One score, many voices.
An agentic operating system for Communications & Corporate Affairs in a large multinational. A Chief-of-Staff orchestrator, 26 task agents in six clusters, a seven-store truth substrate, five judgment gates and five end-to-end workflows — specified to the level of detail required to build it.
Measured autonomy · 98 steps across 5 workflows
78 of 98 discrete steps execute at A3 or A4 — no human in the execution path. The remaining 20 are concentrated at named judgment gates.
Live control room · Ceilings reduced one level on affected workflows. Gate deadlines compress to thirty minutes.
Posture control →Active situations
0
2 currently held at a gate
Gate queue depth
0
7 past deadline · escalate, never auto-approve
Agents in the register
0
six clusters, one output contract each
Canon writes logged
0
7 stores · 0 evidence items
Section 01
The one definition to fix before reading anything else
"80% autonomous" means at least 80% of discrete process steps execute at autonomy level A3 or A4 — without a human in the execution path — with all remaining human effort concentrated at named, pre-declared judgment gates. It does not mean 80% of hours. It does not mean 80% of headcount. It means the human is removed from the conveyor belt and placed at five specific decision points, each of which is a decision, not a review.
BCG’s Workforce Analytics Engine puts the Communications function at 26–36% task-level and 34–47% reshaped productivity opportunity, not 80%. Any program that lets the hours reading stand will fail its first board review.
14
Gate steps
carrying fifteen gate firings between them — one step in WF3 combines G2 and G5 under a single eight-minute deadline
3
Non-gate human decisions
engage-or-hold in WF1, declare-crisis in WF3, escalate-to-the-triad in WF4
3
A2 co-production steps
a human writes and the agent supports rather than the reverse
of workflow steps execute with no human in the execution path
78 of 98 steps
task agents, each with a single output contract
a named failure mode and an evaluation set
judgment gates where every remaining human hour is concentrated
a gate is a decision, not a review
Canon stores holding the machine-readable corporate truth
the only layer the company actually owns
Section 05 · the stack
Five layers, addressed top to bottom
Select a layer to see what it owns, what it is forbidden from doing, and where it lives in this application.
Every agent reads from L0; every completed action writes back to it. Task agents never talk to each other directly — all coordination is via the Chief of Staff.
Layer L1
Chief of Staff
Routes work, sequences it against the corporate calendar, enforces autonomy ceilings, calls the gates, escalates, and writes everything back to the Canon. The Chief of Staff does not write, does not publish and does not decide. It is air traffic control — it owns separation and sequencing, not the aircraft.
L1 · execution
Five workflows, end to end
Every step carries its autonomy level and its gate. These five carry the whole autonomy claim.
| Workflow | Steps | A4 | A3 | A2 | A1 | A0 | Autonomy |
|---|---|---|---|---|---|---|---|
| WF1Value dispute with the franchise system | 22 | 11 | 6 | 1 | 0 | 4 | 77.3% |
| WF2Global position cascade | 17 | 13 | 2 | 0 | 0 | 2 | 88.2% |
| WF3The golden hour | 23 | 14 | 6 | 0 | 0 | 3 | 87.0% |
| WF4The conduct issue | 18 | 8 | 3 | 2 | 0 | 5 | 61.1% |
| WF5Multi-jurisdiction policy campaign | 18 | 10 | 5 | 0 | 0 | 3 | 83.3% |
| Aggregate | 98 | 56 | 22 | 3 | 0 | 17 | 0.0% |
Counting rule · 78 of 98 steps execute at A3 or A4 — 79.6%. WF4 is the lowest at 61.1%, against an A2 ceiling set by the General Counsel. A step counts as autonomous only at A3 or A4 — no human in the execution path. 78 of 98 steps qualify.
L2 · Chief of Staff
Signal intake and next-best-action
Feed a raw signal to the orchestrator. It scores materiality against the six weighted drivers, selects a workflow, names the gates that will fire and drafts the ranked action set.
L1 · Chief of Staff · Intake
Classify an inbound signal
The orchestrator scores materiality, routes to a workflow, and names the gates that must fire. It does not write, publish or decide.
No brief yet. The output contract is a ranked action set with a null option, its drivers exposed, and a deadline — never an undifferentiated summary.
Direct access
The rest of the specification
Why it is built this way
The problem the orchestrator exists to solve
Four numbers on the size of the opportunity, and one sentence from a practitioner.
of the Comms task mix is primed for humans and AI working together — not for either alone
of Comms tasks are genuinely human-dominant. This is the true size of the counsel layer.
of the working day — about one day a week — goes to work AI could meaningfully improve
of consultant time in a global comms advisory firm sits in just five workflows
“If everybody builds a messaging agent, and if everybody builds a… then you've got this proliferation of agents that don't talk to each other and actually have the opposite effect to the one that we're trying to drive, which is consistency and coherency.”
Supports · The most important sentence in the evidence base, and the reason this platform is designed the way it is.