Prove · item 22

Cost attribution and control

An agent estate that cannot say which budget paid for a token is not being managed, it is being subsidized. Every dollar of inference, platform runtime, human review and storage in the open quarter is attributed to a consuming cost center and shown back to the person who owns that budget — with the basis of the attribution stated, because a metered number and an allocated guess are not the same evidence.

Benefit ledger

Health

What the estate costs to run

Spend by how it was arrived at, attribution coverage and the budget caps under pressure.

A benefit realization ledger reconciled to a general ledger needs the customer’s actual baseline figures and account structure. Modeled numbers demonstrate the instrument and prove nothing about the business case.

Run cost, open quarter

$69.12m

54 attributed spend records

Inference

$21.98m

32% of run cost

Attributed to a consumer

96%

$3.07m sits in a shared pool

Metered, not estimated

33%

56% is an estimate with no underlying record

Budgets over cap

4 past their alert threshold

Cost centers with no cap

12 caps cover the rest

What the money buys

Open quarter by category

Human review$38.76m

56% of the quarter

Inference$21.98m

32% of the quarter

Platform runtime$7.47m

11% of the quarter

Storage and egress$901.2k

1% of the quarter

Human review is carried here as a cost of running agents rather than as business-as-usual headcount. An estate that only counts tokens understates what it costs to keep a person in the loop.

How the number was arrived at

The honest weak point of this page

Estimated56% · $38.76m

No underlying record. Never used in a benefit claim.

Metered33% · $22.88m

Measured at the point of consumption.

Apportioned11% · $7.47m

Split by a driver that is a proxy, not a measurement.

56% of the quarter is an estimate. Showback built mostly on estimates will be argued with by every cost center that receives it, and it should be. Moving spend from estimated to metered is the work this page is measuring.

Five-quarter run cost

What the estate has cost as it scaled

FY25 Q3$62.96m

review $38.44m · inference $17.75m · platform $6.04m

FY25 Q4$74.62m

review $39.01m · inference $25.79m · platform $8.77m

FY26 Q1$84.84m

review $39.31m · inference $32.97m · platform $11.21m

FY26 Q2$92.80m

review $38.51m · inference $39.31m · platform $13.37m

FY26 Q3 · open$69.12m

review $38.76m · inference $21.98m · platform $7.47m

Showback by cost center

What each budget owner is being charged this quarter, and whether a cap is watching it

Cost center	Scope	Owner	Inference	Platform	Review	Total	Metered	Cap
CC-3010Engineering	Function	Daniel Okoye	$1.98m	$479.0k	$2.98m	$5.44m	36%	no cap
CC-2010Legal	Function	Sofia Lindqvist	$1.53m	$369.4k	$2.85m	$4.75m	32%	Within budget
CC-1020Marketing	Function	Ravi Menon	$1.67m	$403.0k	$2.67m	$4.75m	35%	Past alert threshold
CC-2030Supply Chain	Function	Sofia Lindqvist	$1.35m	$326.6k	$2.94m	$4.62m	29%	no cap
CC-3020Research and Development	Function	Daniel Okoye	$1.51m	$364.3k	$2.69m	$4.57m	33%	Over budget
CC-5030AI GBS, function line	Function	Sofia Lindqvist	$710.0k	$171.4k	$3.29m	$4.17m	17%	no cap
CC-5020Customer Service	Function	Ravi Menon	$1.26m	$305.3k	$2.30m	$3.87m	33%	Past alert threshold
CC-4010Finance	Function	Marta Alvarez	$776.8k	$187.5k	$2.89m	$3.86m	20%	no cap
CC-4030Information Technology	Function	Daniel Okoye	$1.16m	$280.3k	$2.39m	$3.83m	30%	Within budget
CC-5010Administration	Function	Marta Alvarez	$569.8k	$137.5k	$3.04m	$3.75m	15%	no cap
CC-1030Revenue Operations	Function	Ravi Menon	$176.3k	$42.6k	$3.11m	$3.33m	5%	no cap
CC-6010GBS Tower, Hire to Retire	GBS tower	Sofia Lindqvist	$2.58m	$621.7k	$0	$3.20m	81%	Past alert threshold
CC-6040GBS Tower, Record to Report	GBS tower	Marta Alvarez	$2.50m	$603.6k	$0	$3.10m	81%	Within budget
CC-4020Human Resources	Function	Marta Alvarez	$39.3k	$9.5k	$3.02m	$3.07m	1%	no cap
CC-1010Sales	Function	Ravi Menon	$482.5k	$116.5k	$2.07m	$2.67m	18%	Within budget
CC-2020Procurement	Function	Sofia Lindqvist	$66.4k	$16.0k	$2.52m	$2.60m	3%	no cap
CC-6020GBS Tower, Source to Pay	GBS tower	Sofia Lindqvist	$1.98m	$478.7k	$0	$2.46m	81%	Within budget
CC-6030GBS Tower, Order to Cash	GBS tower	Sofia Lindqvist	$1.63m	$393.3k	$0	$2.02m	81%	Over budget
CC-9020Data and Integration, shared	Shared platform	Daniel Okoye	$0	$597.9k	$0	$1.50m	60%	Past alert threshold
CC-9010Agent Platform, shared	Shared platform	Daniel Okoye	$0	$1.20m	$0	$1.20m	0%	Over budget
CC-9030Trust and Assurance, shared	Shared platform	Ana Duarte	$0	$373.7k	$0	$373.7k	0%	no cap

9 of 21 cost centers have no budget cap at all — Revenue Operations · Procurement · Supply Chain · Engineering · Finance · Human Resources · Administration · AI GBS, function line · Trust and Assurance, shared. Spend in those centers is visible and unbounded, which is a control gap rather than a deliberate allowance.

Budget caps and rate limits

12 caps · 3 over budget · 4 past the alert threshold · 3 that stop nothing

Order to Cash tower, inference104% · Over budget

$2.02m of $1.94m · FY26 Q3 · alert at 80% · 4,200 model calls per minute, burst 6,000

hard capRequests queue rather than fail. Queue depth is published to the tower dashboard and the tower lead is paged at the cap, not after it.

owner Sofia Lindqvist · 4 breaches on record · last alert 11 days ago

Source to Pay tower, inference65% · Within budget

$2.46m of $3.79m · FY26 Q3 · alert at 80% · 2,800 model calls per minute, burst 3,600

hard capRequests queue. Invoice matching falls back to the deterministic matcher, which is slower and has a lower match rate.

owner Sofia Lindqvist · 0 breaches on record · last alert never

Hire to Retire tower, inference87% · Past alert threshold

$3.20m of $3.69m · FY26 Q3 · alert at 85% · 1,400 model calls per minute

hard capRequests queue. Candidate screening stops and the queue is worked by hand until the cap resets.

owner Sofia Lindqvist · 1 breach on record · last alert 9 days ago

Record to Report tower, inference67% · Within budget

$3.10m of $4.66m · FY26 Q3 · alert at 75% · 1,900 model calls per minute

hard capClose-period work is exempt from the cap by standing exception. Outside close, requests queue.

owner Marta Alvarez · 0 breaches on record · last alert never

Agent platform, shared runtime108% · Over budget

$1.20m of $1.10m · FY26 Q3 · alert at 80%

soft capSoft cap only. Spend above the cap is allowed and reported; nothing stops. This is a reporting control, not a spending control.

owner Daniel Okoye · 3 breaches on record · last alert 4 days ago

Data and integration, shared90% · Past alert threshold

$1.50m of $1.66m · FY26 Q3 · alert at 90%

soft capSoft cap only. Sync frequency is reduced by hand if spend runs ahead, which has happened twice.

owner Daniel Okoye · 2 breaches on record · last alert yesterday

Sales, function inference72% · Within budget

$2.67m of $3.70m · FY26 Q3 · alert at 80% · 900 model calls per minute

hard capDraft generation stops. Pipeline scoring continues because it runs on a cached model.

owner Ravi Menon · 0 breaches on record · last alert never

Marketing, function inference74% · Past alert threshold

$4.75m of $6.39m · FY26 Q3 · alert at 70% · 1,100 model calls per minute

hard capContent generation stops. Campaign measurement is unaffected.

owner Ravi Menon · 1 breach on record · last alert 9 days ago

Legal, function inference79% · Within budget

$4.75m of $6.02m · FY26 Q3 · alert at 85% · 600 model calls per minute

hard capClause review queues. Nothing is auto-approved when the cap is reached, by design.

owner Sofia Lindqvist · 0 breaches on record · last alert never

Research and Development, function inference102% · Over budget

$4.57m of $4.48m · FY26 Q3 · alert at 75%

soft capSoft cap. R and D spend is exempt from hard capping by standing exception, which has not been reviewed since it was granted.

owner Daniel Okoye · 2 breaches on record · last alert 11 days ago

Information Technology, function inference69% · Within budget

$3.83m of $5.53m · FY26 Q3 · alert at 80% · 1,500 model calls per minute

hard capTicket triage queues; incident work is exempt and bypasses the cap.

owner Daniel Okoye · 0 breaches on record · last alert never

Customer Service, function inference91% · Past alert threshold

$3.87m of $4.24m · FY26 Q3 · alert at 90% · 3,300 model calls per minute, burst 4,500

hard capDeflection stops and contacts route to the human queue. Service level is expected to breach within eleven minutes of the cap.

owner Ravi Menon · 1 breach on record · last alert 9 days ago

3 of the 12 caps are soft: reaching them raises an alert and stops nothing. They are labeled soft on the card rather than being presented as a control, because a limit that cannot refuse work is a notification.

Actions

What is waiting on a person

Cost to serve is only actionable where the spend is attributed to something somebody owns.

Replace an estimate with a meter

Estimated basis. The figure moves when the meter arrives.

Attribute an unallocated spend record

Real spend with no owning function or service.

Act before a cap breaches

Inside the cap but on the warning threshold.

Answer for a breached budget cap

Spend above the cap written into the record.

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 desk can meter consumption, apportion shared cost by the declared driver, and publish a cost-to-serve figure. It cannot approve spend, raise a cap, or stop a workload — a hard cap on this estate is enforced by the platform that runs the workload, not by this page.

Live observability

What the record shows right now

How each spend record was arrived at. Metered is observed; apportioned and estimated are calculated.

Current distribution

270 records

Metered9535%

Apportioned10539%

Estimated7026%

Is policy and strategy coming to fruition

Whether the written intent is holding here

35.2% of spend records are metered. 3 budget caps are breached.

Not holding on the record

The written intent is a real-time cost to serve for every service this platform oversees. It is partly there: 95 of 270 records are metered, 105 are apportioned by rule and 70 are estimates. 20 records are not attributed to any owner at all. Against 12 budget caps, 3 are breached and 4 are on warning. A cost figure built from apportionment is a defensible allocation, not an observed cost, and it is shown that way wherever it appears.

What this page is, and what it is not

The attribution model, the basis taxonomy, the showback table and the cap-and-alert mechanism are real. The amounts are modeled, and no live metering feed is connected — the metered share shown here is what a connected estate would look like, not a reading taken from one.

56% of the quarter is estimated rather than measured. That is the weakest number on this page and it is stated in its own tile rather than buried, because a showback report is only as strong as the share of it somebody can dispute with a meter reading.

3 budgets are over cap right now and 3 of the caps in force cannot stop anything. 9 cost centers have no cap at all. None of those three facts are presented as satisfied controls.