Microsoft token economics: what AI really costs

Microsoft is moving AI from the license to the meter, and every prompt now draws on a balance of credits, tokens or compute units. You know your Microsoft license cost to the dollar, but probably not what AI cost you last month or who caused it.

1. From hype to reality: Microsoft’s own numbers show when AI hit the invoice

Four years ago, AI was a demo nobody budgeted for. Today it is a line on the invoice, and Microsoft’s results show when the shift happened.

The seat business grows steadily. Azure consumption pulls away from it. The AI business goes from almost nothing to a double-digit share of Microsoft’s revenue in four fiscal years.

Switch the chart to growth and the pattern gets sharper. Microsoft 365 commercial grows at 13% to 18% a year, the pace of a seat business. Azure accelerates again from FY24, and the AI business more than doubles every year.

Azure is about to overtake Microsoft 365. In FY26 Azure reaches about USD 100bn, against 102bn for Microsoft 365 commercial. That is three times its FY22 size, and its share of Microsoft’s revenue rises from about 17% to 30%. Part of that growth is AI: model inference, Copilot and agents all run on Azure capacity, and as AI workloads ramp up, Azure growth climbs back to 41%.

Every point of that growth is consumption that somebody inside a customer pays for. It is the part of your Microsoft bill that no longer follows your headcount.

The FinOps discipline has noticed. Surveys from 2026 show AI spend running past plan, and in the State of FinOps 2026 the FinOps Foundation names AI cost management as the skill teams need most.

The foundation calls this field tokenomics. We deliberately say token economics, because it reads more easily and says what it means. Either way, the field covers turning capital and energy into tokens, consuming them efficiently and mapping the spend to business outcomes. FOCUS, the open billing standard behind many cost tools, is set to add AI token tracking in its next version.

2. From seat to consumption: you now buy work, not access

Microsoft launched Microsoft 365 Copilot in November 2023 at a fixed price per user. The budget was known in advance, and the cost driver was the number of licenses.

Since then, Microsoft has put a meter next to almost every AI seat. On 1 September 2025, Copilot Studio replaced messages with Copilot Credits. Cowork, the Microsoft 365 agent for multi-step work, bills in the same credits. GitHub moved Copilot to usage-based billing on 1 June 2026.

In the consumption world, cost varies per action, the budget needs a forecast, and the cost driver is how agents behave. Your question changes from «how many licenses?» to «how much capacity, and governed by what?»

3. How a credit becomes a bill: three streams, one stops on its own

The question does not set the price of an AI task. What the agent has to read and do before it can answer does. Two prompts that look identical can differ by a factor of ten.

Take one task in our own tenant. The prompt «Create a Word file named Test» is six tokens. Before Cowork can act, it adds the site path and the ID of the document library, another 24 tokens: four times what was written. Then Cowork works in laps: analyze, access, write, check. Each lap carries the whole context forward, so the context grows with every lap, to 10'600 tokens over four laps. For one empty Word file, that came to 220 credits, or USD 2.20 at USD 0.01 per credit.

Six dials set that bill: the model, the size of the context, how many agents run in parallel, how long a run takes, how many tools it calls and how often it repeats. Microsoft prices them, you set them. Give the same prompt 60 pages of context, and it needs more laps and more credits.

Microsoft publishes planning bands for Cowork. A light task uses 100 to 300 credits, a medium task 300 to 700 and a heavy task 700 or more. The credits reach your bill in three streams, and each carries a different risk.

The ad hoc task is the harmless one. A schedule is a standing order. An agent grows with adoption rather than with your headcount, and that is the stream that grows on its own.

4. Five meters, no common currency

Microsoft prices AI in at least five units, and they do not convert into each other. Each one also behaves differently when the budget runs out.

AI meters in the Microsoft stack (USD list prices, October 2026)

Service	Unit	How you buy	When the budget runs out
Microsoft 365 Copilot, Cowork, Copilot Studio	Copilot Credit, USD 0.01	Pay-as-you-go, capacity pack, pre-purchase plan	Copilot Studio enforces at 125% of prepaid capacity, unless pay-as-you-go is linked
Power Platform AI Builder	AI Builder credits, moving to Copilot Credits	Seeded credits end on 1 November 2026 for new and renewed licenses	Falls back to Copilot Credits, otherwise blocked
Security Copilot	Security Compute Unit (SCU)	USD 4 per provisioned SCU and hour; included in Microsoft 365 E5 and E7	Overage at USD 6 per SCU and hour
Azure AI Foundry	Tokens or provisioned throughput (PTU)	Pay-as-you-go or reserved PTU	Keeps running and lands on the Azure bill
GitHub Copilot	GitHub AI Credits, priced by tokens	Included in the plan, extra usage on a budget	Admins cap the spend or pay for extra usage

Watch the naming overlap. A Copilot Credit and a GitHub AI Credit sound alike and live in separate contracts. People using Copilot are subscriptions. Agents doing work are consumption.

5. One tenant, CHF 82'566 a year in Cowork credits

The example is an anonymized demo tenant in the Zation Platform. It spends CHF 82'566 a year on Cowork credits, valued at the rate the tenant actually pays rather than the list rate.

Before any of that comes down, four layers have to work together. Most organizations stop at the first one.

• The invoices give you the amount. Four services, four billing units, several invoices and sometimes two currencies.
• Usage data gives you the cause. It shows who consumed what, per user and per agent.
• Identity makes allocation possible. Entra ID joins usage to teams, departments and cost centers.
• Chargeback changes behavior. Consumption nobody pays for is consumption nobody questions.

Once usage and identity are joined, five patterns show up. Each prices a different part of the spend, so the amounts add up without double counting.

The two largest levers are habits. Chat candidates are people whose tasks stay small and ad hoc, and Copilot in Word, Outlook and Teams covers that work within the Microsoft 365 Copilot license they already hold. Cowork by default describes people who reach for Cowork when the task needs no multi-step work across several sources.

Context scope is the third large lever. An agent pointed at whole sites and long time windows reads far more than the task needs, and the context it reads drives the credit count. So build context deliberately for AI: keep content that agents reuse often, such as policies, product data or project facts, in small, well structured sources of its own, apart from the long tail of documents. An agent that finds a lean, curated source reads less per run and answers more reliably. The schedule levers are smaller in money and the quickest to pull.

None of the five touches a license. The CHF 28'279 come from usage, not from a contract.

6. Adoption cuts both ways

Cutting adoption to cut cost also cuts the return. A seat nobody uses is the most expensive seat you own.

More adoption brings more value and more cost. Awareness turns the cost half back down: people who know what a task costs choose Chat over Cowork, send less context and switch off schedules they no longer need. Most of these levers sit with your users, not in an admin console.

Buying capacity follows the same logic. A Copilot Credit capacity pack costs USD 200 a month for 25'000 credits, billed annually, which is 20% below pay-as-you-go. It pays for itself at 80% use. Unused credits expire at the end of the month, and overage falls back to the list rate.

Microsoft also changes the defaults. From 1 December 2026, new Copilot Business purchases through CSP arrive with pay-as-you-go switched on and a default of 4'000 credits per user and month. The post on 2027 cost shifts runs the numbers for 1'000 users.

7. Five steps from invoice to control

1. InformJoin invoices, usage and identity every month. Seats and activity come from the Microsoft 365 admin center, token spend from Azure Cost Management, credits from the invoices and cost centers from Entra ID. The Zation Platform connects these sources once and runs the join for you.
2. InformPublish a monthly chargeback on an agreed basis. Agree the allocation key once, allocate per user and per agent, and send the result to the budget owners.
3. OptimizeWork the five levers in order of money. Show Chat candidates where Copilot covers the task, scope agents to the sites and time windows they need, and switch off schedules nobody reads.
4. Operate, before you enable a serviceSet the control first. Know which meter stops spend and which only reports it. Give every agent an owner and a cost center on the day it is published, and lower Microsoft’s spending defaults before the first invoice.
5. Operate, after three months of dataBuy capacity on measured usage and track value next to cost. Packs cover the base load, a pre-purchase plan covers the swing, pay-as-you-go covers the rest. Pick one unit per use case, such as cost per resolved case, and report it beside the spend.

Limits and exceptions

• The credit bands are planning estimates. Microsoft names model, context, runtime and tools as the drivers of actual use, so a task can land outside its band.
• The demo figures show potential, not a forecast. The levers and their thresholds are rules of the Zation Platform, applied to a demo tenant.
• List prices are orientation. Microsoft and GitHub publish USD list prices that change, so check the current rate card before any commitment.
• Security Copilot inclusion rolls out in phases. Microsoft 365 E5 and E7 include 400 SCU a month per 1'000 licenses, up to 10'000 SCU, without rollover.
• Prepaid capacity expires. Pack credits expire at the end of each month and pre-purchase units at the end of the term.

Questions to settle before the next invoice

• Which team, user or agent caused last month’s AI spend, and who in your organization can answer that?
• Which of your meters stop spend at the budget, and which only report it afterwards?
• Who owns each published agent, and which cost center pays for it?
• Which scheduled tasks ran last month without anyone opening the result?
• What does one unit of value cost you, per task or per resolved case?

Bottom line

Organizations that see AI usage per user and per agent decide where the spend goes. Organizations that see only the invoice learn their AI budget after the money is gone.