Eighty people in twenty-four roles pass priced artefacts around eleven functions — senior management included. Every artefact costs effort and diary time. Tune the machine, then let AI absorb it role by role — and watch the money.
flow balance λ = (I−Pᵀ)⁻¹·a · ρ = load ÷ capacity
| Role | λ/wk | ρ |
|---|
| Artefact | Human £ | AI £ | Human elapsed | AI elapsed |
|---|
Structure. 24 role groups across 11 functions; 29 work stations pass ~20 artefact types role-to-role.
Steady state solves flow balance λ = (I−Pᵀ)⁻¹·a; a role's utilisation is its total load
(effort + its share of meeting time) against active headcount. The animation runs the same equations as a fluid simulation.
Time. Every artefact has hands-on effort (person-days) and a coordination delay — elapsed days spent aligning diaries across functions, of which ~20% consumes the owner's capacity as meetings. Lead time therefore decomposes into queue waiting + coordinating + working, shown live. AI-run roles cut coordination to ~8% (no diaries) and work queue items in parallel — a backlog at an AI station is cleared in hours at £-per-item, not weeks.
Money. UK salary ranges per role (midpoint used), ×1.28 employer on-costs. AI cost = £1,200/wk platform floor + £250/wk per AI-run role (hosting, evals, guardrails) + a per-artefact execution cost (e.g. build ≈ £28, design ≈ £15, incident fix ≈ £2.50) scaled by the AI unit-cost slider. Savings tick up as released payroll minus AI spend.
Absorption & consolidation. AI takes the hottest eligible role; human review stations above it keep working (senior engineers still review AI code) until their own turn. Leads become eligible once their function's craft roles are absorbed; completed clusters collapse to a single steward — one person running what was several functions' leadership (Design & Assurance · Engineering · Platform · Flow & Governance). Product Management is never automated.
Shocks. The human column re-sizes the original 76-person org to the new demand at 85% target utilisation: recruitment at 20% of salary and ~22 weeks to productivity (14 to hire, 8 to ramp), redundancy at ~3 months' salary with an 8-week consultation; backlog accrues during the gap. The AI column reprices compute at the new arrival rates — effective same day.
Presets & cadence. Three starting organisations (≈14, 76 and 300 people) apportion the same role structure at different scales — small shops run fractional "hats" (½ FTE roles). Demand accepts sub-weekly cadences (1 per fortnight, month or quarter) and governance volume scales with delivery cadence, so summary reporting slows when the machine does. Salaries are editable per role (±£5k on the midpoint), and the whole organisation — role counts and salary ranges — downloads and re-loads as CSV (with an always-available copy/paste window for viewers that block file downloads) so you can maintain it in a spreadsheet. The AI rollout's default strategy is a marginal analysis: every wave, each eligible role is hypothetically absorbed, the network re-solved, and the winner is whichever most improves predicted lead time and weekly run cost together — the most effective route to total absorption, not just the busiest role. Alternatives: most-overworked, biggest backlog, most-expensive-to-run, or a combined score; one to three roles per wave. The PMO is coupled to the value chain three defensible ways: coordination — the PMO runs the diaries, so every project hand-off's coordination time scales with PMO health (healthy ≈ ×0.9, starved ×1.3, AI-run ×0.55); interrupt shielding — a starved PMO leaks ~12% more incidents into the machine; and gating — on by default: funding and release-assurance sign-offs sit on the project path, so a starved PMO queues projects, not just packs (switch it off to model a PMO without gate authority). A fourth effect is real but not animated: a PMO that smooths starts reduces arrival variability, and queues grow with burstiness as well as utilisation (Kingman), so treat the lead-time picture as conservative. Reporting volume itself is still never counted as value — that claim stays out of the model on purpose. While the machine runs pre-AI, an advisory offers the old playbook whenever roles run over capacity: hire the computed gap (priced with recruitment fees and recurring payroll) or raise the standing portfolio kill rate — the gate then culls more of everything that reaches it, doomed work first, viable-but-marginal work next — so the AI rollout reads as the third option, not the only one. Senior management closes the sight loop: board packs flow up to steering, and decisions flow down. A fixed 12% of demand is doomed work; how much dies cheaply depends on the kill chain — flagged at design review, recommended by the PMO, cancelled by senior management — driven by management sight (25% when the PMO is starved, ~80% healthy, 95% AI-run) and management availability — the rest ships blind as waste, split out from value in the shipped numbers. Senior managers are automatable knowledge workers here like everyone else; the two directors are not — UK law and regulators require accountable humans, and the machine cannot want. Deciding automates; owning does not. Default salaries are blended-UK 2026 base pay with London weighting; the ×1.28 on-cost covers employer NI, pension and benefits on top.
All parameters are illustrative defaults, not benchmarks — move them.
Pick a preset, then make it bespoke — role counts, salaries, demand cadence, process weights. Choosing a preset after starting returns you to setup.
| Role | n | salary | ρ |
|---|
Copy this into a spreadsheet to edit, or paste an edited version here and apply. Keep the role_id column intact; salaries accept £ or £k.
Delivery is visibly faster and cheaper. What does demand do?
| Function | Was | Kept | Freed | Payroll released /yr |
|---|