Source: raw/ai_index_report_2026.pdf (Chapter 4 “Economy,” pp. 171-230; Appendix “Chapter 4: Economy” methodology notes, pp. 396-404); McKinsey & Company, “The State of AI in 2025: Agents, Innovation, and Transformation” (verified directly, see ai-research/mckinsey-state-of-ai-2025-november-*.md)
Deep-dive companion to Stanford HAI AI Index Report 2026, which covers only the chapter-map summary. This article extracts the full Chapter 4 content: investment and infrastructure, corporate AI adoption methodology, the consumer-value (“what is generative AI worth?”) study, and the full labor-market/productivity research bibliography — the material needed before citing any of HAI’s economic numbers in a high-stakes proposal.
Key Takeaways
- Global corporate AI investment more than doubled in 2025, reaching 344.66B (+127.5%), with generative AI alone at $170.9B (+200%+, nearly half of all private AI funding).
- US private AI investment (12.41B) in 2025 — though Quid’s private-investment figures exclude Chinese government guidance funds, which deployed an estimated $184B into AI companies between 2000 and 2023, so the true US-China gap is narrower than the headline ratio suggests.
- Frontier AI company revenue is scaling fast: OpenAI ~19B annualized as of early 2026 (log-scale trajectory, both companies reaching 1B crossings). Compute spend is scaling just as fast — OpenAI’s annual compute spend hit 5.8B in 2024); Anthropic’s hit 2.5B).
- “88% organizational AI adoption” and “single-digit AI agent deployment” are the same McKinsey survey, different bars — see Corporate AI Adoption section below. Conflating them overstates how far agentic AI has actually diffused.
- The $172B/yr US consumer-surplus figure is a willingness-to-accept (WTA) choice-experiment estimate, not a revenue or revealed-preference figure — see the dedicated section below.
- AI’s labor-market effects are concentrated in the youngest, most AI-exposed workers: US software-developer employment ages 22-25 fell ~20% from its 2022 peak by September 2025, even as employment for older developers kept growing. The AI-exposed/least-exposed employment gap (16 points, controlling for firm effects) began widening in mid-2024.
- Productivity research is genuinely mixed, not uniformly positive — see the full study table below, which includes a widely cited finding (METR/Becker et al. 2025) that experienced open-source developers were measurably slower with AI assistance despite believing it helped them.
Investment and Infrastructure
Corporate investment (Quid data, all four categories — M&A, minority stakes, private investment, public offerings): 344.66B (+127.5%); M&A grew 132.6% YoY.
Private investment by geography (2025): United States 12.41B, 48.5x the UK’s 757.3B vs. China’s 912B across industries and ~$184B into AI companies specifically between 2000 and 2023 (citing Beraja et al. 2024; Luong et al. 2021). The report flags this directly: “comparisons based solely on private investment alone likely understate how much capital China is directing toward AI.”
Company economics (Epoch AI data on publicly disclosed figures, directional not precise): annualized revenue — OpenAI 19B, xAI 400M, Z.ai 16.30B in 2025 (up from 0.42B in 2022); Anthropic 2.50B in 2024).
Capital expenditures: hyperscaler capex has more than doubled since ChatGPT’s release (Citi Research data). Google led 2025 capex at >240B (2024) toward an estimated ~$430B (2026 est.).
Corporate AI Adoption — the 88% Figure, Precisely
This directly resolves the drain question on the McKinsey adoption-vs-agent-deployment gap.
McKinsey & Company’s 2025 “State of AI” survey (see Methodology section below) asked organizations about AI use at three levels of specificity, and each level produces a different headline percentage:
| Metric | 2025 value | 2024 value | What it measures |
|---|---|---|---|
| Org uses AI in ≥1 business function (“88% adoption”) | 88% | 78% | Broadest bar — any AI tool, anywhere in the org |
| Org regularly uses generative AI in ≥1 business function | 79% | 71% | Narrower — specifically GenAI, “regularly” |
| Org has AI agents in scaled use, by function | single digits across nearly all functions (ceiling ~24%, software engineering in tech sector) | n/a (agentic AI is a new 2025 skill cluster) | Narrowest — agent deployment specifically, at scaled maturity |
The 88% figure is therefore the least strict of the three: it counts an organization as “adopted” if AI is used anywhere, at any maturity stage (experimenting counts the same as fully scaled for this top-line number). The report’s own Deployment Stages breakdown (Figure 4.3.6) shows why the headline overstates depth: even among the largest organizations ($5B+ revenue), only 10% report AI as “fully scaled” — most are still “piloting” (31%) or “scaling” (39%).
The 88% bar has also loosened over time, which matters for any year-over-year trend claim built on it. McKinsey’s own footnote (from the prior, March 2025 edition of the same survey) documents the shifting definition: in 2017 “adoption” meant AI was “a core part of the business or at scale”; in 2018-19 it meant “embedding at least one AI capability”; since 2020 it has meant simply “adopted AI in at least one function” — a much lower bar than the pre-2020 definition. The trend line in Figure 4.3.1 (20% in 2017 → 88% in 2025) is therefore not a clean apples-to-apples series; a meaningful share of the rise reflects a loosened definition, not only real adoption growth.
For AI agents specifically, McKinsey’s own report (Exhibit 2, cited directly rather than through HAI’s paraphrase) states: “In any given business function, no more than 10 percent of respondents say their organizations are scaling AI agents.” Broader context from the same exhibit: 23% report scaling an agentic system somewhere in the enterprise (not tied to one specific function), 39% are experimenting, and 62% are at least experimenting. Stanford HAI’s own restatement of this in-chapter is “AI agent deployment was in the single digits across nearly all business functions,” and the chapter text adds: “Across most business functions, a majority of respondents reported no agent use at all… Even in functions with the most activity, including IT and knowledge management, about two-thirds or more of respondents reported no use.”
The one place the data legitimately exceeds “single digits” is a narrower cut than “any given function”: Figure 4.3.7 breaks scaled agent use out by industry × function specifically (not function alone, aggregated across all industries), and within that finer cut, the technology sector’s own software-engineering function reaches 24% scaled use, with IT at 22% and service operations at 21% — still the ceiling of the entire dataset, not the average, and consistent with McKinsey’s “no more than 10% in any given function” statement once you account for the fact that averaging across all industries pulls the tech-sector-specific peak back down.
Bottom line for citing this stat: “88% of organizations use AI” is true but (a) describes breadth, not depth or agentic maturity, and (b) is measured against a bar that’s been lower since 2020 than it was 2017-19. If the claim is about agentic AI specifically, McKinsey’s own number is “no more than 10% scaling agents in any given function” (23% scaling something agentic somewhere in the enterprise) — cite that quote directly rather than HAI’s “single digits” paraphrase, which is accurate but less precise.
Other adoption findings: China and Europe posted the largest YoY adoption increases (+13pp and +11pp respectively). Highest-use industry/function pairs: business/legal/professional services knowledge management (58%), technology-sector IT (56%) and software engineering (58%… wait — the exact figures are IT 56%, software engineering 58% per Figure 4.3.3), consumer goods/retail marketing and sales (51%). Reported organizational outcomes: 64% say AI improved innovation, 45% report improved employee/customer satisfaction; only ≤7% report AI worsened any measured outcome.
Diffusion-speed context: generative AI reached 53% adoption within 3 years of its mass-market introduction (2022) — faster than the internet (measured multiple ways, ~76-91% by year 3 depending on source) or the personal computer (~69% by a comparable point), per The Project on Workforce at Harvard, 2025, using genaiadoptiontracker.com data. Platform-level Claude usage data (Anthropic Economic Index) shows computer/mathematical tasks as the largest usage category (~35-40% of activity throughout 2025) and a shift from augmentation-dominant toward automation-dominant conversation patterns over the year (41% automation share in January 2025 → peaked 49% in August → 45% by November).
What Is Generative AI Worth? — The $172B Consumer-Surplus Methodology
This directly resolves the drain question on the $172B/yr methodology (willingness-to-pay vs. revealed preference).
It is a stated-preference willingness-to-accept (WTA) study, not a revenue figure and not a revealed-preference/behavioral study. Source: Brynjolfsson, Collis, Eggers, Kazinnik & Nguyen, “What is Generative AI Worth?” (2026), available at SSRN.
Methodology: the study ran online choice experiments — not observational/revealed-preference analysis — surveying US adults in 2025 (N=1,400) and early 2026 (N=2,000). Rather than measuring productivity effects or observing actual purchases, respondents were directly asked how much compensation they would need to accept to give up access to all generative AI tools for one month. This “compensating variation” / consumer-surplus measure is the standard technique for valuing goods that are largely free and already in consumers’ possession (where there’s no market price or transaction to observe).
Findings:
- Total US consumer surplus from generative AI grew from 172B (2026), a 53.6% increase.
- US adults using generative AI grew from 95M to 115M (+21%) over the same period.
- Median consumer surplus per user tripled: 11.40 (2026), a 236% increase.
- Average consumer surplus per user grew 27.6%: 125 (2026).
- The gap between median and average (both far below revenue-scale numbers) indicates a right-skewed distribution — a subset of heavy users capture disproportionate value.
- This consumer-surplus estimate dwarfs AI companies’ actual revenue, which the report frames as consistent with Nordhaus (2004): innovators historically capture only ~3% of the total social returns their technologies generate (i.e., the 44B combined OpenAI+Anthropic revenue captured by producers).
- The strongest individual-level predictor of surplus was usage frequency, followed by work-use, number of different products used, and paid-subscription status. Practical-guidance, technical-help, and information-seeking use cases were associated with higher surplus.
Practical implication: this number should never be characterized as “AI companies generated 172B/yr to them, based on what they’d need to be paid to give them up for a month.”
Jobs — Labor Market and Productivity
AI Labor Demand and Hiring
Lightcast job-posting data: AI-skill job postings hit new highs in 2025, led by Singapore (4.69% of all postings), Hong Kong (3.48%), Luxembourg (3.43%), Spain (3.31%); the US reached 2.56%. Within the US, Python appeared in 258,674 AI postings (+391% vs. 2013-15 baseline); the fastest-growing specific skill categories were scalability (+733%), workflow management (+818%), and Amazon Web Services (+1,358%) — evidence that demand is shifting from “familiarity with chat tools” toward “skills to build and operate AI systems at scale.” Agentic-specific postings grew fastest of all: “Agentic AI” mentions grew from 0.69% to 18.90% of AI-agent-related postings share (+2,643%) between 2024 and 2025.
California leads US state-level AI job postings (170,881 postings, 17.2% of national total, though down from >25% in 2012), followed by Texas (80,547, 8.1%) and New York (66,029, 6.6%). By posting-density relative to each state’s total job market, Washington DC (6.18%) and Delaware (4.43%) lead.
LinkedIn hiring/talent data: Indonesia (+31.7%), Croatia (+27.8%), and Belgium (+21.5%) had the highest relative AI-vs-overall hiring growth in 2025. Israel has the highest AI talent concentration among LinkedIn members (2.10%), followed by Singapore (1.82%) and Luxembourg (1.60%). The US is a net importer of AI talent (+1.22 net migration per 10,000 LinkedIn members) but trails Luxembourg (+5.23) and the UAE (+4.40). Gender: women represent 30.5% of AI talent globally; in the US specifically, 34.3% (unchanged for years — “gender ratios have for the most part stayed flat since 2016, despite an expanding AI workforce”).
Productivity Trends — the Full Study List
This directly resolves the drain question on the productivity-study methodology/list behind the 14-26% figures.
Micro-level studies (individual worker/task productivity, Figure 4.4.27):
| Study | Occupation | AI application | Productivity change | Who benefited most |
|---|---|---|---|---|
| Reimers & Waldfogel (2026) | Authors | LLMs for content | +200% output volume (releases tripled) | New entrants (drove quantity); pre-AI authors (maintained quality) |
| Shen & Tamkin (2025) | Software engineers | Learning new libraries | 0% (statistically insignificant) | High scorers (65%+) who used AI for conceptual inquiry, avoiding “learning penalties” |
| Becker et al. (2025), METR | Developers | Open-source tools | -19% (slower) | None — significant gap between perceived help and actual performance |
| Brynjolfsson et al. (2025) | Support agents | Conversational assistant | +14-15% (issues resolved/hour) | Less experienced/skilled agents (30-35% gains) |
| Cui et al. (2025) | Software developers | GitHub Copilot | +26% (completed pull requests) | Junior and less-experienced workers |
| Ju & Aral (2025) | Marketing teams | Multimodal ad creation | +50% (output/worker) | Human-AI teams (shifted focus from social coordination to task execution) |
| Choi & Xie (2025) | Accountants | AI-based accounting | +55% (weekly client throughput) | Experienced accountants (used AI confidence scores to target oversight) |
Note the METR finding (Becker et al. 2025) is the most widely cited counter-example: experienced open-source developers became measurably slower using AI assistance, with a disconnect between how helpful they believed the tools were and how they actually performed. METR has not been able to replicate this in a later study, “primarily due to a growing reluctance among developers to work without AI” and because developers by late 2025 were likely already sped up by AI relative to the original study period — a methodological caveat worth carrying forward with the headline stat.
Macro-level studies (firm/economy-wide, Figure 4.4.28):
| Study | Scope | Insight | Productivity/employment impact |
|---|---|---|---|
| Aldasoro et al. (2026) | 12,000 European firms (2019-2024) | AI adoption increases efficiency without reducing short-run employment; training significantly boosts gains | +4% labor productivity; +5.9pp gain per 1% spent on training |
| Yotzov et al. (2026) | 6,000 executives (US, UK, DE, AU) | High adoption but minimal realized productivity impact to date | +1.4% projected productivity boost; +0.8% projected output increase; -0.7% projected employment reduction (over 3 years) |
| Brynjolfsson (2026), FT | US economy | Output “decoupling” from labor input visible; “J-curve” hypothesis (costs absorbed before gains show up) | 2.7% US productivity growth in 2025 (nearly double the 1.4% prior-decade average) |
| Filippucci et al. (2025), OECD | G7 economies, 10-year horizon | Projected gains vary by sectoral specialization (high in finance/ICT, low in manufacturing) | +0.4 to +1.3pp (US/UK) vs. +0.2 to +0.8pp (Italy/Japan) annual labor productivity growth |
| Frank et al. (2026) | LinkedIn profiles + US unemployment insurance records | Deterioration in AI-exposed labor markets began pre-ChatGPT, in early 2022 | Negative entry rates for AI-exposed roles |
| Brynjolfsson et al. (2025), “Canaries in the coal mine” | US payroll data (ADP), through 2025 | Large employment declines for junior workers in AI-exposed fields | -15% to -16% employment for early-career workers |
| Hosseini Maasoum & Lichtinger (2025) | 62M workers / 285,000 US firms | ”Seniority-biased technological change” — AI substitutes junior labor, leaves senior roles intact | Sharp decline in junior employment driven by slower hiring |
| St. Louis Fed (2025) | US general labor market | Back-of-envelope calc from self-reported time savings | +1.1% to +1.3% labor productivity increase |
| Penn Wharton Budget Model (2025) | US economy | Projects AI’s current contribution to total factor productivity | +0.01pp contribution (negligible) |
Workforce Impact — the 22-25 Age-Group Decline
The report’s headline youth-employment stat traces to Brynjolfsson et al. (2025), “Canaries in the coal mine? Six facts about the recent employment effects of artificial intelligence” (Stanford Digital Economy Lab), using normalized US payroll (ADP) headcount data through September 2025:
- Software developers ages 22-25: employment fell ~20% from the 2022 peak by September 2025, even as headcount for developers 26+ continued growing over the same period.
- The same generational pattern holds for customer-support agents.
- Controlling for firm-by-time fixed effects (within-firm, same-period comparisons, isolating AI exposure from broader shocks like interest-rate pressure or sector slowdowns): employment among 22-25-year-olds in the most AI-exposed occupations fell ~16% relative to the least-exposed, and this gap began widening in mid-2024 and has grown steadily since.
Important nuance the report itself flags: this is not simply “AI is destroying jobs across the board.” Unemployment data (Felten et al. 2021 exposure crosswalk; Eckhardt & Goldschlag 2025) shows unemployment rose for every AI-exposure quintile between 2022 and early 2025 — and rose more for the least-exposed workers (quintile 1: +0.94pp) than the most-exposed (quintile 5: +0.30pp). The report’s conclusion: “AI exposure alone does not seem to be driving recent unemployment trends, but it appears to play a part in broader macroeconomic conditions and organizational changes.” The clean story is about hiring (entry-level pipelines shrinking in AI-exposed fields), not about broad-based unemployment.
Employer expectations point toward acceleration: per McKinsey’s survey, one-third of organizations anticipate a workforce decrease in the coming year (35% at $1B+ revenue orgs vs. 30% at smaller firms); in nearly every business function, expected headcount decreases for the next year exceed actual decreases observed over the past year — most pronounced in service operations, supply chain/inventory management, marketing and sales, and software engineering.
Worker sentiment vs. actual deployment (Shao et al. 2026, 844-task survey across 104 occupations, N=3,618): 46.1% of workers actively want AI to take over specific tasks they do — mainly because it would free time for higher-value work, the task is repetitive/tedious, or it would improve output quality. But actual Claude.ai usage doesn’t track this desire: tasks with the highest “automation desire” scores account for only ~1.3% of Claude.ai usage, suggesting the gap between what workers want automated and what’s actually being automated is still wide.
Robot Deployments (brief)
International Federation of Robotics (IFR) 2025 data: 542,000 industrial robots installed globally in 2024 (+0.2% YoY, essentially flat); operational stock reached 4.664M (+9% from 4.282M in 2023). China installed 295,000 (54.4% of the global total, up from 20.8% in 2013) — 6.6x Japan (44,500) and 8.6x the US (34,200). Collaborative robots grew from 2.8% of new installations (2017) to 13.6% (2024).
McKinsey Methodology (Primary Source)
Directly resolves the drain question on McKinsey’s own report and methodology. Verified directly against McKinsey’s own site and PDF (not just HAI’s citation of it).
McKinsey’s report is “The State of AI in 2025: Agents, Innovation, and Transformation,” published November 5, 2025, by Alex Singla, Alexander Sukharevsky, Bryce Hall, Lareina Yee, and Michael Chui, with Tara Balakrishnan (QuantumBlack, AI by McKinsey). mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai. No newer edition exists as of this writing (2026-07-02).
- Fielded: June 25, 2025 – July 29, 2025 (online survey), per McKinsey’s own methodology note, verbatim: “garnered responses from 1,993 participants in 105 nations… Thirty-eight percent of respondents say they work for organizations with more than $1 billion in annual revenues… data are weighted by the contribution of each respondent’s nation to global GDP.”
- HAI also draws on eight prior annual editions of the same survey for trend lines back to 2018: The State of AI in Early 2024, …How Organizations Are Rewiring to Capture Value (2024), …Generative AI’s Breakout Year (2023), …and a Half Decade in Review (2022), plus the 2021, 2020, 2019 (“AI Proves Its Worth, But Few Scale Impact”), and 2018 (“AI Adoption Advances, But Foundational Barriers Remain”) editions.
This is self-reported survey data (McKinsey, not independently verified), which the AI Index explicitly caveats: “the results are self-reported and should be viewed as directional rather than comprehensive.”
McKinsey’s own productivity/value metric is distinct from the academic productivity-study table above. McKinsey does not run controlled productivity experiments; its internal proxy is a self-reported EBIT-impact bracket, where “AI high performers” are organizations reporting >5% EBIT impact attributable to AI plus self-assessed “significant value” (roughly 6% of the 1,993 respondents, n≈109), analyzed via a relative-weights analysis — a correlational, not causal, method. The specific study-level percentages cited earlier in this article (customer support +14-15%, software development +26%, marketing +50%, etc.) are Stanford HAI’s own literature synthesis of independent academic studies (Brynjolfsson et al., Cui et al., Ju & Aral, and others, each individually cited above) — not McKinsey survey output. Conflating “McKinsey’s 88%/EBIT figures” with “the academic 14-26% productivity-study figures” is a common secondary-source error; keep the two source families separate when citing.
Try It
- For “how mature is enterprise AI, really” conversations: use the three-tier McKinsey breakdown (88% / 79% / single-digit-agents) instead of citing “88% adoption” alone — it’s a materially different claim depending on which tier you mean.
- For consumer-value framing in SaaS pitches: cite the WTA methodology explicitly (“users say they’d need $125/month on average to give this up”) rather than implying it’s revenue or market size.
- For labor-market risk framing: the 22-25 age-cohort decline (Brynjolfsson et al. 2025) is the most defensible single stat; pair it with the unemployment-by-exposure-quintile finding to avoid overclaiming (“hiring pipeline effect,” not “broad unemployment effect”).
- For productivity ROI decks: pull the specific study matching your use case from the micro-level table (support: Brynjolfsson et al. +14-15%; dev: Cui et al. +26%; marketing: Ju & Aral +50%) rather than citing an unsourced “14-26%” range — cite the study, not just the range.
Open Questions
- No OECD-specific AI Index or comparable annual publication was found in the Chapter 4 works-cited list; the OECD Filippucci et al. (2025) productivity study is a working paper, not an “OECD AI Index.” See Chapter 8 deep-dive for the fuller OECD policy-observatory answer.
- McKinsey’s underlying microdata (beyond what’s published in the report PDF) was not accessed, so a single precise blended ”% of orgs with any agent at scale, across all functions combined” figure is not available — only the per-function ceiling (“no more than 10% in any given function”) and the enterprise-wide “scaling something agentic somewhere” figure (23%).
Related
- Stanford HAI AI Index Report 2026 — parent top-takeaways article
- Chapter 8 — Policy and Governance Deep-Dive
- Marketing Cuts — cross-topic companion in ai-marketing
- AI Agents Unleashed — 2026 Playbook
- 2026 AI-Work Restructuring