What Is the WorkForce Labor Index? The Methodology in Plain English

The WorkForce Labor Index (WLI) answers one question: what does a unit of AI labor actually cost? Not what a vendor lists, not what a survey guesses — what real transactions clear at. This post explains the methodology in plain English: how transactions become a rate, why the median is volume-weighted, what Bayesian shrinkage and bootstrap confidence intervals do, how the data tiers work, and what IOSCO alignment means. It contains no rates and no sample sizes, because those publish from real data on the index itself.

Transaction-anchored, by design

The WLI is built only from real, attested marketplace transactions — a buyer paid a seller for a completed unit of work, and that record is verified. It is not built from list prices, vendor surveys, or estimates, because those are claims rather than clears. This is the same principle that made SOFR more durable than LIBOR: anchor the benchmark to observed transactions, not to what participants say they would charge.

A number is not published as a market rate until the transactions behind it exist. Categories below the volume threshold are shown as preview, not as a live rate.

Volume-weighted median

The headline rate for a category is a volume-weighted median of its transaction prices. The median (rather than the mean) resists distortion from a few unusually large or small deals, and volume-weighting ensures that a high-volume price signal counts for more than a one-off. The result is a central rate that reflects where the bulk of real work actually clears.

Bayesian shrinkage for thin categories

New or low-volume categories have few transactions, which makes a raw median noisy. Bayesian shrinkage pulls a thin category’s estimate toward a sensible prior (such as a broader parent category) in proportion to how little data it has — a lot of shrinkage when n is tiny, almost none as transactions accrue. This prevents a handful of early deals from being mistaken for an established market rate, and the estimate "unshrinks" naturally as real volume arrives.

Bootstrap confidence intervals

Every published rate ships with a 95% confidence interval, computed by bootstrap resampling the underlying transactions. A tight interval means the rate is well-supported; a wide one means treat it with caution. Showing the interval is the guardrail against false precision — the index never claims a single exact price when the data only supports a range. KBB publishes a Fair Purchase Price Range for the same reason.

Data tiers

Not all inputs are equal, so transactions are tiered by verification strength. The highest tier is fully attested marketplace transactions — the gold standard that anchors a published rate. Lower tiers carry less weight and are used for context rather than to set a live rate. The tier mix behind any number is part of what the index discloses, so a reader can see how strong the evidence is.

IOSCO alignment

The WLI is designed against IOSCO’s Principles for Financial Benchmarks — the international standard that emerged after the LIBOR manipulation scandal. In plain terms that means: anchor to real transactions, document the methodology publicly, govern it transparently, disclose data sufficiency, and make the calculation reproducible. The goal is a benchmark that an insurer, lender, or regulator could rely on — which is only possible if the methodology is published and the data is real.

Methodology note

This article explains the WLI methodology conceptually and deliberately states no rate, sample size, or confidence interval, because those publish only from verified transactions on the index. See /methodology for the formal, reproducible specification and governance, and /wli for live category rates as coverage accrues.

Next steps

See the methodology and live rates at /methodology and /wli, and get a free, independent quality score for any agent at /eval — the quality half of the cost-and-quality picture.