The Variable Cost
of Copilot
Cowork went GA and consumption is metered in Copilot Credits. A task now costs model plus context plus tools plus runtime, not headcount times a number. This is the method to allocate, charge back, cap, and forecast the bill before it forecasts itself.
"Battle-tested in production environments."
Jared Spataro, Microsoft's CMO for AI at Work, citing the author's public Copilot prompt library on LinkedIn. The same standard runs through this book and kit: every figure dated, sourced, and labelled, so you always know what you can bank on.
Per-seat broke. A task is now four moving parts.
For two years Copilot was the easiest line in the stack to forecast: seats times a number. Once consumption is metered in Copilot Credits, the same person running the same kind of task can cost wildly different amounts. A task's cost is model plus context plus tools plus runtime, and none of those show up on a headcount report.
Model
Which engine the task routes to. A heavy reasoning model and a fast cheap one are not the same line item.
Driver 1Context
How much you feed it. Big grounded documents and long histories cost more per call.
Driver 2Tools
What the task reaches out to. Tool calls and agent actions add up, quietly.
Driver 3Runtime
How long it runs and how often. Recurring and scheduled work compounds where nobody is watching.
CompoundsFour drivers. One metered invoice. This is the FinOps of the consumption era, the part nobody hands you a dashboard for.
Allocate it. Charge it back. Cap it.
The book treats Copilot as what it now is, a consumption product, and walks the whole surface in the order you have to work it.
Map the cost surface, then pick an allocation base on purpose (per cost centre, per team, per use case) instead of letting the bill land wherever the licence happens to sit.
Run showback for a cycle or two, then flip to chargeback only when reconciliation holds. Bill on the base your CFO will defend; track on the base that explains the spikes.
Set the spending limit and the anomaly alert at the same time. A cap with no alert means you find out at the limit, not before it.
Cost controls are not governance controls
A spending cap stops spend. It does not stop oversharing, and it does not authorise an agent. Cost is one axis of a governance triangle: cost, data, agent. The book is explicit about where it stops, so you control the bill without mistaking a budget alert for a data-governance control. That honesty is the credibility.
A book, ten templates, and two calculators that actually compute.
~230 pages and a working operating kit. A method built in two layers, so it survives the next price change instead of going stale the week Microsoft reprices.
12 chapters, ~52,000 words
Why per-seat pricing broke, how a Copilot Credit accrues across model, context, tools, and runtime, and a working FinOps-for-AI method to allocate it, charge it back, cap it, and forecast it. PDF (tagged and bookmarked) plus an epubcheck-clean EPUB.
10 fillable Word templates (A01–A10)
Allocation model, chargeback policy, allocation-base decision matrix, cost-centre mapping, spending-limits policy, Cowork access-approval workflow with a security and data-governance gate, FinOps-for-AI operating model, budget and alert runbook, anomaly-response playbook with a war-room card, monthly showback statement.
2 working spreadsheets (B01–B02)
A TCO and ROI calculator and a consumption and forecast tracker. Both compute live, with built-in validation self-checks and zero formula errors. You change the inputs, they do the math.
Governance and onboarding
A START-HERE launchpad with a Choose-your-path guide, a RISKS refresh checklist, a claim-confidence register, a QA summary, and four executive one-pagers.
Every figure is labelled
The claim-confidence register marks each Microsoft claim so you always know what you can bank on.
- Verified · sourced to a first-party page
- Dated · a June-2026 snapshot, isolated and logged
- Assumption · a stated, defensible default
- UNKNOWN · flagged, not faked
Two working calculators
Live .xlsx models, not screenshots.
- TCO and ROI calculator with scenario inputs
- Consumption and forecast tracker, month by month
- Validation tab that checks its own formulas
Built two-layer, honest about its limits.
The FinOps method is durable. The dated facts are isolated and logged. The governance boundary is stated plainly. You are buying a method, not a snapshot.
The method does not date. When Microsoft moves a number, the product refreshes in minutes, not a rewrite.
The credit price and the Cowork GA position are logged in the claim register with a refresh checklist in the box.
A cap stops spend, not oversharing. Cost, data, and agent are three controls. The book never pretends a budget alert is governance.
The two spreadsheets compute live and self-check their own formulas. Change the inputs and watch the totals move.
A consumption bill you can allocate, defend, and forecast.
For the person who now owns the meter.
You are numerate, busy, and done with vendor savings claims you cannot reproduce. This gives you a model you build yourself. It assumes you know FinOps; it will not define showback or chargeback from first principles.
The companion: the per-seat baseline
This is the consumption-era half of a two-book line. The Real Cost of Copilot ($29) covers the per-seat baseline: what a seat truly costs once you count licences, adoption, and the hidden overheads. Read that to defend the seat decision; read this to govern the meter.
Stop forecasting a flat line. Start governing the meter.
- The full 12-chapter book, PDF and EPUB
- 10 fillable Word templates plus a combined reference PDF
- 2 working spreadsheets with built-in validation
- A method that survives the next pricing change