Microsoft 365 Copilot Cowork reached general availability in June 2026, and it arrived with a billing model that differs fundamentally from the Copilot most organizations have deployed. Where Microsoft 365 Copilot is a per-user subscription with predictable monthly cost, Cowork bills on usage through Copilot Credits. Alongside it, Microsoft introduced Scout, positioned as an always-on personal agent.
Customers on the Frontier programme were required to configure usage-based billing by June 30, 2026 to maintain access. More broadly, the Work IQ API also reached general availability on June 16 with consumption-based billing through Copilot Credits and no separate per-user licence, requiring administrators to enable pay-as-you-go billing in the Microsoft 365 admin center for any customer using Work IQ in custom agents.
Read together, these are the same signal: Microsoft's AI portfolio is bifurcating into a subscription tier for predictable assistant use and a consumption tier for agentic work. Most organizations will end up running both.
Why consumption pricing needs different governance
A per-user subscription has a natural ceiling. You know the seat count, you know the rate, and the worst case is that you paid for licences nobody used. Consumption has no ceiling by default. An agent in a retry loop, a poorly scoped workflow triggered by every inbound email, or an enthusiastic pilot team can generate cost at a rate nobody modelled, and the discovery mechanism is the invoice.
This is a solved problem in Azure, where cost management, budgets, alerts and tagging are standard practice. It is a new problem in Microsoft 365, where the admin centre has historically been a place to assign licences rather than to control spend. The organizational consequence is that the people administering Copilot are frequently not the people who have ever managed a consumption budget.
The controls to put in place first
Before enabling pay-as-you-go, four things should exist. An owner: a named person accountable for Copilot Credit consumption, with the authority to disable a workload. A budget: an explicit monthly figure derived from a modelled estimate of the intended workload, not from an aspiration. Alerts at meaningful thresholds — fifty, seventy-five and ninety percent — routed to a channel someone actually monitors. And a scoping decision: which users and which agents are permitted to consume, enforced through assignment rather than through policy documents.
Beyond that, three practices matter in the first ninety days. Pilot with a bounded population and measure actual credit consumption per user per week, because published estimates will not match your workload mix. Instrument the agents themselves so consumption can be attributed to a specific workflow rather than appearing as a single tenant-level number. And review weekly at first — monthly review of a consumption service means a four-week feedback loop on a cost that can move in days.
Where the value actually is
The consumption model is not a disadvantage. It is the correct pricing for work that is bursty and unevenly distributed, and it removes the central objection to per-seat AI licensing — that you pay the same for the heavy user and the person who logs in twice a month. For an organization with a small number of high-value automated workflows, consumption is substantially cheaper than licensing every potential participant.
The failure mode is not the model, it is deploying the model without the operational discipline that consumption services require. Organizations that already run Azure cost management well will find this straightforward. Organizations that do not should build the practice on Copilot Credits at small scale before the numbers get large.
A note on Scout
Microsoft Scout, introduced as an always-on personal agent, raises a governance question worth settling early: an always-on agent observing a user's work has a standing view of that user's data. The data residency, retention and audit position for Scout should be understood and documented before it is enabled broadly, particularly in regulated environments and in the EEA.
How Lorexus engages
We configure pay-as-you-go billing with budgets, alerting and attribution before enablement, run the bounded pilot that produces real consumption data for your workload mix, and build the weekly review cadence into your operations. Where Copilot Credits and per-seat Copilot both make sense, we model the split. Book a free 15-minute call with our senior engineers.