The metric

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.

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 build and growth. ← 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.

01 · Executive summary

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. McKinsey
• The 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. Census
• Agents are the frontier and the shakeout. 23% are scaling agents somewhere; Gartner expects 40%+ of agentic projects canceled by end-2027. Gartner
• Where it works, gains are large. Productivity grew ~4x faster in the most AI-exposed industries; revenue per employee 3x faster. PwC
• A 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

02 · Key findings

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 · Crunchbase
• AI lifted productivity ~27% in the most exposed industries — while cutting entry-level employment ~13%.PwC · Stanford
• CEOs 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

03 · Market adoption

Near-universal, unevenly deep

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

04 · Economic impact

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

05 · By business function

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

06 · AI-native companies

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

07 · Failure & limits

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

08 · The category

"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 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

09 · From the Leapd platform

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.

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.

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.

Source: Leapd platform data, 2026 (n ≈ 1,000+ businesses).

10 · Method & sources

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.