I presented on Copilot Cowork at the AI Community Conference in New York, and the room was packed. People were genuinely excited about what the technology could do, and that excitement was easy to feel. But one concern kept coming up repeatedly: consumption-based pricing. That instinct is the right one. Cowork introduces something most knowledge workers have never had to consider before: every AI task now carries an economic decision. Instead of asking only, “Can AI do this?” organizations also have to ask, “Is this particular task worth paying for?” Most organizations are not ready for what that change actually means.
From Flat-Rate to Pay-Per-Use
Until now, Microsoft Copilot has been easy to budget. Copilot Chat is included with an eligible Microsoft 365 subscription, available on the web and governed by standard fair-use access rather than a true unlimited tier. Microsoft 365 Copilot runs on a per-user subscription model, so the cost stays fixed no matter how heavily someone leans on it. Early usage limits existed, but Microsoft has relaxed most of them as the product has matured. The point is that none of it had variable costs. You knew the number before the month started.
Cowork breaks that pattern. It is billed on consumption, denominated in Copilot Credits, priced at one cent per credit under pay-as-you-go. You pay for the work you actually hand it. Developers already know this rhythm from Anthropic's Claude Code and GitHub Copilot, both of which moved to consumption billing, but it is a new and unfamiliar concept for most knowledge workers. Cowork is also not included in the Copilot seat. It sits on top of the license, off by default, and an admin has to enable billing before anyone can run a single task.
Developers are still learning how to evaluate the cost of tokens, and that same evaluation is now landing on knowledge workers. Token-based consumption models are a newer concept outside of developer workflows, so most knowledge workers haven’t had a reason to weigh them before. Weighing token consumption against the value of a finished deliverable is a difficult judgment call in any role. For a knowledge worker deciding which AI tool fits a given task, it's an entirely new decision-making layer.
The Ten-Dollar Question
Here is a concrete example from my own work. I needed to reformat my conference presentation into the event's official speaker template. Normally, I skip that kind of busy work because reformatting a deck by hand means wrestling with fonts, colors, spacing, widgets, and slide masters, and the payoff rarely justifies the effort.
This time I handed it to Cowork. I gave it the existing deck and the new template, and I was strict about readability and a professional finish, since I would be standing in front of a room presenting it. It ran for roughly 45 to 50 minutes, taps into the frontier models, and runs quality checks that some other services skip. The output was strong enough that I used it at the conference without hesitation.
The cost was just over 1,000 Copilot credits. At a penny a credit, that is about ten dollars and some change.
Ten dollars for a conference-ready deck sounds like a fair trade, and on its face it is. But it raises a question that is harder to answer than it looks: Who confirms the work product was worth the spend? Once you start asking that, you are factoring in what you make per hour and what else you could have finished in that same window. For a developer, that calculation is becoming second nature. For most knowledge workers, it is brand-new territory, and they will be making that call dozens of times a week without a clear sense of what good judgment even looks like yet.
The Enterprise Governance Gap
On a personal plan, the limits stay manageable. You hit your ceiling, you wait a couple of hours, or you buy a little more, and either way, you control the spend. I can plan my own workaround those constraints without much friction.
Inside an enterprise, that control fragments fast. Procurement, IT, governance, and risk all hold a stake in how the tool gets approved, funded, and monitored, and that means more approval, funding, governance, and monitoring requirements sit between the user and the outcome. Microsoft has built real safeguards into the Cost Management dashboard in the admin center, where admins can set budgets, spending limits, hard caps, and alerts at the tenant, group, or individual level, and users can type a quick command to see what a task has consumed so far. Those controls help contain spend. However, they do not answer the harder question, which is how you tie a specific task to its credit cost and prove, repeatedly, that the investment paid for itself.
For CFOs, measuring that value is not straightforward. Most organizations will need someone monitoring usage closely and writing internal guidelines on when to reach for Cowork versus standard Microsoft 365 Copilot, which is already covered by the existing seat. Routing a job to Cowork that a standard entitlement could have handled is an easy mistake to make, and an expensive one, especially in the early days before anyone has developed good judgment. This is the same challenge showing up across the industry as consumption-based AI spend becomes the norm – a discipline some are now calling AI FinOps, tying variable AI cost to the outcomes it actually produces. The credits also pool with Copilot Studio, so the spend you are watching is not always coming from where you think. All of this friction can slow adoption at exactly the moment a company is trying to speed it up.
Start with Rapid Prototyping
One framing helps organizations get off on the right foot: Treat Cowork as a rapid prototyping tool rather than an everyday assistant handed to everyone at once. Unlike a shared agent that serves an entire team, Cowork is scoped to the individual, built to feel like a personal assistant. That personal, exploratory nature makes it especially well-suited to high-value work rather than routine tasks.
Think about the front end of a software project. Requirements gathering often drags on for weeks while business users try to describe what they need and developers try to interpret it. Now imagine handing those same users a tool like Cowork. They can show how they want something to work, connecting across Monday.com, SharePoint, Outlook, and Miro, steering the whole thing in plain language, and learning from each pass in minutes instead of meetings. Compress weeks of back-and-forth into a day or two, and the value becomes measurable: faster time to market, lower project overhead, and a cleaner handoff to IT when it is time to build the enterprise-ready version.
Orchestrating work across services and shaping the output through natural language is what makes Cowork approachable, even for people who would never call themselves technical. Accessibility is the part worth paying attention to, and it is easy to miss if all you see is the variable cost.
The Takeaway
Cowork is a real shift in how AI gets delivered and billed, and the excitement around it is earned. But consumption pricing introduces complexity you should not underestimate. The ROI is still difficult to quantify, mistakes will happen, and building the governance to manage usage well will take real time and effort. Now that Cowork is generally available, that work is no longer hypothetical – it applies to every organization running it, not just the tenants who piloted it through the Frontier program.
The right posture right now is experimental. Go in with curiosity, track what you spend against what you get back, and resist the two easy traps: dismissing the tool because the cost varies, or assuming it is the right answer for every job. Focus on scenarios where Cowork delivers outcomes that standard Microsoft 365 Copilot cannot, such as rapid prototyping, where the speed and capability gains are concrete and measurable. Tie each task to a measurable outcome, and the ten-dollar question gets a whole lot easier to answer.
For a deeper look at how organizations are tackling AI FinOps, governance, and readiness at scale, check out the State of AI 2026 report.
