Value · Economics · Sequencing · Transfer
The step count, in financial units
Seventy-eight of ninety-eight steps run without a human. This page states that in the units finance uses: hours by role band, cost per situation resolved, reinvestment share, unit costs, sequencing and transfer.
HOURS RECOVERED
15,862 / yr
COST / SITUATION
-56.6%
REINVESTMENT
44.0% trailing
PLATFORM RUN
$300K / yr
Platform run cost · USD
0K / yr
Fixed. Whole roster.
Per-person cap · USD
0 / mo
Backstop, not primary control
Value & reinvestment ledger
The step count in hours, cost and reinvestment
Step autonomy describes the design. Hours, cost per resolved situation and reinvestment share are the financial reading.
Hours recovered · per year
0.0K h
73.5% of 21,568 instrumented hours
Cost / situation · after · USD
0.0K
-56.6% against $42,212 before
Reinvestment · trailing 3mo
0.0%
Cumulative 49.1%. Target 50%. Falling.
Agency spend displaced
0.00M
-45.3% of the external portfolio, one line up
79.6% of steps run without a human but only 73.5% of hours come out — a 6.0-point gap, because the steps a human keeps are the long ones. And the hours that do come out are not evenly distributed: 90% of analyst hours against 17% of executive hours. Senior attention concentrates rather than reduces, because the autonomy ladder keeps judgment steps with a human.
The translation nobody does out loud
Steps that run without a human
79.6%
78 of 98 specified steps across the five workflows. This is the architecture number, and it is the one everybody quotes.
Gap
6.0
points
Hours that actually come out
73.5%
15,862 of 21,568 annual human hours across 143 situations. This is the number that reaches a budget line.
The gap is not an error. Steps are not equal in weight: the ones a human keeps are the long ones — the gate reads, the market adaptation calls, the conversation with counsel. Automating four cheap steps and leaving one expensive one moves the step count a long way and the hour count much less. Any vendor quoting step autonomy as a saving is quoting the wrong unit.
Where the hours come from · by role band
Analyst
$85/h90%
of band hours recovered
Monitoring, first drafts, list building, deck assembly, status reporting.
Manager
$145/h72%
of band hours recovered
Editing, channel planning, stakeholder handling, agency briefing.
Director
$235/h47%
of band hours recovered
Position calls, market adaptation, escalation judgment, gate holding.
Executive
$420/h17%
of band hours recovered
The five gates. Irreversible decisions. The hours the platform protects rather than removes.
90% of analyst hours come out. 17% of executive hours come out. That asymmetry is the design working, not failing — the platform concentrates senior attention onto the five gates and the irreversible calls, and removes the assembly work underneath it. A program that recovered executive hours at analyst rates would be a program that had automated judgment, and this one is built specifically not to.
Per workflow · step autonomy against hours recovered
Steps without a human · 17 of 22 · ceiling A3
Hours out · 97h of 133h per situation
gap 4.3 points · 3,686 hours a year
Analyst
64 → 7
hours / situation
Manager
41 → 11
hours / situation
Director
22 → 13
hours / situation
Executive
6 → 5
hours / situation
The analyst hours were franchisee sentiment collation and evidence hunting. Both are now continuous and cited. Director hours barely move — someone still has to decide what we concede.
What this ledger does not cover
Function base
148,000h
Modeled annual task hours across the whole Communications and Corporate Affairs function.
Instrumented by the five
14.6%
21,568h of that base sits inside a specified workflow today.
Recovered against base
10.7%
15,862h out, worth $1.94M at loaded rates.
Specified addressable
26-36%
Task level. 34-47% if the function is reshaped, with 20-28% cost impact.
10.7% against a specified 26-36% band reads like a shortfall and is not one. Five workflows are not the function. The distance between what is instrumented (14.6%) and what is addressable is the roadmap, not the miss — and quoting the addressable band as though it were already banked is the single most common way these business cases are oversold.
The published analog
A global communications advisory firm whose plan rests on a review of 400K hours of time entries. 400,000 hours of time entries reviewed; roughly 300,000 hours out of a 1,162,000-hour base, about $12M incremental EBITDA and 1-2 margin points by 2029 — with roughly 50% of the recovered capacity deliberately refilled rather than banked.
A 1,162K-hour professional-services function with a different revenue model. The mapping to a company of this scale is not linear.
Section 27 · Unit economics
Budget on workflow completions, not on prompts
Agentic workflows consume multiples of a chat query. Every component below is a driver you can move.
Driver
Note
Driver
Note
Driver
Note
Driver
Note
Cost per active user / month
Modeled 2027 → 2030, by usage class
Cost controls
Budgets attach to situations, not to people
A runaway scanning loop surfaces as a situation over budget, with a name and an owner, within hours.
Every live situation — a value dispute, a policy filing, a crisis — carries a budget assigned at intake and consumed by the agents working it. A runaway scanning loop shows up as a situation over budget, with a name and an owner, within hours.
A monthly per-user cap in the region of $200 catches pathological individual usage without throttling the situations that legitimately need depth. Caps set low enough to bind on normal work produce users who route around the platform.
Continuous sub-five-minute monitoring on a 40-entity watchlist costs an order of magnitude more than hourly on the same list. Reserve continuous for the watchlist and for elevated or crisis posture.
A full hostile-panel simulation across sixteen audience classes is expensive by design. Gate it on materiality — below the floor, run a three-class screen instead.
The business case
The efficiency case and the coherence case
The efficiency case is countable. The coherence case is larger and unprovable in advance, so the case does not rest on it.
26–36% of communications task time is addressable at task level; 34–47% when workflows are reshaped rather than accelerated in place. Cost impact runs 14–18% at task level, 20–28% reshaped. In the closest published analog — a global communications advisory firm whose plan rests on a review of 400K hours of time entries — roughly 300,000 hours come out of a 1,162K-hour base, and the plan targets approximately $12M of incremental EBITDA and one to two points of margin by 2029 after choosing to refill about half the freed capacity.
This case is real, it is defensible, and it is the one the CFO will hear.
Harder to size, larger. Stakeholder simulation in an analogous industrial setting was scoped against protecting against >2% stock-price impact, on the observation that comparable sectors have seen 4–15% negative impact following large-scale adverse media events. One avoided misstep at a company of this scale exceeds the entire efficiency case.
It is also unprovable in advance, which is why it must never be the number the business case rests on.
Two-thirds of leaders report that GenAI enables genuinely new work in this function. Fewer than one in ten reinvest recovered capacity into building it.
If the reinvestment share is not a reported metric from month one, it will be zero by month twelve and the function will have bought a cost reduction and called it a transformation.
“This is not a productivity play… It’s about fundamental business transformation. The structures will be reimagined around agent employees as well as human employees.”
Section 28 · Thirty months
Thirty months, sensing first and drafting last
Each phase exits on conditions, not on dates.
Stand up the Narrative Canon, Evidence Ledger, Policy & Risk Register and Signal Lake. Deploy the six sensing agents (A1–A6) at A4, the reporting agent (F1) and the narrative-drift detector (B4). No production agents. No drafting. The function's first experience of the platform is that it finally knows what is happening and why it matters.
- The CCO brief is generated end to end with no manual assembly for 30 consecutive days.
- Materiality scoring agrees with human triage on a blind sample at a declared threshold.
- Evidence Ledger holds every claim in the current external narrative with an owner and an expiry.
Deploy the Chief of Staff orchestrator, all five gates, and just enough production capability to run one workflow completely. Recommendation: the global position cascade (WF2) — highest autonomy ceiling, lowest blast radius, and it forces the Canon change-cascade to work, which is the hardest technical problem in the platform. Do not start with crisis.
- One position change has propagated through every downstream artifact automatically, with the diff reviewed and accepted.
- Median G5 dwell under one hour in normal posture.
- Zero published artifacts citing a superseded Canon version.
The remaining production and engagement agents, the full stakeholder-simulation stack with calibration running from day one, and the crisis cluster. Crisis posture is rehearsed monthly against live-fire scenarios before it is trusted on a real incident. All five gates live. Autonomy ceilings raised only on evidence from the Gate Ledger.
- Three live-fire crisis rehearsals completed within the golden-hour deadlines.
- Simulator Brier score beating the naive baseline for two consecutive quarters.
- Gate load per executive flat or falling against rising output volume.
Second business unit or second sector via the swap protocol. Ceiling ratchet program driven by override data. Gate deletion program for gates approving above 95% unamended. Agent sunset program. The roster should get smaller in this phase, not larger.
From here the roster should shrink, not grow. Gates approving above 95% unamended are deleted; agents below their adoption threshold are sunset.
Starting with drafting. Drafting demos beautifully, gets funded instantly, and produces an agent that writes fluent copy against a narrative nobody has agreed and claims nobody has substantiated. Six months later the function has a faster way to publish things it has not decided. Sensing first is slower to applaud and is the only order that works — you cannot orchestrate coherence around a truth substrate that does not exist yet.
Section 30 · Transferability
Seven things swap. The architecture does not.
The sector vector here is a global quick-service restaurant operator. Moving to banking or pharma changes the contents of seven stores — not the layer model, the gates or the orchestrator.
| What swaps | Where it lives | Quick-service restaurant | Banking | Pharma |
|---|---|---|---|---|
The audience with a veto | Stakeholder Graph | Franchisees — ~95% of restaurants, a de facto veto on pricing and operations | Regulators and supervisors — a veto on products, capital and pace | Prescribers and payers — a veto on adoption regardless of the label |
The substantiation regime | Evidence Ledger + Policy Register | Advertising and franchise-disclosure law; nutrition and sourcing claims | Securities disclosure, fair-lending, forward-looking-statement rules | Promotional-claims regulation; off-label prohibition; adverse-event duty |
The mandatory legal gate | Policy & Risk Register | G3 on franchise-relations and value-claim exposure | G2 fires on every forward-looking statement, no exception path | G2 and G3 both mandatory on any efficacy or safety claim |
The localization boundary | Narrative Canon | ~114 countries, three operating segments, ~80 licensee markets | Jurisdictional legal-entity boundaries; what one entity may say about another | Label variation by market — the same molecule says different things legally |
The crisis archetypes | Precedent Library | Food safety, labor and conduct, pricing, franchisee dispute, supply chain | Outage, breach, conduct, credit event, run risk, sanctions | Safety signal, recall, trial failure, pricing, supply interruption |
The scanning surface | Signal Lake | Franchisee forums, review platforms, local news, food-safety trackers | Supervisory publications, analyst notes, credit and CDS markets, prediction markets | Clinical registries, regulatory dockets, patient communities, KOL channels |
The autonomy ceilings | Policy & Risk Register | A3 on most workflows, A2 on conduct, A1 on journalist pitches | A2 on anything touching disclosure; A3 elsewhere | A2 ceiling across the board on anything clinical; A3 on corporate |
Populate the Stakeholder Graph for the new sector; identify the veto-holding audience explicitly and name it in the Canon. If nobody can name it, stop — the deployment is not ready.
GC writes the substantiation regime and the mandatory gate rules into the Policy & Risk Register. This is a legal drafting exercise, not a configuration task.
Ingest 24–36 months of the company's own published output into the Voice Corpus; ingest the current external narrative into the Evidence Ledger with owners and expiries.
Re-point the Signal Lake at the sector's scanning surface; retire irrelevant sources rather than carrying them.
Seed the Precedent Library with the sector's crisis archetypes and the company's own last five years of incidents, including the ones it would prefer not to document.
Set the initial autonomy ceilings one level below where the evidence suggests, and ratchet up from Gate Ledger data. Never the reverse.
Re-calibrate the simulator against the new sector's audiences from zero. Prior calibration does not transfer and must not be inherited.
What does not transfer, ever: simulator calibration, precedent, voice, and the materiality weights. The twelve-stage NBA engine transfers unchanged; the weights inside stage 4 do not — business exposure matters more in pharma, stakeholder influence matters more where a regulator holds the veto, velocity matters more in consumer.