Cost at the unit · the price of the next point
Autonomy economics
The estate runs at 71.4% touchless. A weekly total cannot tell a function head whether point 72 is worth buying. This page prices it one sub-function at a time: what one percent of that sub-function's weekly volume costs to move from the person to the machine, what it saves once moved, what the machine then costs to run, and how long the build takes to pay for itself. The marginal cost of the next point of touchless is arithmetic on the unit economics each sub-function already publishes: the cost of a unit before automation, the cost of a unit after it, the volume, and the ceiling the function itself set. One point of touchless is one percent of weekly volume moving from the first to the second. The build cost is modeled, not quoted, and rises the closer a sub-function already sits to its own ceiling. The running cost of an automated unit is held at 1.35 times the sub-function’s published inference cost, because an automated unit calls the model on every case rather than only the ones a person escalated.
What the next point is worth, by verdict
Over the 78 sub-functions whose published cost pair can be decomposed. The other 76 are counted in the second column and nowhere else.
A build is called worth doing where the modeled build cost pays back inside twenty-six weeks, marginal where it pays back inside seventy-eight, and not worth doing beyond that or where the machine costs more to run than the person it replaces. At its ceiling means the sub-function has already reached the autonomy limit it set for itself, so there is no next point to buy without moving the ceiling first.
The twelve fastest to pay for themselves
Ranked by payback. Every row is one percentage point of that sub-function's own weekly volume, not one point of the estate.
| Sub-function | Now → ceiling | Units a point | Person / machine | Net a week | Build | Payback |
|---|---|---|---|---|---|---|
Field Service Dispatch Customer Service | 75% → 86% 11 points | 149 | $49.84 / $7.25 | $5,347 less $999 to run | $36k | 6.8 weeks Build it |
Supply Planning Supply Chain | 75% → 84% 9 points | 156 | $47.00 / $6.76 | $5,886 less $391 to run | $56k | 9.5 weeks Build it |
Code Review & Quality Engineering | 82% → 88% 6 points | 96 | $47.08 / $7.34 | $3,584 less $231 to run | $39k | 10.9 weeks Build it |
Learning & Development Human Resources | 73% → 82% 9 points | 178 | $51.63 / $7.25 | $6,790 less $1,111 to run | $74k | 10.9 weeks Build it |
O2C Control Tower AI GBS & Global Business Services | 84% → 92% 8 points | 230 | $35.47 / $4.86 | $6,672 less $368 to run | $73k | 10.9 weeks Build it |
Campaign Orchestration Marketing | 78% → 86% 8 points | 184 | $29.03 / $6.36 | $3,731 less $440 to run | $41k | 10.9 weeks Build it |
S2P Control Tower AI GBS & Global Business Services | 85% → 93% 8 points | 204 | $26.89 / $4.05 | $4,234 less $424 to run | $48k | 11.3 weeks Build it |
R2R Control Tower AI GBS & Global Business Services | 79% → 89% 10 points | 197 | $37.04 / $6.57 | $5,367 less $635 to run | $61k | 11.4 weeks Build it |
H2R Control Tower AI GBS & Global Business Services | 78% → 91% 13 points | 255 | $26.00 / $1.91 | $5,501 less $642 to run | $64k | 11.7 weeks Build it |
Demand Generation Marketing | 80% → 88% 8 points | 93 | $51.87 / $8.68 | $3,573 less $444 to run | $43k | 12.2 weeks Build it |
Warehouse & Fulfillment Supply Chain | 84% → 88% 4 points | 183 | $33.05 / $2.80 | $5,212 less $324 to run | $64k | 12.3 weeks Build it |
Data Science & Modeling Research & Development | 75% → 83% 8 points | 150 | $45.54 / $3.63 | $5,594 less $692 to run | $69k | 12.4 weeks Build it |
Where the headroom sits
Net weekly return of one point, summed over the sub-functions of each function that still have room.
76 sub-functions publish a pair that cannot be decomposed
Shown, and held out of every headline on this page.
What holds a ceiling down
Each sub-function states in prose why it will not go further. These are those statements sorted into five kinds.
This sorting is keyword matching over the rationale each sub-function wrote, not a maintained taxonomy, and it shows: risk appetite absorbs 68% of the estate because it is where anything unmatched lands. The distinction matters for who can lift the ceiling — a regulator, a controller and a data owner are three different conversations — so the sorting is worth doing properly and is not yet.
What an exception costs
The same service, priced two ways: chains that ran clean, and chains that raised at least one exception. Sampled week of 10 to 16 August 2026.
| Service | Clean | With an exception | Multiple | Excess a week | Most common cause | Payback |
|---|---|---|---|---|---|---|
Requisition to receipt Measured | $19.26 316 chains · 24.7 h | $48.01 44 chains · 53.7 h | 2.49× | $63,259 $1,265 in the sample | S2P-CAT · 22 The dominant code is S2P-CAT, on 22 of the 44 sampled chains that raised anything at all. | 2.5 weeks $94k to build |
Invoice to payment Measured | $23.52 251 chains · 40.2 h | $61.38 49 chains · 95.9 h | 2.61× | $92,755 $1,855 in the sample | R2R-UNR · 20 The dominant code is R2R-UNR, on 20 of the 49 sampled chains that raised anything at all. | 2.4 weeks $132k to build |
Order to cash Too few exception chains to measure | $12.91 187 chains · 7.2 h | $45.12 13 chains · 26.4 h | 3.49× | — $20,936 if the mean held | O2C-ATP · 4 The dominant code is O2C-ATP, on 4 of the 13 sampled chains that raised anything at all. | — not costed |
Deduction to settlement Too few exception chains to measure | $11.68 153 chains · 28.7 h | $55.27 17 chains · 209.3 h | 4.73× | — $37,047 if the mean held | O2C-PRF · 7 The dominant code is O2C-PRF, on 7 of the 17 sampled chains that raised anything at all. | — not costed |
Hire to productive Measured | $26.57 148 chains · 89.9 h | $52.45 42 chains · 137.3 h | 1.97× | $54,350 $1,087 in the sample | H2R-CTR · 12 The dominant code is H2R-CTR, on 12 of the 42 sampled chains that raised anything at all. | 5.4 weeks $176k to build |
Close to report Measured | $55.48 98 chains · 85.0 h | $95.79 67 chains · 127.5 h | 1.73× | $135,051 $2,701 in the sample | R2R-EST · 19 The dominant code is R2R-EST, on 19 of the 67 sampled chains that raised anything at all. | 1.5 weeks $125k to build |
Query to resolution Measured | $23.29 161 chains · 60.1 h | $59.53 49 chains · 97.6 h | 2.56× | $88,786 $1,776 in the sample | ITS-P1E · 13 The dominant code is ITS-P1E, on 13 of the 49 sampled chains that raised anything at all. | 1.7 weeks $91k to build |
Concept to shelf Measured | $54.09 115 chains · 139.1 h | $85.13 40 chains · 264.2 h | 1.57× | $62,080 $1,242 in the sample | C2L-SPC · 11 The dominant code is C2L-SPC, on 11 of the 40 sampled chains that raised anything at all. | 2.7 weeks $101k to build |
The widest gap in the sample is Deduction to settlement at 4.73 times, and it is one of the 2 services with too few exception chains to call it a measurement: it raised 17 against a floor of twenty. The widest gap that can be measured is Invoice to payment at 2.61 times, on 49 exception chains against 251 clean ones. The most expensive is Close to report at $135,051 a week of avoidable cost, and it is also the fastest to build against, at 1.5 weeks on a modeled build of $125k. The candidate on every measured row is named, so the argument is about one agent and one code rather than about exceptions in general.
A clean chain and an exception chain of the same service are priced with the same rate card, so the multiple between them is a difference in the work, not in the pricing. A service is only marked measured where at least twenty chains of each kind exist in the sampled week; below that the mean is a curiosity rather than a measurement, and the page says so on the row. The weekly excess scales the sampled difference by fifty and assumes an agent removes 60 percent of it, which is an assumption and is stated as one.
Actions
What is waiting on a person
Each count below is a sub-function whose next point of touchless has a price and a return, or a service where the exception path costs a multiple of the clean one. Every figure is arithmetic on unit economics the estate already publishes.
Correct a published cost pair that cannot be decomposed
The cost before and after automation these sub-functions publish implies an automated unit cheaper than the model call it makes. They are shown and held out of every headline.
Fund the next point of touchless where it pays back inside two quarters
Each of these returns its build cost in 26 weeks or less on the sub-function’s own published unit economics.
Put an agent on the exception path of a measured service
The fastest is Close to report, 1.5 weeks against a modeled build of $125k.
Retire a ceiling that no longer earns its place
The next point of touchless costs more than it saves here, or pays back beyond 78 weeks. That is a reason to revisit the constraint, not to spend.
Sample more chains before pricing an exception path
Fewer than twenty chains of one kind in the sampled week, so the multiple on the row is a curiosity rather than a measurement. The widest multiple in the whole set, 4.73 times on Deduction to settlement, is one of them.
Operations
What this desk is allowed to start
A surface that only reports is not operable. This is the work this page can set in motion, and the bound it runs into.
Trigger and bound
This page can price the next percentage point of touchless for every sub-function, separate the ones whose published economics can carry the arithmetic from the ones that cannot, and set the fully loaded cost of an exception path beside the clean path for the same service. It cannot raise an autonomy ceiling, commission a build, or change a constraint. Ceilings are set on the autonomy pages by the function that owns the risk.
Live observability
What the record shows right now
The next point of touchless for every sub-function whose published economics can be decomposed, an exhaustive partition of all 78 of them into 4 verdicts: each carries exactly one. The 76 sub-functions whose published cost pair cannot be decomposed are excluded here on purpose.
Current distribution
78 sub-functions
Is policy and strategy coming to fruition
Whether the written intent is holding here
The next point of touchless across 73 sub-functions costs $4.5m to build and returns $188k a week net, a payback of 24 weeks.
Holding on the record
The position this layer exists to test is that autonomy should be bought a point at a time on evidence, not committed to as a program. On the arithmetic it holds. The estate runs at 900,433 transactions a week across its service towers, and one more point of touchless in each of the 73 sub-functions that have headroom and decomposable economics saves $210k a week, costs $23k a week more to run, and nets $188k against $4.5m of modeled build. 32 of the 78 decomposable sub-functions pay back inside 26 weeks, 35 between 26 and 78, and 6 should not be built at any horizon. The exception path says the same thing louder: across 8 services an exception costs between 1.57 and 4.73 times a clean chain, $554k a week of excess, of which $496k sits on the 6 services with enough chains of both kinds to measure. What does not hold is the evidence underneath. 76 of 154 sub-functions publish a cost pair that cannot be split at all, 2 of 8 services are too thin to measure, and 105 of the 154 ceilings resolve to risk because the constraint kind is keyword matching on a rationale sentence rather than a maintained taxonomy. The economics are actionable. The classification behind them is not yet a control.