AI GBS
The shared transaction spine
Repeatable, rules-bound, high-volume work is lifted out of the functions and run once for the enterprise, across 12 service towers with one control set and one cost-per-transaction number. This page reads the spine as an instrument: what it is carrying, what it is costing, what it is asking a person to do, and whether the policies written above it actually reach it.
Service towers
12
14 of 14 functions feed them
Transactions / week
900,433
76% touchless, weighted
Cost to serve / wk
$9.1M
was $27.9M
Run-rate saving
$980.2M
annualized at current volume
Inference / wk
$1.9M
21% of cost to serve
Headcount on the spine
2,219 FTE
down from 3,642
Health
Four dimensions, each with a printed rule
A dimension is marked degraded because a stated condition is true, not because somebody judged it so. The weakest reading under each panel is named rather than averaged away.
Coverage
holding
Service towers running
12
Feeder sub-functions
120 of 154 · 77.9%
Functions draining into a tower
14 of 14
Towers with no ratified position behind them
0
34 sub-functions keep their repeatable work inside the function. Most are judgment-bound rather than rules-bound, which is a defensible reason not to centralize — but it is a reason nobody has written down per sub-function.
Unit economics
holding
Transactions per week
900,433
Cost to serve per week
$9.1M
Was, before the spine
$27.9M
Inference share of cost
21%
Every cost figure on this page is a modeled unit rate multiplied by a modeled volume. Nothing here has been reconciled to an accounts-payable ledger or a cloud invoice, so treat the $18.9M weekly delta as the size of the prize, not as booked saving.
Service commitments
degraded
Towers reporting attainment
12 of 12
Towers with a signed service level
4 of 12
Below 95% attainment
6
Breaches still open
15
8 of the 12 towers publish an attainment percentage without a signed service level behind it. A number with no counterparty and no penalty is a dashboard reading, not a commitment.
Concentration risk
degraded
Weighted touchless
76%
Towers fed by fewer than three functions
3
Headcount on the spine
2,219 FTE
Down from
3,642 FTE
A shared spine concentrates risk as fast as it concentrates saving. 3 towers rest on fewer than three feeder functions, so a single function changing its volume moves that tower's cost per transaction directly.
Operations
What each tower is carrying this week
Volume, touchless rate, unit cost against the pre-agent rate, the queues running right now, and which functions drain into it. Feeder count is the honest measure of how broad a tower's base is.
H2R · Hire-to-Retire Tower
Every repeatable people transaction in the enterprise: payroll, time, benefits, employee data and statutory filing.
Tx / wk
126,636
Touchless
70%
Cost / tx
$10.26
was $28.80
Of which inference
$2.13
Queues running now
S2P · Source-to-Pay Tower
Requisition through payment: purchase orders, receipting, invoice capture, three-way match, supplier master and the payment run.
Tx / wk
142,093
Touchless
80%
Cost / tx
$8.06
was $26.34
Of which inference
$1.62
Queues running now
O2C · Order-to-Cash Tower
Order validation through cash: billing, cash application, collections, disputes and deductions.
Tx / wk
107,549
Touchless
80%
Cost / tx
$8.52
was $27.61
Of which inference
$1.76
Queues running now
R2R · Record-to-Report Tower
Close and control: journals, reconciliations, intercompany, consolidation and the disclosure pack.
Tx / wk
87,512
Touchless
72%
Cost / tx
$13.09
was $37.12
Of which inference
$2.92
Queues running now
P2P · Plan-to-Perform Tower
The planning cycle as a transaction: budget load, forecast refresh, variance packs, headcount plans and the management reporting run.
Tx / wk
25,194
Touchless
79%
Cost / tx
$15.42
was $46.04
Of which inference
$2.73
Queues running now
M2D · Make-to-Deliver Tower
Supply execution: supply plans, order promising, inventory moves, freight tendering, customs paperwork and warehouse fulfillment.
Tx / wk
81,095
Touchless
79%
Cost / tx
$10.44
was $34.72
Of which inference
$2.03
Queues running now
ITS · IT Service Tower
Service desk, joiner-mover-leaver provisioning, device and license administration, and the change and release pipeline.
Tx / wk
85,277
Touchless
80%
Cost / tx
$8.19
was $28.86
Of which inference
$1.57
Queues running now
LGL · Legal Operations Tower
Contract intake through execution, obligation extraction, outside counsel billing, privacy requests and compliance filings.
Tx / wk
40,630
Touchless
68%
Cost / tx
$9.89
was $28.61
Of which inference
$2.41
Queues running now
A2L · Awareness-to-Lead Tower
Campaign build and launch, content production, list hygiene, lead scoring, routing and the qualification handoff to sales.
Tx / wk
51,156
Touchless
77%
Cost / tx
$11.94
was $35.84
Of which inference
$2.43
Queues running now
C2L · Concept-to-Launch Tower
Experiment intake, laboratory and pilot operations, invention capture, regulatory dossiers and the launch readiness pack.
Tx / wk
51,438
Touchless
73%
Cost / tx
$9.86
was $27.81
Of which inference
$1.88
Queues running now
AIO · AI Operations Tower
The estate that runs the estate: service catalog intake, automation build, model and agent operations, and continuous improvement.
Tx / wk
68,456
Touchless
79%
Cost / tx
$9.45
was $28.38
Of which inference
$2.29
Queues running now
D&A · Data and Analytics Tower
Master data stewardship, integration pipelines and the reporting layer every other tower reads from.
Tx / wk
33,397
Touchless
75%
Cost / tx
$14.04
was $45.29
Of which inference
$3.44
Queues running now
Actions
What the spine is asking a person to do
Ordered by size. Nothing in this list clears itself, and each row opens the surface where the work is actually held.
Items in a tower queue that need a person
evidence to read, holds to release, aged exceptions
Sub-functions with no tower route
repeatable work still run inside the function
Service level breaches still open
$15K in service credits not yet settled
Tower queues past their service window
exceptions aged beyond 48 hours
Towers running without a signed service level
attainment reported, nothing committed
Tower queues at risk of breaching
inside the window, trending out of it
AI GBS incidents open or on watch
actions outstanding or monitoring running
Live observability
What has actually gone wrong on the spine
Breaches against signed commitments, incidents raised inside AI GBS, and escalations an agent handed to a named person. These are the events the tower dashboards are built to hide.
Service level breaches
29 on record · 15 still open
The planning system returned partial result sets for two days without erroring, so the line processed an incomplete population and the shortfall was found at reconciliation.
missed by 6.5% · 2,408 units · source-system
The investigation was started and never finished. It is shown as unresolved because closing it without a cause would misrepresent the estate.
missed by 5.9% · 2,269 units · unknown
The planning system returned partial result sets for two days without erroring, so the line processed an incomplete population and the shortfall was found at reconciliation.
missed by 4.3% · 892 units · source-system
Intake quarantine held 340 items over the weekend because the release decision needs a named human and none was rostered.
missed by 8.0% · 1,995 units · data-quality
The planning system returned partial result sets for two days without erroring, so the line processed an incomplete population and the shortfall was found at reconciliation.
missed by 8.3% · 432 units · source-system
A drift alarm on the classification agent fired mid-month. Autonomy was reduced by the operator to supervised, and the queue lengthened as a direct consequence.
missed by 9.3% · 140 units · model-degradation
Incidents inside AI GBS
3 on record · 1 open or on watch
INC-GB023 · medium · 554 units affected
INC-GB024 · high · 3,464 units affected
INC-GB025 · medium · 990 units affected
Escalations on the spine
1 on record · 0 waiting on a person
ESC-GB155 · tier tier-3 · Owen Marchetti · Head of Automation
Is policy and strategy coming to fruition
Does the intent above the spine reach the transaction inside it
A tower executes. It does not decide. The question is whether the work it executes is governed by a position the leadership ratified and a policy written from that position — counted per feeder sub-function, where each step only counts if the step before it did.
Feeder sub-functions routed to a tower
repeatable work lifted out of the function
…carrying a ratified position
a stance the leadership signed, not a draft
…and an active policy reaching a live workflow
a rule with a version, an approver and something that runs
Chain by tower
how much of each tower's feeder base is governed all the way down
| Tower | Feeder subs | Ratified position | Active policy | Governed share | Signed SLA |
|---|---|---|---|---|---|
| H2RHire-to-Retire Tower | 13 | 3 | 3 | 3 | |
| S2PSource-to-Pay Tower | 16 | 8 | 8 | 3 | |
| O2COrder-to-Cash Tower | 15 | 4 | 4 | 3 | |
| R2RRecord-to-Report Tower | 9 | 2 | 2 | 3 | |
| P2PPlan-to-Perform Tower | 10 | 4 | 4 | none | |
| M2DMake-to-Deliver Tower | 8 | 3 | 3 | none | |
| ITSIT Service Tower | 11 | 3 | 3 | none | |
| LGLLegal Operations Tower | 7 | 1 | 1 | none | |
| A2LAwareness-to-Lead Tower | 7 | 2 | 2 | none | |
| C2LConcept-to-Launch Tower | 8 | 4 | 4 | none | |
| AIOAI Operations Tower | 8 | 5 | 5 | none | |
| D&AData and Analytics Tower | 8 | 1 | 1 | none |
40 of 120 feeder sub-functions carry the full chain — a ratified position, an active policy written from it, and a workflow that runs under that policy. The remaining 80 are executing transactions at scale under a general standard rather than a rule with a version and an approver. Fixing that is a governance task, not an engineering one.
What this page is, and what it is not
Every figure here is computed live from the platform's own tower records, which makes them exact and makes them narrow. The estate is modeled: no volume, unit cost or attainment reading has been reconciled against an enterprise resource planning system, a payroll register or a service management tool. A shared spine concentrates risk as much as it concentrates saving — one bad policy version in a tower touches every function feeding it, which is why each tower carries its own control set, rollback path and blast-radius number rather than inheriting a function's. That trade is deliberate: fewer failure modes, much better instrumented, instead of fourteen weakly-observed copies of the same process.