The argument
SG&A is not one function.
It is a collection of functions,
and a significant heartbeat of the company.
This is a model of how it runs.
Every company already does this work. more
The case
Why a company would move its functions onto a network
Corporate functions have been reorganized twice in thirty years and both times the unit of change was the transaction. more
Where this is heading
Three eras, and only the third one changes the function
Each era took the same corporate work and moved one thing about it. more
Centralize the transaction
Move the same work to a cheaper place.
Invoices, payroll runs and journal entries were lifted out of the countries and put into a handful of centers. more
Where it stops · The saving is bounded by the wage gap, and the wage gap closes. Nothing about how the work is decided changed.
Automate the keystrokes
Take the typing out of the work that was already moved.
Towers were drawn, service levels were signed, and scripts were pointed at the screens people were filling in. more
Where it stops · Automating a step does not remove the step. The exception queue grows in proportion to the automation, and it is staffed by people.
Run the function
Agents hold the position, write the rule, watch the outcome and run the work.
The unit of automation stops being the keystroke and becomes the sub-function. more
Where it stops · This is the era this platform models end to end. Modeling it is not running it, and the pages say so wherever a figure would imply otherwise.
The three eras below are an argument about where corporate functions are going, not a measurement. more
The operating metaphor
A function is a transit network, and throughput is the whole point
A subway does not go faster because the trains are faster. more
What maps to what
Five pieces, nothing metaphorical about any of them
Line
A workflow inside a sub-function
Station
A stage with an owning agent and a dwell commitment
Gate
A station the train cannot leave until a person signs
Interchange
A station where work crosses into another function
Depot platform
A station whose work runs in the shared transaction depot
Where the throughput goes
Timetable against actual, on the estate as it stands
262 run behind, and the volume they carry turns that into 1,767,391 lost hours a week.
Dwell across the whole network sums to 26,835 hours against 47,623 hours of commitment. Every station is inside its own commitment while 262 lines run behind theirs, which the estate does not reconcile.
Against 1,463,698 items a week entering the network. A function that cannot answer this question is running without a signal box.
Why the metaphor earns its place
Corporate functions do not lose time inside the steps. more
Drawing the estate as a network makes the gap a physical object. It becomes an interchange with an owner, a dwell it is held to, and a signal that goes amber when the work has been standing too long. 143 of those exist here, and every one of them is a place the old operating model had no name for.
The second thing the network buys is arithmetic. Once every line has a timetable you can subtract, and the answer is 1,767,391 hours a week on this estate. That number is only available because somebody drew the line first.
where the work is waitingDwell, timetable and actual hours are stored on the station and line records. more
What the platform actually covers
Strategy, policy, observation, the work, and the depot
This is the part that separates an agentic operation from an automation program. more
Agents that take positions, argue them, and get held to the outcome.
209 positions authored
Open the positions registerIt starts here, not at the invoice. An agent researches the options, writes the position with its trade-offs and its dissenting read, attaches the evidence, and puts a named human on the ratification. Most automation programs never touch this layer at all, which is why they only ever produce a cheaper version of the same function.
The written rules every other agent is bound by, with version and owner.
465 policy items under version
Open the policy corpusA position that nobody wrote down binds nobody. The policy layer turns the ratified position into versioned text with an owner, an effective date and an authority behind it, so that every other agent in the estate is executing against a rule somebody signed rather than a habit somebody remembers.
Continuous sensing of the outside world and of the agents themselves.
70 signals on the grid
Open signal intakeThen the function watches. Signals come off the outside world and off the agents themselves, and the question the layer exists to answer is whether the outcome the position predicted actually moved. This is the layer that closes the loop, and it is the one that most often finds that the position was wrong.
Doing the work: lines, stations, and live units moving through them.
475 lines · 2,748 stations
Open the live work floorOnly now does the work run, and it runs sub-function by sub-function rather than function by function. Each sub-function is a set of lines, each line is a set of stations, and each station has an owning agent, a dwell commitment and, where the risk warrants it, a gate that stops the train until a person signs.
Repeatable transaction processing run once for the whole enterprise.
12 towers · 900,433 transactions a week
Open the transaction depotThe repeatable part of every line drains into one depot that runs once for the whole company, instead of every function keeping its own version of the same posting, matching and settlement work. This is the layer that changes the unit cost, and it is the last one, not the first.
Cost and efficiency
What the depot does to the unit economics
The depot is the fifth layer and the only one that moves the cost line … more
Read the difference carefully
What the number is, and what it is not
$18.85m a week is the arithmetic difference between two fields on the same twelve records: what the depot costs to run, and what the same volume was authored as costing before it existed. It is a modeled delta, and it is shown here because the shape of it matters, not because anybody has banked it.
The honest way to read it is as a ratio rather than a total. 67.5% of the before cost comes out, and 21.0% of what remains is compute rather than labor. That second figure is the one that behaves differently from every previous era: it scales with volume instead of with headcount, and it is the first thing that moves when an agent is promoted to a higher autonomy level.
Headcount on the depot records moves from 3,642 to 2,219, a difference of 1,423 and 39.1% of the starting figure. Those are authored fields on a modeled estate. No person in any company has been affected by anything on this page.
23 of 655 benefit lines carry a verification, and the verifier is this platform checking baselines it also holds. The benefit ledger is the page that refuses to add the rest into a total.
open the benefit ledgerWhy one depot and not one each
851 of the 2,748 stations on the network run their work inside the depot, and 274 of the 475 lines feed one. That is the mechanism: the line stays inside its function, and only the repeatable station drains out.
A function keeps the judgment, the policy and the accountability. more
what a tower chargesBefore and after cost per transaction, and headcount before and after, are fields on the twelve depot records. more
The argument for moving
Five things that change, each with the count behind it
These are the reasons a multinational would put its functions on this operating model rather than buy another tool. more
The saving stops coming from the wage gap
$31.01 → $10.08 per transaction
Across 12 depot towers carrying 900,433 transactions a week, the modeled cost per transaction moves from $31.01 to $10.08. That is a change in how the transaction is produced, not in where it is produced, so it does not evaporate when a labor market catches up.
Unit costInference is a cost line, and it is on the same page
21.0% of the run cost
$1.90m of the $9.08m weekly run cost is inference. more
Autonomy economicsHandoffs become scheduled interchanges instead of emails
143 interchanges on the network
Work crosses a function boundary at 143 places in this estate. more
HandoffsControl stops being an audit and becomes a gate in the line
508 gates standing in 2,748 stations
A gate is a station the train cannot leave until the named approver signs. more
ControlsCapacity is released at the top of the function, not the bottom
209 positions and 465 policy items authored by agents
Authoring a position is most of the labor in a corporate function, and it is the part every previous era left alone. 209 positions and 465 policy items on this estate were drafted by agents for a human to ratify or refuse. more
Strategy positionsAnd the reason it has to be all five layers
Any one of these on its own produces a tool. more
the modeled enterpriseThe other side
What would have to be true before any of this is proof
Four things stand between the argument above and evidence. more
Not one of the 26 source systems is connected
Every figure on this page is read from the modeled estate. 0 systems of record are exchanging data and 0 of 19 write-back routes are open, so no conclusion this platform reaches has ever landed in the system that holds the record.
632 of 655 benefit lines are unproven
The economics section is the weakest part of the case, and it is deliberately placed after the operating model rather than before it. more
62 of 251 checks were signed from outside the authoring team
The rest were signed by the team that wrote the thing being checked, and none by anybody outside the company. more
The before figures were authored, not observed
Cost before, headcount before and the timetable each line is graded against are fields on the records, written by the team that built this. more
Actions
What is waiting on a person
This page makes a case. more
stations reporting inside their dwell while the lines they sit on run late
Every one of the 2,748 stations is recorded inside its own dwell commitment, and 262 of the 475 lines those stations sit on are recorded as running behind their timetable. more
benefit lines with nothing outside the platform proving it
23 of 655 lines carry a verification, and the verifier is this platform checking its own baselines. more
lines running behind their timetable
1,767,391 hours a week are lost across them, weighted by the volume each line carries. more
checks signed from inside the team
Of 251 verification checks on the record, 62 were signed by somebody outside the authoring team, which is 24.7%. None was signed by anybody outside the company at all.
source systems described and not one of them connected
19 write-back routes are designed and 0 are live, so nothing this platform decides has ever reached the system that holds the record. more
Operations
What this desk is allowed to start
A surface that only reports is not operable. more
Trigger and bound
This page starts nothing. more
Live observability
What the record shows right now
The five layers of the operating model, sized by how many records exist behind each one. more
Current distribution
1,231 records
Is policy and strategy coming to fruition
Whether the written intent is holding here
The model is complete. The migration has not started.
Not holding on the record
The case this page makes is that a corporate function is a network, that the network can be run by agents under a control tower, and that the economics change when the repeatable part drains into one depot instead of one per function. more
What this page is
This page argues a position. more