Every number on this site carries its method.
Source, sample size, field dates, and sponsor — inline, every time. Government, academic, and standards-body figures lead; vendor-sponsored numbers appear only as corroboration and are labeled vendor-sponsored where they stand. A statistic that can't meet that bar doesn't ship.
Memento mori.
Every claim on this site carries its date and, where it can age, a review-by date. Superseded content gets a banner, never a silent edit. Models die; guidance dies (SR 11-7, rescinded April 17, 2026); consultants die. The record is built to survive all three. The long-form version of this discipline is the field note The Gold Watch.
Load-bearing figures: government, academic, standards.
| Figure | Source |
|---|---|
| 19.8% of U.S. businesses use AI; 37% at 250+ employees; 32% at 100–249 (mid-market); Finance & Insurance 33.9%, Information 39.7% | U.S. Census Bureau, Business Trends & Outlook Survey (~200,000-firm panels), period ending May 3, 2026 |
| ~41% of workers use generative AI for work (Nov 2025); ~78% of the labor force works at AI-adopting firms | Federal Reserve FEDS Note, "Monitoring AI Adoption in the U.S. Economy," April 3, 2026 |
| Documented AI incidents rose 233 (2024) → 362 (2025); organizations with no responsible-AI policy fell 24% → 11% | Stanford HAI AI Index, 2025/2026 editions |
| Average human-AI combinations underperform the better of human or AI alone — oversight must be designed, not assumed | Vaccaro, Almaatouq & Malone (2024), meta-analysis of 100+ experiments |
| AI raises average productivity ~15%, novices +30–35%; quality declines for the most experienced workers | Brynjolfsson, Li & Raymond, QJE 2025 (N=5,172, field experiment) |
| High-risk AI must be "effectively overseen by natural persons," including automation-bias guards; enforcement powers Aug 2, 2026 | EU AI Act, Regulation (EU) 2024/1689, Art. 14 |
| "Valid & Reliable" is a necessary condition of trustworthy AI | NIST AI RMF (AI 100-1); Generative AI Profile (AI 600-1) |
| Supervision obligations apply "irrespective of the technology" | FINRA Regulatory Notice 24-09 (June 2024) |
| SR 11-7 rescinded; revised interagency guidance places generative and agentic AI outside its scope; RFI forthcoming | OCC Bulletin 2026-13, April 17, 2026 — Briefing #1 |
| AI is the #1 mid-market investment destination (28% of allocation); cybersecurity the top named risk | NCMM Year-End 2025 Middle Market Indicator (N=1,000 executives, $10M–$1B revenue; fielded Dec 2025) |
| Documented public incident: Deloitte Australia refunded the final installment (~AU$97K) of an AU$439K DEWR report containing AI-fabricated citations and a fabricated court quote | October 2025; primary reporting incl. Fortune (Oct 7, 2025), CFO Dive; AI Incident Database #1193 — cited as a dated event, not a survey |
Corroboration only — always labeled.
| Figure | Source + label |
|---|---|
| 88% adoption; 39% report any EBIT impact; ~6% high performers; high performers run defined validation 65% vs 23%; 51% had at least one AI incident | McKinsey State of AI, Nov 2025 (N=1,993; fielded Jun–Jul 2025) — vendor-sponsored |
| 59% in production; 16% high measurable value; 92% agree agents need guardrails while 48% have defined rules; legacy systems the top scaling barrier (69%) | HBR Analytic Services / Appian pulse, Apr 29, 2026 (N=385; fielded Mar 2026) — vendor-sponsored |
| 6% fully trust AI agents with core processes; 43% routine-only | HBR pulse / Workato & AWS (N=603; fielded Jul 2025) — vendor-sponsored |
| 91% mid-market generative-AI adoption; expertise gap as key risk | RSM US Middle Market AI Survey 2025 (N=966; fielded Feb–Mar 2025) — professional-services vendor |
| 40% received workslop in the prior month; ~1h56m to resolve each; ~$186/employee/month | Stanford Social Media Lab + BetterUp (N=1,150 desk workers; Sept 2025) — survey-based self-report |
| Augmentation 52% vs automation 45%; task success falls roughly 60% → 45% as task length grows | Anthropic Economic Index, Jan 2026 — vendor data |
What we deliberately don't cite: the widely quoted "95% of AI pilots show no P&L impact" figure. Its methodology — 52 interviews and a six-month P&L window — doesn't meet this page's bar. The defensible version of that story is McKinsey's 39%/6%, labeled above.