The three meters
- Seats (M365 Copilot per-user): predictable, visible, and wasted in silence — the meter where INACTIVITY is the cost problem. Discipline: the monthly recycling funnel (usage-analytics concept).
- Credits (agent/consumption billing): elastic, invisible until the invoice, and SUCCESS is the cost problem — a popular custom-engine agent reaching unseated users (taxonomy concept's licensing fact) scales spend with adoption. Discipline: per-agent burn dashboards + budget alerts at thresholds you set BEFORE launch (Studio concept), plus a unit-cost habit: credits per conversation per agent, tracked.
- Agent 365 (~$15/user/mo governance overlay): a step-function you opt into when fleet size makes identity-grade agent governance necessary — the trigger is the census crossing your ungovernable line, not a calendar date. Its ROI case is risk (one confused-deputy incident pays for years of it), which finance accepts when written that way.
The reconciliation view (build it once)
One page: seats (count × price × utilisation), credits (by agent, with trend), Agent 365 (population × price), against the value ledger (usage-analytics concept). This is the artifact that catches the classic drifts: seat spend flat while credit spend compounds quietly; a 'free' maker experiment becoming the biggest line; governance spend lagging the fleet it should cover.
Cost-shaping levers, ranked by effort
Seat recycling (cheap, monthly), agent targeting rings (deployment concept — don't org-wide a credit-burner on day one), per-agent budget caps/alerts where the platform offers them, model/knowledge scope tuning on expensive agents (narrower scopes retrieve less and spend less), and the lane decision itself (developer-paths: a declarative rebuild of an over-engineered custom engine is sometimes the biggest saving available).
What to watch (proofs)
- The reconciliation page, monthly, dated — three meters + value, one screen; the renewal negotiation starts from it.
- Credit anomalies: week-over-week per-agent burn deltas — the compound quietly is caught by the delta report, never by the invoice.
- Utilisation-weighted seat cost: price ÷ weekly-active rate per role family — the number that reallocates seats better than any survey.
- The Agent-365 trigger metric: governed vs ungoverned agent count — when the ungoverned line grows, the overlay's case writes itself.