Reading agentic markets

Three people are guessing right now. A procurement lead comparing agent vendors with no comparable numbers. A fund analyst sizing the agentic market from vendor press releases. A policy economist asked what AI did to productivity this year, answering from surveys. The ledger replaces all three guesses, first in business terms, then in economics.

The story, in business language

Every settled unit of work is a receipt with both sides of a market on it: what a buyer was willing to pay (the budget, committed before execution), and what delivery actually cost, all-in. Enough receipts, and the questions that matter stop being rhetorical. Procurement compares vendors on delivered value per dollar on their own unit types. The analyst sizes the market in minted quirqs instead of bookings. The economist reads productivity off the ratio of value delivered to cost, quarter over quarter, instead of waiting three years for tax data. Exploratory, and labeled open in the claim ledger: no market publishes this today. The point of this page is what becomes readable the day markets do.

What each reader gets, quantified

Using the worked quarter from the dashboard (illustrative arithmetic, consistent across this section):

The market objects, mapped

Where budgets clear against real money, B is a demand observation, not a survey answer. Where c_q falls with tenure, the supply curve is shifting outward in real time, timestamped.

The economics, as you scroll deeper

A demand elasticity, worked. In the dashboard quarter, cost per quirq fell from 0.32 to 0.18 while monthly units rose from 2,100 to 4,800. The arc elasticity of demand for agentic work on those unit types:

ε = Δln(Q) / Δln(p) = ln(4800/2100) / ln(0.18/0.32)
  = 0.827 / (−0.575)  ≈  −1.44

Demand is elastic: a 44% price fall more than doubled quantity. That single number, estimated on real ledgers instead of this illustration, is the demand-expansion thesis of The Future of Work made falsifiable: latent demand exists exactly where |ε| > 1 holds as c_q keeps falling.

Supply dynamics. From the cost model, c_q(t) = c_exec + k·H(intent | M_t) per quirq-dollar: supply shifts outward not from better models alone but from accumulated environment memory, and the ledger separates the two (a model swap moves c_exec discontinuously; tenure decays H smoothly). An industry supply curve is the c_q distribution across builders, observed per window.

Productivity, growth-accounted. Audit-corrected QER* is output over input for agentic work, so its log change is a total-factor-productivity reading:

g_TFP(T) = Δ ln QER*(T)        worked: ln(5.6/3.1) ≈ 0.59 over the quarter

Fifty-nine log points in a quarter is a technology-adoption transient, not a steady state; the steady-state series, across firms, is the number the productivity-paradox debate has been missing.

The substitution margin, estimable. For each check type i, let s_i(t) be the share settled by agents rather than routed to humans. Fit the diffusion

s_i(t) = 1 / (1 + e^{−r_i (t − t_i)})

and the parameters are the automation frontier: r_i is how fast task type i is moving, t_i is when it crosses half-automated. Ranked r_i across an economy's check types is a live substitution map, observed as a side effect of settlement rather than surveyed.

Structure. Concentration is HHI over builder shares of minted quirqs per unit-type market; entry shows up as mass moving into the c_q distribution's left tail; price wars show up as α compressing toward c_q (the participation band of the builder page).

The honest caveats

Ledgered firms are early adopters: selection bias until coverage widens. Notional budgets are weaker observations than cleared ones and should be flagged in any series. Goodhart applies to markets too: only audit-corrected (QER*) series deserve publication. And every aggregation must respect the relativity rules: ratios and trends travel, raw totals do not. These are the standard defects of administrative data, and the standard econometrics applies.

One more frontier, further out: quirqs and quantum computing.