LensReading which lens this session carries.
Start here

Fourteen corporate functions, run by agents, under a control tower you can audit.

This is not a dashboard sitting on top of an operation. It is the operation. Every function is modeled to the sub-function, every sub-function to the workflow, every workflow to the stage, and every stage to the work unit standing in it right now. Agents hold positions, write the policy they are bound by, watch themselves, run the work, and drain the repeatable part into a shared transaction spine. What you see below is that estate as it stands today, with the parts that are not finished named on the same page as the parts that are.

The shape is deliberate. Corporate functions have been reorganized twice and both times the unit of change was the transaction, so both times the judgment stayed exactly where it was. Here the unit is the sub-function, the estate is drawn as a transit network so that every handoff is a station with an owner rather than an inbox with none, and the repeatable part of every line drains into one depot instead of fourteen. The case for that, with the live counts behind it, is on the page next door.

14
Functions
154
Sub-functions
1,361
Agents modeled
475
Workflows
200,240
Items per week
72.9%
Mean autonomy

The idea

What you are actually looking at

Three things separate this from an automation program. Read these before you read a single number.

Agents do judgment work, not only transactions

The comfortable version of this story is that agents handle the paperwork and people keep the thinking. That produces a cheaper version of today. Here an agent authors the position — the options, the trade-offs, the dissent and the evidence — and a named human ratifies or rejects it. Authoring the position is most of the labor, and that is the capacity this releases.

Agents are managed like staff, with the same instruments

Volume, quality, speed, fully loaded cost including inference, risk and trust trend — the same scorecard whether the worker is an agent or a team. An agent whose override rate climbs gets flagged, demoted an autonomy level and put on the equivalent of a performance plan. Promotion needs a passing evaluation on the record, not an opinion.

Every claim on screen has to survive being checked

A benefit is only counted when it lands against a baseline with a general ledger account behind it. An autonomy promotion is only counted when the evaluation run that justified it is attached. Where the evidence does not exist yet, the page says so in the same type size as the good news.

How a function is built

Five layers, and a function is only agentic when it has all five

Automating the transactions is the last layer, not the first. A function is agentic when it can hold a position, write the rule it is held to, watch itself, do the work, and push the repeatable part into a spine that runs once for the enterprise.

01

Strategy

Agents that take positions, argue them, and get held to the outcome.

209 positions authored
Open a function and read its layers
02

Policy

The written rules every other agent is bound by, with version and owner.

465 policy items under version
Open the authority register
03

Observe

Continuous sensing of the outside world and of the agents themselves.

70 signals on the grid
Open signal intake
04

Workflow

Doing the work: lines, stations, and live units moving through them.

475 lines · 2,520 units in flight
Open the live work floor
05

AI GBS

Repeatable transaction processing run once for the whole enterprise.

Twelve service towers on one spine
Open the transaction towers

The tour

How a single piece of work moves, end to end

Six stations. Follow one work unit through all of them and you have seen the whole platform. Each station links to the live page where that step is actually happening.

01

Something changes outside

A price moves, a regulator publishes, a supplier slips, a customer escalates. Observation agents watch named external and internal sources continuously and raise a signal with a confidence and a source, not a rumor.

Accountable · Observation cell, per function

Signal intake
02

A position gets authored, or a rule gets applied

If the signal needs a judgment, a strategy agent writes the position with alternatives and dissent and routes it to the executive who owns it. If an existing policy already covers it, the policy layer decides and records which version of which rule applied.

Accountable · Strategy cell · policy owner

Executive decisions
03

The work enters a line

Workflows are drawn as lines and stages as stations. A work unit is a train. You can see which station it is standing at, how long it has been there, what is behind it, and how far the delay spreads if it does not move.

Accountable · Sub-function orchestrator

Live network
04

An agent stops and hands the decision back

When a unit hits the edge of an agent's authority, it stops, states why, recommends an option, attaches the evidence and puts a clock on it. It arrives in a person's inbox as a decision, not as a ticket to investigate.

Accountable · Named approver, by authority grant

Decision inbox
05

The repeatable part drains into the transaction spine

High-volume, rules-based work does not get automated fourteen times inside fourteen functions. It drains into four shared towers — hire to retire, source to pay, order to cash, record to report — and rejoins the line afterwards.

Accountable · Tower control, AI GBS

AI GBS towers
06

The money is proven, or it is not counted

The saving only exists once it has been matched to a baseline cost with a ledger account behind it and signed off by the finance business partner for that function. Claims that have not cleared that bar are shown separately and are never added to the total.

Accountable · Finance business partner

Benefit ledger

For the operator

Your first hour, if you actually run this

Not a product tour. This is the sequence a chief executive, chief financial officer or chief operating officer would work through on the first morning, and the sequence to repeat every Monday afterwards.

0—10 min

See the estate and find the two functions that are off the pace

The function estate ranks all 14 functions by how much work agents close without a person. Mean autonomy is running at 72.9% and mean touchless throughput at 71.4%. The gap between a function's bar and its committed target is the only thing to look at first — a low number with a low target is a planning problem, a low number against a high target is an execution problem.

10—20 min

Find where work is physically stuck right now

There are 2,520 live work units in the estate and 35 open escalations where an agent stopped and asked for a person. Held units are listed with the blast radius behind them, so you can tell a stalled invoice apart from a stalled invoice that is holding up a quarter close.

20—30 min

Clear one decision that is genuinely yours

Open the inbox and take the oldest item that carries your name. Each one states what the agent recommends, what it is not allowed to do without you, what happens if you do nothing, and who else could have decided it. Deciding writes an evidence record with your identity and the authority you used.

30—40 min

Check the money before you repeat it to anyone

Start at the baseline, not the benefit. The baseline page shows which function costs have a ledger account and an owner behind them and which do not. The benefit ledger then shows only the value that has cleared that bar, with everything still unproven held in a separate column that never rolls into the total.

40—50 min

Check the seatbelt is actually fastened

The control tower is where autonomy is constrained: what each agent may do alone, what it must escalate, what stops it. 23 fairness alarms are currently unclosed and 6 platform components are a single point of failure. These are the numbers a general counsel or a regulator will open first, so open them first.

50—60 min

Decide what earns more autonomy next week

Autonomy is granted against evidence, not enthusiasm. 178 agents are inside the managed lifecycle with a stage, an evaluation history and an arming state for automatic demotion. Promote the ones that cleared the floor, freeze the ones that did not, and note that 1,183 agents in the estate are still outside that program.

Guided walkthrough

The flash demo, if you would rather be shown

A guided run through the argument: the three eras of corporate functions, the five layers an agent has to hold before it earns the transaction, the network the work moves through, the depot underneath it, the cost case, and a closing chapter on what none of it proves. The prose is authored. Every figure beside it is read from the same live counts as the rest of the platform.

14 chapters, about 6.6 minutes

Narrated by a synthesized voice and paced to it. Auto-advances, or mute the voice and step through it at your own pace with the arrow keys.

01

SG&A is not a function. It is a collection of many functions

Selling, General and Administrative, and no specificity behind any of it

02

Full transparency, or it is just another deck

The six things this platform lets you see

03

Reorganized every three years, and the work never moves

Why the argument is always about structure

04

Eighty percent, function by function

What the target actually means

05

It does not start at the invoice

Strategy, policy, observation, workflow, depot

06

A named desk behind every layer

Who is actually doing this

07

Policy is written, versioned and binding

Shaping the rule, not just following it

08

Is the strategy and the policy actually landing

The layer that grades the other four

09

The estate runs like a network

Lines, stations, gates and interchanges

10

Handoffs are the thing that breaks

Throughput across functions, not inside one

11

One depot for the whole company

End to end business services

12

The reason a board would sign this

Cost and efficiency

13

People move to where judgment matters

The decision queue

14

What it delivers, and how you build it

Core systems to outcomes, and the payoff

Start the demo

The rest of the estate

Where everything else lives

The left-hand navigation is ordered the way the work is: see it, run it, govern it, improve it, prove it. If you only remember five destinations, remember these.

What this platform is, and what it is not

This is a fully modeled operating estate, not a running enterprise. Every function, workflow, agent, control and figure on these pages was authored by the team that built the platform, and the behavior you see is the behavior we specified. That is enough to judge whether the operating model is the right one. It is not evidence that any of it has survived contact with a real ledger, a real regulator or a real works council. The list below is the honest version, and it is on this page on purpose.

No source system is connected

26 systems of record are described with their owner, auth mode and freshness commitment. None of them is exchanging data. Every figure downstream is therefore a modeled figure.

The evidence chain is not externally anchored

Evidence records are hashed and versioned inside the platform, but nothing is notarized to an outside authority, so the chain proves internal consistency rather than tamper resistance.

23 fairness alarms unclosed

Of 23 alarms raised against people-affecting decision points, 0 have been closed. Detection works. The closure discipline behind it does not exist yet.

16 of 28 service levels carry no consequence

A commitment with no credit, fee at risk or termination right attached is a statement of intent. They are shown as commitments, never as guarantees.

6 platform components are single points of failure

Recovery times published on the continuity page are measurements of a rehearsal on a modeled estate, not proof that a customer's production estate recovers in the same time.

1,183 agents sit outside the managed lifecycle

The lifecycle program covers 178 of 1,361 agents. The rest have no stage, no evaluation history and no automatic demotion arming.

No agent has been audited by a third party

Every evaluation ran against golden sets we wrote ourselves. A passing suite proves an agent behaves the way we specified, not that the specification is right.

The interface is English only

Policy text is being translated and legally reviewed jurisdiction by jurisdiction, but no interface string has been extracted or translated yet. Right-to-left layout is untested.

9 of 49 accessibility criteria evaluated

A static markup scan covers what markup alone can prove. Nothing requiring contrast measurement, keyboard walkthrough or assistive technology has been tested, and no audit has been passed.

Enterprise sign-on is built but not connected to a directory · pilot lifecycle excludes human resources and legal on purpose · benefit claims outside the ledger are shown separately and never added to the total