Cost at the unit · one outcome, priced line by line
Cost per outcome
Every one of the 1,750 outcome chains assembled from the week of 10 to 16 August 2026 is priced here one rate line at a time … more
What one outcome is made of
The five components of a fully loaded outcome, summed across the priced sample and divided down to one chain.
The outcome pages carry $30,762 for the same 1,750 chains. This page carries $57,118. The $26,357 difference is entirely method: people are priced at three loaded rates rather than one, model calls at three rates rather than one, and two components the outcome pages never carried at all — platform license at $13,921 and management overhead at $10,495 — are added here. No new spending happened between the two pages.
Loaded rates, license costs and overhead bases are the customer’s figures, and no customer figure has been supplied. more
The evidence behind every dollar
Each cost line declares which of three things its rate is. more
Cost per outcome, by service
8 named services, sorted by what one outcome costs fully loaded. more
Two comparable teams, and why one costs more
Same function, weekly volumes within a factor of two of each other. more
The gap is $19.59 a unit. The largest single driver is what one person-handled unit costs, worth $17.19 of it. The four drivers sum to the gap exactly because they are a decomposition of it, not an explanation offered after the fact. Both sit inside Legal and run within a factor of two of each other on weekly volume. That is the whole test. It is not a claim that the two do identical work, and the four drivers are arithmetic on the published unit economics rather than an opinion about difficulty.
The gap is $16.68 a unit. The largest single driver is what one person-handled unit costs, worth $9.32 of it. The four drivers sum to the gap exactly because they are a decomposition of it, not an explanation offered after the fact. Both sit inside Procurement & Source-to-Pay and run within a factor of two of each other on weekly volume. That is the whole test. It is not a claim that the two do identical work, and the four drivers are arithmetic on the published unit economics rather than an opinion about difficulty.
The gap is $14.71 a unit. The largest single driver is what one person-handled unit costs, worth $10.10 of it. The four drivers sum to the gap exactly because they are a decomposition of it, not an explanation offered after the fact. Both sit inside Human Resources and run within a factor of two of each other on weekly volume. That is the whole test. It is not a claim that the two do identical work, and the four drivers are arithmetic on the published unit economics rather than an opinion about difficulty.
Inference, attributed to the decision it served
The estate reports $1.9m a week of inference. more
The gap between the inference the chains can account for and the inference bill the estate reports. more
Four weekly totals that do not reconcile
The estate already publishes its cost to serve in three places. more
These are four different objects, not four attempts at one number. more
One of the four is nonetheless the one to quote. The tower rollup counts each transaction once against a published rate and reconciles tower by tower against the daily ledger, so it is the number this platform names — $9.1m a week, $471.9m a year. Which Cost Is The Cost shows the reconciliation tower by tower and why the other three are not it. Annualized figures are the weekly figure multiplied by fifty two. They carry no seasonality, no holiday effect and no growth. The modeled year the ledger holds runs 2.3 percent above the published week, so a fifty two week multiplication is a slight understatement of the modeled year rather than an overstatement.
Operations · what if inference costs more than we modeled
Inference is 21.0% of the weekly cost to serve · the modeled rate is the softest input on this page, so it is stress tested here
Every rate on this page is metered except the inference rate, which is modeled from published token prices and average prompt sizes — so it is the input most likely to be wrong. At five times the modeled price the estate still serves a transaction for $18.53 against $31.01 before agents, and the saving only disappears if inference turns out to be 10.9 times dearer than modeled. That is the honest shape of the exposure: material enough to meter properly, not large enough to change the decision. The outcome pages price a chain at one blended human rate and two flat system rates. These pages price each stage at the rate for the function that ran it, the model class it called and the system it ran in, then add the two components the outcome pages never carried at all: platform license and management overhead. The totals therefore differ, and the difference is a change of method rather than new spending. Both numbers are shown side by side wherever they appear.
Actions
What is waiting on a person
Each count below is a cost line, a service or a published pair of numbers that cannot be defended as it stands. more
Defend a divisor on an apportioned line
A real invoice exists behind each of these, but the units it bought are not counted, so the unit cost is whatever the chosen divisor makes it.
Put a meter behind an estimated cost line
These lines carry $35k, 62.1 percent of the priced sample, on a rate nobody counted and a quantity nobody observed.
Answer why two comparable teams cost different amounts
Same function, volumes within a factor of two, and the wider pair still differs by 852 percent a unit. more
Give a service a value denominator
A service that books no value anywhere in its chain has a cost per outcome and nothing to divide it into, so no cost per unit of value can be formed at all.
Explain the inference bill no outcome accounts for
$1.8m a week, 94 percent of the reported inference spend, is attributed to no assembled chain at all.
Operations
What this desk is allowed to start
A surface that only reports is not operable. more
Trigger and bound
This page can price one outcome instance line by line at the rate for the function, model class and system that actually ran it, declare on every line whether that rate is metered, apportioned or estimated, and set two comparable teams beside each other and decompose the gap between them. more
Live observability
What the record shows right now
Every dollar in the priced sample by the evidence behind its rate, an exhaustive partition of $57k across 3 tiers: each of the 10,939 cost lines carries exactly one. more
Current distribution
57,118 dollars in the sample
Is policy and strategy coming to fruition
Whether the written intent is holding here
A fully loaded outcome costs $32.64, 1.86 times what the outcome pages carry, and 62.1 percent of that is estimated rather than measured.
Not holding on the record
The position this layer exists to test is that a weekly total is not something a function head can act on, and that the cost of one outcome is. more