The short version
- 01
Most companies now use AI, but few can show much for it. Across 14 reports, a median of 71% of organizations use AI in some form. In the six reports that measure it, a median of 39% see any measurable business value, and in five reports about 6% see significant value.
- 02
Spending keeps rising anyway. In all six reports that ask, most companies plan to spend more on AI next year (median 83%).
- 03
What holds companies back is mostly organizational. Governance and accuracy concerns, security and privacy, skills, and integration each appear in five to seven reports, usually near the top of the list.
- 04
People feel faster than they measurably are. Self-reported productivity gains are large and nearly universal. The few controlled measurements are mixed, and executives put the firm-level effect so far at well under 1%.
- 05
Executives expect job cuts more than the data shows them, except at the entry level. A median 32% of executives expect AI to shrink headcount, yet only 4% of AI-using service firms say they have laid anyone off. Early-career employment in AI-exposed jobs is where the clearest decline appears.
Use vs. value
Use figures range from 18% to 99%, and most of that spread is about who is counted. Government surveys of all U.S. businesses (Census, Federal Reserve) find about one in five using AI. Surveys of executives at larger companies find 70% to 90%. Value figures are lower and more consistent: wherever a report asks about measurable returns rather than satisfaction, fewer than half say yes, and only a small group reports substantial, sustained value.
What holds companies back
We mapped every barrier each report names to eight shared categories. Governance, risk and accuracy concerns appear most often and rank in the top three in five of seven reports. Security and privacy rank in the top three in every report that lists them. Data quality is the top barrier in the one survey built around mid-market firms (RSM). Cost and unclear ROI come up often but usually rank low, except among CFOs (Deloitte), where it ranks first.
Felt vs. measured productivity
Self-reported gains are consistent: across 13 survey findings, between 42% and 86% of workers say AI makes them faster or more productive (median 69%). Measured results vary with the task. In a randomized trial, experienced developers were 19% slower with AI while believing they were 20% faster (METR). Engineering telemetry shows more output per person but no company-level delivery gain (Faros AI). On narrower, well-defined tasks, experiments show large gains. At the level of the whole firm, executives estimate the realized productivity effect so far at an average of 0.29% (NBER).
69%
median share of workers reporting productivity or time gains, across 13 survey findings
- Accenture: 81% · Employees who feel AI tools have increased their overall productivity at work
- Anthropic: 86% · Reported productivity gains in speed from using AI
- BCG: 42% · Save at least 8 hours (one workday) per week with AI
- CPA.com: 84% · Agree AI saves them time
- Deloitte: 66% · Benefits achieving today from AI: improve efficiency and productivity
- Glean: 75% · Say AI makes them more productive
11
measured results, from randomized trials, telemetry and labor data
- Anthropic: +80% · Estimated reduction in task completion time with Claude vs. without AI (average across tasks)
- Anthropic: +84% · Median estimated time savings per conversation
- DX: +37% · Change in median weekly throughput (PRs per engineer per week), 1.42 to 1.94 over four quarters
- Faros AI: +21% · Change in tasks completed per developer on high-AI-adoption teams (telemetry)
- Faros AI: +98% · Change in pull requests merged per developer on high-AI-adoption teams (telemetry)
- METR: -19% · Change in task completion speed when AI tools allowed (randomized; negative = slower)
- NBER: +75% · Share of the education-based performance gap closed by AI access (0.548 SD without AI to 0.139 SD with AI)
- OpenAI: 1.12× · Speedup when an expert tries GPT-5, reviews output and redoes it if needed, vs expert alone
- OpenAI: 1.18× · Cost advantage of the same GPT-5 try-review-fix workflow vs expert alone
- PwC: +34% · Labour productivity growth 2018–2025 at companies in the most AI-exposed sectors
- PwC: +163% · Labour productivity growth 2018–2025 at the top 20% most AI-exposed ('super-star') companies
Jobs
Executive surveys point toward smaller workforces: a median 32% expect AI to reduce headcount. Hard data shows less so far. Only 4% of AI-using service firms have laid off workers because of AI (New York Fed), though about 15% hired fewer than they otherwise would have. Announced job cuts that name AI are real and rising: 116,175 so far in 2026, about 22% of all announced cuts (Challenger), though announcements are plans, not confirmed layoffs. The clearest measured effect is at the entry level: employment for 22–25-year-olds in the most AI-exposed jobs is running about 19% below peers in less-exposed work (Stanford Digital Economy Lab). One counterpoint: firms that spend the most on AI grew entry-level headcount 12% over two years (Ramp).
- Accenture: +52% · Leaders who expect to hire more entry-level talent due to AI
- Atlanta Fed: -0.8% · Expected change in employment in 2026 as a result of AI
- EY: +69% · Believe AI investments will lead them to maintain current employment levels or hire new talent over the coming year
- Indeed Hiring Lab: +15% · Change in software development postings since late Feb 2025 (launch of agentic coding tools)
- McKinsey: +14% · AI contributed to an overall decline in workforce size in the past year
- New York Fed: +4% · Laid off workers due to AI in the past six months
- New York Fed: -15% · Hired fewer workers than they would have without AI
- New York Fed: +13% · Hired more workers due to AI
- PwC: +52% · Headcount growth at the most AI-exposed companies, 2018 to 2025
- PwC: +35% · Growth since 2019 in AI-exposed entry-level roles requiring senior-level skills ('seniorised' roles)
- PwC: -10% · Change since 2019 in other entry-level roles
- Ramp: +10.2% · Headcount growth over the two years after AI adoption at high-intensity adopters
- Ramp: +12% · Entry-level headcount growth over the two years after adoption at high-intensity adopters
- Schwab: +9% · AI adoption avoided the need to hire new staff
- Stanford DEL: -19% · Employment shortfall vs. where it would be had it kept pace with less-exposed peers
- Stanford DEL: -11% · Change in employment, Nov 2022 – Jun 2026
- Stanford HAI: -20% · Change in employment of software developers aged 22–25
Money
Plans to spend more are the most consistent finding in the library: 60% to 85% of companies plan to increase AI spending, depending on the survey. The dollar figures are not comparable with each other, so we list them rather than chart them: enterprise spending on generative AI reached an estimated $37 billion in 2025 (Menlo Ventures); the top 1% of AI-spending businesses paid a median $7,205 per employee per month in August, down from $7,976 in July (Ramp).
Who is asking whom
Most business AI research is published by organizations that sell AI products, AI services or software that uses AI. That does not make the data wrong, but it shapes which questions get asked and which number leads the press release. Very little of it is about companies under $100M in revenue, and fewer than a third of reports measure outcomes directly (experiments, usage data, labor or government statistics) rather than asking executives.
- Publisher sells AI products, services or AI-enabled software71
- Government, universities, nonprofits and others29
| Company size | Reports |
|---|---|
| Under $100M revenue only | 4 |
| $1B+ only | 5 |
| Mixed sizes | 26 |
| Not disclosed | 30 |
29% of reports measure outcomes directly (experiments, usage data, labor or government statistics). The rest ask people.
What it means for $10–100M companies
Don't benchmark against the big adoption numbers. The 70–90% figures come mostly from executives at larger companies. Among all U.S. businesses, the share using AI is closer to one in five.
Measure one or two processes before and after. Most companies report value without measuring it, and the research consistently shows self-reported gains running ahead of measured ones. A simple before-and-after on a real workflow puts you ahead of most of your peers.
Budget for the unglamorous work. The barriers the research agrees on are governance, security review, data cleanup, integration and training, not the choice of AI model. That work is where a $10–100M company's first dollars are best spent.
Keep hiring and developing junior people deliberately. The clearest labor-market signal is fewer early-career roles in AI-exposed work. Companies that keep training the next generation of employees will have an easier time hiring in three to five years.
What to watch
Predictions are dated and scored in a later edition, including the ones we get wrong.
Executive surveys published through January 2027 (including BCG's annual value study, the State of AI Report, and Deloitte's and PwC's January surveys) will again find fewer than half of companies reporting measurable value from AI.
Check by January 31, 2027The median share of companies planning to increase AI spending will stay above 75% in surveys published through January 2027.
Check by January 31, 2027Challenger's count of announced U.S. job cuts attributed to AI will end 2026 above 130,000 but below 25% of all announced cuts.
Check by January 15, 2027No controlled study published by January 2027 will measure a firm-wide (not task-level) productivity gain from AI above 5%.
Check by January 31, 2027
Method and data
The same question worded differently produces very different numbers. 99% of mid-market executives say their company measures the ROI of AI (RSM), while 12% of senior leaders say they consistently assess the value AI produces against what it costs (KPMG). Both are accurate; they measure different things. That is why we compare figures only within one defined measure, show medians and ranges, and never average across questions.
- We extracted 217 comparable figures from 58 of the 68 reports published through September 30, 2026. Each figure is checked against the publisher's own page or PDF; 11 unverified figures are excluded.
- Figures are grouped by a shared vocabulary (for example "uses AI in core processes" or "cites data quality as a barrier"). Only organization-level percentages are plotted together.
- This is a structured review, not a statistical pooling of the data. We show medians and ranges, never averages across different questions.
- Every dot links to the report summary, and from there to the original.