically significant associations” with the likelihood of above-average current and future performance and with current sales increases. Functional breadth shows the strongest association.
Three cautions apply. First, the authors describe these as regression-style summaries “with no causal interpretation.” Firms that were already performing well may simply be the ones able to integrate AI more broadly. Second, the associations are not all benign for workers: functional breadth and operational investment are also positively associated with the likelihood of current and future employment declines, while worker-task use shows no significant relationship with those outcomes. Third, the authors suggest that structural investments are likely to drive long-term productivity gains “even if they depress measured performance in the short run,” a pattern economists call a productivity J-curve. They flag causal identification as future work.
Other evidence points the same way without settling the question. A September 21, 2026 Stanford Digital Economy Lab working paper by Bharat Chandar and Bouke Klein Teeselink, which infers AI adoption from job advertisements involving generative-AI use across 41 countries, finds shifts in workforce composition at foreign affiliates of adopting firms and a model suggesting modest productivity gains. The Bureau of Labor Statistics reported that U.S. nonfarm business labor productivity rose at a 1.4% annualized rate in the second quarter of 2026. That is an economy-wide measure, and BLS does not attribute it to AI or any other technology.
A sequence employers can test
None of this yields a guaranteed return. It does suggest an order of operations that makes an AI investment measurable, whatever the outcome:
- Start with the task, not the tool. Document how long the work currently takes, where errors occur and which errors are costly.
- Identify the data. Decide what information the tool will see and what must stay out. Privacy and security concerns were cited by 20% of firms in the Census supplement that were not planning to use AI in the near term.
- Decide where human review stays. Specify which outputs need checking and by whom before the rollout, not after a mistake.
- Measure end to end. A faster first draft is not a gain if review time or error rates rise. Compare total process time and quality against the baseline.
- Track people effects separately. Record changes in training needs, staffing mix and hiring alongside output, so the organization can see whether AI is changing the composition of work.
For workers, the Census findings suggest that in most firms, AI is still being used to assist existing tasks. Where employers do redesign workflows, the job changes more than the tool does, and training becomes the relevant question. Only about 15% of AI-using firms reported training staff as an adjustment.
What to watch next
The Census Bureau publishes core AI-use estimates every two weeks and is planning questions on AI’s effects on hiring for its next AI supplement, pending approval from the Office of Management and Budget. Firm-level evidence that follows the same firms over time, rather than comparing adopters with non-adopters at one point, would be needed to test whether redesign-first adoption actually outperforms tool-first adoption. Until then, the defensible claim is narrower: most firms using AI have not yet reorganized work around it, and those that measure the work before and after will be better placed to judge whether it paid off.
Sources
- U.S. Census Bureau, Center for Economic Studies, “The Microstructure of AI Diffusion: Evidence from Firms, Business Functions, and Worker Tasks (CES-WP-26-25; Bonney, Breaux, Dinlersoz, Foster, Haltiwanger, Pande),” 2026-04. https://www.census.gov/library/working-papers/2026/adrm/CES-WP-26-25.html
- Stanford Digital Economy Lab, “How Does AI Change Labor Demand? Evidence from 41 Countries (Bharat Chandar, Bouke Klein Teeselink), working paper,” 2026-09-21. https://digitaleconomy.stanford.edu/publication/how-does-ai-change-labor-demand/
- U.S. Bureau of Labor Statistics, “Productivity and Costs, Second Quarter 2026, Revised,” 2026-09-03. https://www.bls.gov/news.release/prod2.nr0.htm
Featured photo: A shared coworking space at Common Desk in Dallas. Representative image. Photo: Thecommondesk, via Wikimedia Commons (CC BY-SA 4.0). Resized from the original.
.jpg){kind=link}