How this index is built, and exactly how much to trust it

The question

For which economically valuable tasks has AI pushed the expected cost of producing an acceptable outcome below what the market pays for that outcome, and how quickly is that set expanding? A task i is inside the AI Economic Viability Frontier when

where V is willingness to pay for a successful outcome, R is the probability the AI-native process produces an acceptable one, and C is the total expected production cost: inference, tools and APIs, human review, capital overhead, and expected error or liability cost. The primitive metric is TCEVO, token cost per economically viable output, C_inference / R; the all-in analog is FCSO, C_total / R; the headline per-task statistic is the viability ratio Φ = R·V / C.

The reason to measure production cost against willingness to pay, rather than tracking benchmark scores or token prices, is that neither of those crosses an economic threshold. A million tokens of a weak model can be worth less than fifty thousand from a strong one, and a 7-point benchmark gain matters less than a single task crossing from Φ < 1 to Φ > 1. The crossing is the event.

What is observed and what is modeled

Every number in the index belongs to one of four classes, and the dataset labels each:

Nothing in the index to date is a measured trial. The research specification this index implements calls for 20 to 100 graded trials per task with blind expert grading and confidence intervals; that is Phase II. Until then, the correct reading of any single row is "a sourced, disciplined estimate," and the correct reading of the aggregate is directional.

Reliability priors

R means the probability that a professional would accept the output, not a benchmark score. Three gaps separate the two, and the priors discount for all three: benchmarks hand the model a fully specified task while real work fails at scoping; win-or-tie against one human draft is weaker than sign-off under liability; and contamination inflates top scores, most cleanly measured by the roughly 35-point gap between SWE-bench Verified and the contamination-resistant SWE-bench Pro. Where production data exists it dominates: the support prior tracks Intercom Fin's published 76% average resolution rate and its case-study range of 42 to 50%, not tau-bench. Each task carries a low, base, and high R, and a minimum reliability floor; the index reports both the share of tasks that clear cost alone and the share that also clear their floor. The second number is the honest one. Priors are also calibrated across releases: when a batch of new cards is added, each new prior is checked against the closest existing analog (a deterministic form-filling task against the existing form-filling tasks, an open research memo against the existing research memos) so that structurally similar tasks carry comparable R regardless of who authored the card. The v0.3 batch was re-anchored this way; the raw drafts ran about ten points more pessimistic than the v0.2 cards for equivalent work, and the calibration closed roughly half of that gap, leaving the genuinely judgment-heavy new domains (government determinations, construction, pharmacovigilance, root-cause analysis) where the drafts put them. The v0.4 batch (100 cards, 16 new domains including Agriculture, Architecture and Engineering, Trade and Customs, Litigation Support, Financial Crime, Entertainment, and Clinical Services) repeated the exercise: raw drafts ran about four points under analog medians overall and thirteen points under in the creative Entertainment batch; calibration closed roughly half the gap on structurally comparable tasks and left judgment-heavy or legally gated cards (tariff classification, export control, workplace investigations, audiobook performance) where the drafts put them.

Costs

• Inference: per-task input and output token estimates, times a retry multiplier, priced at live per-Mtok rates for the selected model, with a prompt-cache hit-rate assumption (default 50%) applied to input.
• Tools: search, enrichment APIs, OCR, code execution, per task.
• Human review: minutes of role-appropriate review (attorney $180/hr, CPA $120, engineer $110, domain specialist $130, analyst $85, and so on) at a global multiplier the reader controls. In most rows this, not inference, is the dominant cost. That is finding one.
• Capital: 10% of direct cost for orchestration, storage, compliance overhead.
• Risk: an expected-value error cost per task, scaled by risk class, adjustable 0x to 3x.

Known biases, stated plainly

1. Basket selection. The tasks were chosen because they are discrete, gradable, and plausibly AI-producible. The frontier share of this basket therefore runs far ahead of the economy's. Do not quote the headline number without this caveat. The basket is deliberately open-ended: the index adds tasks release by release (v0.1 shipped with 50 tasks in 8 domains, v0.2 with 100 in 16, v0.3 with 200 in 32, v0.4 with 300 in 48), and every aggregate recomputes over whatever the universe holds at the time.
2. Historical-price anchoring. V uses current human-market prices. Where AI supply is already compressing prices (SEO articles, translation, deposition summaries), V overstates durable willingness to pay and Φ overstates the opportunity. The spec's own warning applies: price is not value.
3. No demand response. The index says nothing yet about how much more of a task gets consumed when its price collapses, which is where much of the real economic impact lives.
4. Regulatory viability is not modeled. A tax return needs a preparer's signature; a legal opinion needs a bar member. Several rows are technically and economically viable but legally gated, noted per task.
5. Correlated failure. Treating R as independent per attempt understates risk in workflows where errors compound silently, bookkeeping being the documented case.

Confidence grades

The AI-native institutions test

This instrument exists to test a specific thesis: that AI's economic impact concentrates where production is rebuilt around abundant machine cognition rather than where incumbents buy copilots. The index's cost stack is an AI-native stack; an incumbent adopting the same model inside its existing workflow keeps most of its labor, coordination, and overhead, so its cost line barely moves. Where the table shows 10x to 100x compression against the human market price, the gap is not a productivity improvement, it is a different production function, and it is the opening through which an AI-native firm underprices an industry while keeping software margins. Tasks with high Φ, high autonomy scores, and deterministic verification are the places to look first.

Roadmap

1. Phase II, measurement: run the 10-task pilot from the spec with frozen acceptance criteria, 20+ trials per task, blind grading, and confidence intervals; replace the anchored priors with measured R.
2. Phase III, coverage: expand toward a representative basket and add labor-compensation and transaction-volume weightings alongside value weighting.
3. Phase IV, time series: re-run quarterly; the frontier-expansion chart, share over time, is the artifact this project exists to produce.

Reproducibility

The full dataset (task cards, sourced prices with URLs and retrieval dates, model pricing, reliability evidence) and the compute engine are in the GitHub repository. The site is static; every figure on the index page is recomputed in the browser from the published data, so there is no place for a number to hide.