AEVF Index · v0.4 ·
The AI Economic Viability Frontier
Welcome to the AI Economic Viability Frontier.
Nearly $1.8 trillion of cumulative AI investment is underwriting one proposition: that machine intelligence will diffuse through the entire economy, physical, digital, financial, and knowledge work alike, as a materially superior means of producing economically valuable goods and services. Yet that is not how progress is measured. Benchmarks measure capability. Demos establish possibility. Token prices measure one input.
Silicon Valley's question is: can the agent do X. The more precise question is: can the agent do X in an economically viable way, producing the output consistently, at the quality threshold the market actually demands, within the reliability, oversight, and liability constraints that bind in production, at an all-in cost that materially improves on the incumbent method.
Capability asks whether a task is possible. Economic viability asks whether it has become reproducible as an economic output. That is the threshold that matters for diffusion, and the crossing that changes production regimes.
This index is an attempt to steer the conversation on justified AI capex toward that measurement: a more directionally precise account of AI progress, marked at the point where capability becomes economically viable production.
The task universe below currently holds 50 tasks and expands release by release, C / R < V; every number recomputes live as you change the assumptions.
Assumptions · adjust and everything recomputes
The frontier right now
Each point is one task under the current assumptions. Horizontal position is the probability an AI-produced output is accepted; vertical position is all-in cost per successful output as a share of what the market pays, on a log scale. The line at 100% is the frontier itself: below it, producing the outcome costs less than buyers pay for it.
The task universe · 50 tasks, expanding
A preview of the universe: the 15 tasks with the highest viability ratio under the current assumptions, across 48 industries. Click a column header to re-rank the preview, or a row for the task card. The full universe, searchable and filterable, is one click below.
| Task | Market price | AI cost / success | TCEVO | Φ | Compression | R | Grade |
|---|
The equations
Ask the index
A chat analyst grounded in the full corpus: every task card in the index, sourced prices, reliability evidence, model pricing, and the methodology. Bound by the same honesty rules as this page.
The thesis · why this index exists
The capex bet. Global AI investment is forecast to exceed $1 trillion in 2026 (Goldman Sachs Research), with cumulative investment since 2022 approaching $1.8 trillion by year end. The entire return case for that capital rests on one assumption: that in the interim before any RSI-to-ASI transition, AI agents actually diffuse into the economy, performing economically valuable tasks cheaper, faster, and better than the humans who perform them today. If agents do not diffuse, the capex is not justified by anything. Diffusion is the whole game, and diffusion is a measurable claim.
The industry measures the wrong thing. Token prices, benchmark deltas, and intelligence-per-dollar ratios are the standard proxies for AI progress, and none of them crosses an economic threshold. Cheaper tokens are not cheaper accepted work. Benchmark gains invite optimization against the measure itself rather than true generalization, a dynamic formalized in Qiu, Yu, and Xu, "On Benchmark Hacking in ML Contests" (2026). A model can climb every leaderboard while the economic question, does anyone accept and pay for the output, goes unanswered. This index exists because the underlying objective of the buildout is not benchmark supremacy; it is displacement of human production cost, and that has to be measured in dollars per accepted outcome or it is not being measured at all.
The Neo Firm endgame. The parent thesis behind this instrument holds that as the marginal cost of compute trends toward zero, the cost of producing an economically viable service falls toward, more or less, the price of compute, and the winners are Neo Firms: institutions built AI-native from zero that sell the outcome directly, rather than software sold into incumbents who keep their legacy cost structure. In that regime a specific inversion happens to model choice. Below the viability threshold, capability is everything. Above it, once a task's reliability floor and cost test are both cleared, model selection collapses into procurement: volume flows to whichever model is most token-efficient at producing that accepted output. Which means the market will need a precise definition of token efficiency, and dollars per million tokens is not it. TCEVO, inference dollars per accepted economic output, is this project's candidate. Every row in the table above is that number, computed live.
In most rows the binding cost today is not inference at all; it is human review, and the binding constraint is reliability. That is what pre-diffusion looks like in the data: tokens are already cheap enough, and the frontier now moves on R, not on price. Where the table shows 10x to 100x compression with deterministic verification, that is where Neo Firms get built first. Full framing and every caveat in the methodology.
Share the index