Copilot licensing generates more confusion per dollar than anything else in the Microsoft stack, and the confusion is expensive in both directions: orgs that license everyone and watch half the seats go idle, and orgs so spooked by the price tag that they give nobody anything. Both are math failures. Here’s the math.
The two-tier model
As of mid-2026, there are exactly two ways a person in your tenant gets Copilot:
Tier 1: Microsoft 365 Copilot Chat — included with your M365 subscription. Free, web-grounded chat for everyone. It can reason, draft, summarize pasted content, and search the public web. What it does not include by default: grounding on your tenant data, and it doesn’t put Copilot inside Word, Excel, Outlook, PowerPoint, or Teams. Work-grounded queries and agents that touch tenant data are available in Chat — but metered, billed in Copilot Credits.
Tier 2: The Microsoft 365 Copilot add-on license — per user, per month. This unlocks embedded Copilot in the apps (Word, Excel, Outlook, Teams, PowerPoint), work grounding across the Microsoft Graph, plus the Researcher and Analyst agents. Licensed users also use tenant-grounded agents without consuming credits.
The mental model: the license is an all-you-can-eat pass; credits are the à la carte menu. Everything in tier 2 can be approximated for tier-1 users through metered consumption — except the embedded in-app experiences, which are license-only. If someone’s value from Copilot lives inside Word and Outlook, no amount of credits substitutes.
Copilot Credits: the meter explained
Credits are the unit for everything pay-as-you-go. Two ways to buy them:
- Prepaid capacity packs — 25,000 credits per month, per pack. Predictable budget line, use-it-or-lose-it each month.
- Azure pay-as-you-go — billed per credit consumed, no commitment. You can pair this with packs as overage protection, which is the configuration most orgs should run: packs for the baseline, PAYG for spikes.
What consumes credits varies enormously by operation. The rough hierarchy (check current published consumption rates before budgeting — Microsoft adjusts these):
| Operation | Relative credit cost |
|---|---|
| Simple scripted agent response | Cheapest — single credit territory |
| Generative answer, no tenant grounding | Cheap |
| Answer grounded on tenant data (SharePoint, Graph connectors) | Expensive — tens of credits per answer |
| Autonomous agent actions | Expensive |
That middle distinction is the whole game. A declarative agent grounded only on its instructions and the public web costs nothing — for anyone, licensed or not. The moment an agent reads SharePoint or a Graph connector, every answer it gives to an unlicensed user burns credits at the expensive rate. Licensed users? Free — it’s covered by their license.
The licensed-vs-metered rule in one sentence: tenant-data agents are free for licensed users and metered for everyone else. This single rule should shape both your license list and your agent designs.
A worked example
A 200-person professional services firm. Assume the add-on at roughly $30/user/month list and tenant-grounded answers at roughly 30 credits each (verify current rates; the structure of this math survives price changes).
Option A: license everyone. 200 × $30 = $6,000/month. Telemetry from orgs that did this: typically a third of seats show meaningful weekly usage. You’re paying $6,000 for maybe $2,500 of engaged value.
Option B: segment. Forty people are heavy daily users — consultants drafting documents, partners living in Outlook and Teams. License them: 40 × $30 = $1,200. The other 160 get free Copilot Chat. You ship one HR-policy agent grounded on SharePoint; those 160 unlicensed users ask it an average of 20 questions a month:
- 160 users × 20 answers × 30 credits = 96,000 credits/month
- Four capacity packs (100,000 credits) ≈ $800/month at typical pack pricing, with PAYG as overage backstop
Total: $2,000/month versus $6,000 — and the 160 unlicensed people arguably get more useful Copilot than they would have gotten from an ignored license, because the agent is aimed at a question they actually have.
The crossover math is simple: a license costs the same as roughly 100 tenant-grounded metered answers per month. Anyone consistently above ~100 grounded interactions a month is cheaper to license. Anyone below is cheaper to meter. And the embedded-app experiences put a thumb on the scale: if someone needs Copilot in Word or in meeting recaps daily, license them regardless of the raw count.
The decision framework by persona
| Persona | Signal | Right answer |
|---|---|---|
| Document-heavy professional (consultant, PM, analyst) | In Word/PowerPoint/Outlook hours daily | License. Pays for itself in the first recovered hour per month |
| Meeting-heavy manager | 15+ meetings/week, drowning in recaps and follow-ups | License. Teams Copilot alone justifies it |
| Analyst working in Excel + data | Needs Analyst agent, in-Excel Copilot | License |
| Frontline / occasional user | Opens Office a few times a week, asks occasional questions | Chat + metered agents. Do not license; revisit if usage data argues otherwise |
| Everyone, for a specific workflow | ”We all need to query the policy library” | Build one agent, fund it with capacity packs. Cheaper than 500 licenses by an order of magnitude |
| Developer / IT | Building agents, not consuming them | Chat + a license for the builders who need to test embedded experiences |
Two failure modes to avoid:
Licensing by seniority. Executives get licenses first in most rollouts and are reliably among the lowest users — their assistants and chiefs of staff are the ones doing the document work. License by workflow intensity, measured, not by org chart.
Treating credits as scary because they’re variable. The agent usage estimator in Copilot Studio exists precisely to forecast credit consumption before you ship. Run the estimate, buy packs to cover the baseline, enable PAYG overage, review monthly. Variable cost you can forecast is not risk; it’s just arithmetic.
The operating rhythm
Licensing isn’t a one-time purchase decision; it’s a quarterly portfolio review:
- Start narrower than feels comfortable. It’s organizationally easy to add licenses and politically miserable to take them away.
- Pull usage data monthly from Copilot analytics in the admin center. Sustained near-zero usage after 60 days means reassign — the license isn’t a status symbol, it’s a tool sitting in a drawer.
- Watch the metered side too. An unlicensed user burning 150+ grounded answers a month through agents is a license candidate hiding in your PAYG bill.
- Design agents with the meter in mind. If an agent can answer from its instructions and embedded knowledge instead of live SharePoint grounding, it drops from the expensive rate to free. Sometimes the cheapest infrastructure decision is a better-written agent.
The orgs that get this right spend less and report higher satisfaction, because every license sits with someone who wanted it and every agent is funded at its actual usage. The orgs that get it wrong bought “Copilot for everyone” in one procurement cycle and are now explaining the renewal to a CFO who has seen the usage report. Be the first kind.