One number for a new economy #
Adoption surveys count whether a company touches AI. None measure whether AI actually runs the business. The AI-Run Business Index (ARBI) is the first standardized metric that does — a 0–100 composite index scoring execution, not experimentation, maintained by Leapd as the category's annual benchmark.
The evidence on AI's business impact is scattered across dozens of surveys that each capture a fragment. Leapd's role is to vet and reconcile that fragmented evidence into one comparable standard. Its spine is 30+ external benchmarks; the published score can be reproduced from those alone. Leapd's own platform data is a secondary calibration layer. Every weight is disclosed so it can be quoted, reproduced, and challenged.
Each dimension reduces to a measurable indicator, so an individual business's score can be computed directly from these six inputs — the basis for a per-company AI-Run Score — while the economy-level reading normalizes public benchmarks.
| Dimension | How a business is scored on it | Weight |
|---|---|---|
| Automation depth | % of recurring tasks executed autonomously, without human approval | 25% |
| Value capture | Share of revenue & gross margin attributable to AI-run functions | 20% |
| Revenue leverage | Revenue per employee vs. the sector median | 20% |
| Speed to revenue | Time from idea → live → first revenue | 15% |
| Function coverage | # of core functions (build, market, sell, support, ops) running end-to-end on AI | 10% |
| Reliability (penalty) | Human intervention rate + rollback / abandonment rate | −10% |
Why these weights. The weighting follows this report's central finding — that execution, not adoption, separates a business run by AI from one that merely uses it. Automation depth carries the most weight (25%) because autonomous execution is the definitional core of "AI-run." Value capture and revenue leverage are next (20% each) because they separate real transformation from efficiency theater — the 88%-vs-6% gap this report documents. Speed (15%) and coverage (10%) reward businesses that run end-to-end and fast, not in a single function. Reliability is a penalty (−10%) because failure and ungoverned agents are large and measured (§7); ignoring them would overstate maturity. Weights are fixed for the 2026 edition and published so the score can be recomputed, audited, or contested.
The 2026 bands
0–20 · Experimenting— isolated tools, no workflow change.** 20–40 · Adopted, not run**— AI in a few functions; little autonomy or value capture.← the mainstream economy (~30).40–70 · Executing— AI runs whole functions with measurable ROI. 70–100 · AI-run**— AI runs buildandgrowth.← the AI-native frontier (~80).
ARBI is directional by design — it compares maturity bands, not false-precision point scores. All inputs and weights are disclosed below.
On the AI-Run Business Index, the mainstream economy scores ~30/100 while the AI-native frontier scores ~80 — a 50-point execution gap between using AI and being run by it.
Adoption is saturated. Transformation is rare. #
Adoption is effectively saturated but shallow. 88% of organizations use AI; nearly two-thirds have not begun scaling it; only ~6% tie it to real profit.McKinseyThe headline depends on definition. Self-reported surveys say 88%; the U.S. Census Bureau's firm-level measure is ~18–20%. Both are right — they measure different things.CensusAgents are the frontier and the shakeout. 23% are scaling agents somewhere; Gartner expects 40%+ of agentic projects canceled by end-2027.GartnerWhere it works, gains are large. Productivity grew ~4x faster in the most AI-exposed industries; revenue per employee 3x faster.PwCA new archetype is visible: the ultra-lean, AI-native company — Cursor, Lovable, Midjourney — posting revenue-per-employee 10–100x the SaaS norm.Governance gates the next phase.~80% of organizations lack a mature model for autonomous agents even as ~74% plan to use them within two years.Deloitte
The defining contradictions of 2026 #
AI adoption is 88% — but only ~6% of companies capture real profit from it.McKinsey** 95% of enterprise AI pilots fail — yet AI took ~half of all global VC (~$211B) in 2025.MIT·CrunchbaseAI lifted productivity ~27% in the most exposed industries — while cutting entry-level employment ~13%.PwC·StanfordCEOs report efficiency gains (56%) far more than profit (34%) or revenue (32%).PwC CEO Survey A ~50-employee company (Cursor) reached ~$2B revenue — about $40M per employee.**AI Business
Near-universal, unevenly deep #
| Measure | Figure | Evidence |
|---|---|---|
| Enterprises using AI in ≥1 function | 88% | McKinsey |
| U.S. firms using AI operationally | 18–20% | Census |
| Large firms (250+ employees) | 37% | Census |
| Small firms (<20 employees) | <20% | Census |
| EU firms using AI / intensively | 70% / 7% | ECB |
| Scaling AI agents somewhere | 23% | McKinsey |
| Touching fully autonomous agents | 15% | Gartner |
The definition gap, explained. McKinsey's 88% surveys enterprise leaders about any use in any function; the Census measure asks a representative panel about operational use in the prior two weeks. The gap is a measurement artifact, not an error — the Census figure is the conservative operational floor. By sector, Information leads at 39.7% and Finance at 33.9%, versus Retail at ~14% (national 19.8%). Census · ECB
88% of enterprises report using AI — but fewer than one in five U.S. businesses actually run it in day-to-day operations, and most that do use it in three functions or fewer.
Efficiency first, profit much later #
Productivity growth nearly quadrupled in the most AI-exposed industries (7% → 27%); revenue per employee grew 3x faster than in the least exposed. But the bottom line lags badly.
Among the winners, the pattern is consistent: PwC found firms with mature AI foundations were 3x more likely to report meaningful returns, and AI-in-products firms saw ~4pp higher margins. AI-skilled workers command a 56% wage premium. PwC
Labor: augmentation vs. displacement
Entry-level U.S. workers (22–25) in the most exposed jobs saw a 13% relative employment decline since late 2022, while older workers held steady or grew 6–9%. The Census reports AI-related job cuts in only 2% of firms; the ECB found AI-intensive EU firms were more likely to hire. The displacement is real but concentrated at the entry level. Stanford · Census · ECB
AI lifted productivity ~27% in the most exposed industries while cutting entry-level employment ~13% — the gains and the losses are landing on different people.
Where AI actually does the work #
Engineering: Copilot is used by 90% of the Fortune 100 and generates ~46% of code — though a rigorous trial found experienced devs 19% slower on complex work; gains concentrate in routine code. Customer support is the clearest end-to-end case: Klarna's assistant does the work of ~853 agents (~$60M/yr) — but the company rehired humans for complex cases after over-automating. Sales & marketing is the #1 adopted function (52%) yet shows the weakest ROI relative to spend. GitHub · METR · CX Dive
AI's measured ROI is highest in customer support and coding — and weakest where the most money is spent: sales and marketing pilots.
The headcount-to-revenue rule, broken #
The strongest evidence that "AI-run business" is real is revenue per employee — output per human, plotted below on a log scale against the traditional SaaS norm.
For context, top-tier SaaS historically targeted ~$200K–$500K per employee. Cursor's ~$40M is two orders of magnitude higher. No verified one-person billion-dollar company exists yet — but the gap to it is narrowing fast. AI Business · VC Corner The headcount-to-revenue rule is breaking at the frontier — a roughly 50-person company (Cursor) now generates around $40M in revenue per employee, ~100x the traditional-SaaS norm.
The value collapses between adoption and impact #
The incumbent picture is not all collapse — firms with real foundations see real returns (PwC 3x; ECB hire-not-fire; Agentforce resolving 70–82% of cases at named enterprises). But the funnel from adoption to value is brutal, and failure concentrates in pilots bolted onto unchanged workflows.
Gartner estimates only ~130 of thousands of "agentic" vendors are real ("agent washing"); ~80% of firms lack mature agent governance; and a rigorous trial showed developers believed AI sped them up 20% while it slowed them 19% — self-reported ROI is systematically unreliable. Gartner · METR
95% of enterprise AI pilots deliver no measurable profit — and they fail in the org, not the model: the winners redesigned workflows, the losers bolted AI onto unchanged ones.
"AI-run business" is forming, not formed #
Two models anchor the category. The AI economy stack runs Tools → Functions → Orchestration → Autonomous business; most of the economy is stuck at Layers 1–2. The autonomy spectrum maps how far a business has climbed:
| Stage | What AI does | 2026 prevalence |
|---|---|---|
| 1 · AI-assisted | Drafts, suggests, accelerates tasks | ~75% |
| 2 · AI-augmented | Runs whole tasks, human supervises | 23% |
| 3 · Semi-autonomous | Runs whole functions end-to-end | 15% |
| 4 · AI-run | Runs build and growth | frontier only |
Stage 4 is nearly empty today. "AI-run business" is best read as a direction the data is moving, not a population that already exists at scale. This report measures the slope, not a finished state.
The capital signal is unambiguous: AI took ~61% of global VC in 2025 by the OECD's measure, rising to ~80% in Q1 2026 — even as 95% of pilots failed. OECD · Crunchbase
AI took roughly half of all global venture capital in 2025 even as 95% of pilots failed — capital is betting hard on a category that barely exists yet.
What we see across 1,000 businesses running on Leapd #
We analyzed activity across 1,000+ businesses running on Leapd — what the agents executed, where founders intervened, and how fast each business went from idea to revenue. These businesses skew earlier-stage and AI-native, so they read as a leading indicator of where the category is heading rather than a sample of the whole economy.
A business goes from a one-line idea to a live, operating company — website, backend, checkout, and live ad, email, and social campaigns — in under ten minutes, against the weeks a founding team would normally spend on setup.
On Leapd, an idea becomes a live, operating business — website, checkout, and live ad, email, and social campaigns — in under ten minutes.
Once live, the median active business runs about 50 autonomous tasks a week across engineering, content, prospecting, outreach, and operations — work that would otherwise sit on a founder's list. Most of it runs without a human in the loop. Founders start by reviewing roughly 20–30% of what the agents do, but that supervision drops quickly: within about two weeks, most turn on auto-approval and let the system run.
Where they keep a hand in is consistent — the visual and brand choices on their site, and approving LinkedIn posts before they publish. The most fully automated work sits at the opposite end. Prospecting and outreach run with little oversight, because founders quickly learn to trust the AI-written copy, and so does the build itself: unlike coding tools that need constant prompting, Milo ships a scalable, on-brand application on its own.
| What we measure | 2026 reading (Leapd platform) |
|---|---|
| Idea → live business (site, checkout & campaigns running) | under 10 minutes |
| Autonomous tasks per active business / week | ~50 |
| Time to first revenue | ~6 weeks, then ~30% WoW |
| Founder intervention rate | 20–30% → auto-approve in ~2 wks |
| Most-automated functions | Prospecting, outreach & site build |
| Still active at 90 days | ~60% |
Revenue follows. Most businesses reach their first revenue in roughly six weeks, then grow about 30% week over week — though the curve is bumpy and depends heavily on the market. SMB-focused businesses move fastest; those selling into enterprise, regulated industries, or long sales cycles take longer. About 60% are still active at 90 days.
Adoption varies sharply by agent. Cassy, the LinkedIn agent, is used by 70% of businesses for prospecting, engagement, and campaigns, and by 30% for drafting posts. Milo runs outbound: 90% have used it for automated email — finding their ideal customer, reaching out, handling replies, qualifying leads, and booking meetings — and 20% let it produce and manage video ads across Meta and X. Alex, which handles visibility in AI search, is used by about 20% overall, but the split is sharp: 80% of businesses that arrive with an existing company turn it on, versus only ~10% of those starting from an idea. Established businesses already feel the pressure of being found inside ChatGPT, Gemini, and Perplexity; idea-stage builders haven't yet.
| Agent | What it runs | Adoption |
|---|---|---|
| Cassy — LinkedIn | Prospecting, engagement & campaigns · post drafting | 70% · 30% |
| Milo — Email | ICP discovery, outreach, reply handling, lead qualification & meeting booking | 90% |
| Milo — Paid & video | AI-generated video ads, Meta & X campaigns | 20% |
| Alex — AI-search visibility | Visibility tracking, site audit, AEO article generation | 20% overall 80% of existing-business signups · ~10% of idea-stage builds |
Source: Leapd platform data, 2026 (n ≈ 1,000+ businesses).
How this was built #
Prioritizes 2024–2026 data. Source tiers: (1) primary research and official statistics — Census, the Fed, OECD, McKinsey, Deloitte, PwC, Gartner, MIT, Stanford; (2) financial-data trackers — Crunchbase, CB Insights, Carta; (3) company disclosures (labeled self-reported). Where figures conflict, the report shows the range and the cause rather than picking one.
Method notes: the published readings can be reproduced from the public benchmarks alone; the Leapd platform layer (~1,000 AI-run businesses) is a supporting calibration input, not a global sample; readings are reported as maturity bands, not point scores. Leapd publishes the index and operates in this market — the public spine and open weighting let the result be reproduced or contested independently.
Full citations: McKinsey · U.S. Census Bureau · Federal Reserve · ECB · Gartner · Deloitte · PwC (Jobs Barometer + CEO Survey) · OECD · MIT NANDA · Stanford Digital Economy Lab · METR · Crunchbase · CB Insights · Salesforce · Klarna · GitHub. Links are inline throughout each section.