A Copilot Credit costs one penny.
That is the number Microsoft wants enterprise buyers to anchor on, and it is the clearest signal yet that the company is done pretending AI is just another line item on a per-seat invoice. Last week we asked whether SaaS repricing would force enterprises to consolidate seats or unbundle them. This week Microsoft answered with both: a unified Copilot platform for everyday work, and a separate credit meter for everything agentic.
The relaunch collapses Chat, Cowork, Code, and Autopilot into a single experience organized around three tabs: Home, Code, and Autopilot. CEO Satya Nadella called it "a new OS for work" in a LinkedIn post. The company's chief marketing officer for AI at Work, Jared Spataro, framed the change as pairing a new layout with a new billing logic: everyday chat stays on a per-user license, while work performed by agents is billed by use.
everyday chat stays on a per-user license, while work performed by agents is billed by use
That split is the story. The Copilot user subscription license, at $30 per user per month paid yearly, covers Chat and Copilot inside Word, Excel, PowerPoint, Outlook, and Teams. It comes with usage limits. Microsoft says most people will never reach them. Heavy users get a warning, then a choice: switch to Auto mode at no extra cost, or move to credits. Everything else—Cowork, Code, Autopilot, advanced SharePoint features, and frontier models like Fable and Astra—runs on usage-based billing paid in Copilot Credits. You need a seat license to use any of them. The agent work is metered separately.
The credit system is not a la carte pricing. It is a full FinOps layer. Administrators can create spending policies that control access to specific agents and services, configure organizational and user-level spending limits, set threshold notifications, and use prepaid credits, pay-as-you-go billing, or a combination. There is a Cost Management dashboard in the Microsoft 365 admin center that monitors consumption by spending policy, group, user, agent, service, and funding source. Microsoft recommends setting a monthly per-user limit to prevent a single person from spending all available credits—a small detail that tells you exactly what early enterprise deployments looked like.
This is the direct evidence for the enterprise seat-consolidation question we raised last week. Microsoft is not asking CIOs to choose between per-seat and usage-based. It is asking them to run both, with the seat as the entry ticket and credits as the meter for anything that actually does work autonomously. The company also added Copilot Managed Runtime, a governed hosting environment for apps built with Code, Cowork, and Copilot Studio. That runtime keeps AI-built applications inside the customer's Microsoft 365 environment and under IT's rules. It is in public preview.
The consolidating logic is straightforward: give employees one place to describe what they want, let Copilot route the task—to Chat for quick questions, Cowork for delegated multistep work, Code for building, Autopilot for persistent agentic tasks—and bill the expensive parts by consumption. Analysts quoted by InfoWorld said the practical shift is from AI that knows information to AI that can connect what happened, why it happened, and what needs to happen next. That requires enterprise context. Fabric IQ, which connects Copilot to business data including Power BI reports, is now generally available in Chat and Cowork. Work IQ, which brings in data and workflows from Dynamics 365 and Power Platform, enters public preview over the next month.
But the credit model introduces its own complexity, and the evidence from adjacent platforms shows what happens next. Make.com renamed its billing unit from Operations to Credits at the end of August 2026. The switch was one-to-one for standard modules—ten thousand Operations became ten thousand Credits, plan prices unchanged. But AI modules changed the math. AI modules with your own provider key still cost one Credit per run because you pay token costs directly to the provider. Make's own AI features and AI Agents consume dynamically. Depending on the model and token volume, a single call can cost several times one Credit.
The practical cost traps Make documents should be required reading for anyone deploying Copilot agents. An Iterator that breaks an array into individual elements—say, fifty rows from an Excel file—charges a Credit for every element that passes through another module. A scenario that reads fifty rows and runs three modules per row generates 150 Credits or more from a single run. Polling instead of webhooks consumes Credits even when there is nothing new; at a five-minute interval, that is theoretically up to 288 requests per day per scenario regardless of actual data volume. Error loops and retry logic multiply consumption within hours. And AI modules without measurement are the new cost trap since the credit switch. An AI step billed through the platform no longer has a fixed price. The longer the prompt and response, the more Credits. An agent that makes several tool calls per request adds up quickly.
Microsoft's own documentation acknowledges this risk by building spending policies and approval routing into the admin center. But the structural problem remains: the cost driver is not business volume, it is how the workflow is technically built. A daily stock reconciliation scenario that polls every fifteen minutes, processes twenty orders with three line items each through an iterator, and writes status back to the shop consumes roughly 256 Credits per day—around 7,680 Credits over thirty days—for one scenario. At a penny per credit under standard pay-as-you-go pricing, that is about $76.80 per month for a workflow that handles twenty orders a day. Add two or three similarly built scenarios and the entry-level quota is exhausted. The business volume stays modest. The credit consumption does not.
This is where the pricing philosophy question becomes urgent. Elena Verna's argument for a pricing constitution—a single document that maps out key trade-offs and the reasoning behind them—was written before this Microsoft announcement but reads like a direct response to it. Her principle at Lovable: never manufacture pain to monetize it. Charge for more value, not for removing friction you intentionally created. Her second principle is harder for AI products: monetize the outcome, not the steps required to get there. Every time you monetize an input, you give customers a reason to ration the exact behavior that leads them toward value. If someone needs to perform a hundred actions to get one valuable outcome and you charge for every action, you have created a pricing model that encourages people to use your product less.
Microsoft is monetizing steps. Credits are consumed by agent runs, tool calls, model selections, token volumes. The company mitigates this with Auto mode, which routes each request to the model and level of reasoning Microsoft judges is the best balance of accuracy, speed, and cost. It mitigates with included models—OpenAI's GPT-5.6 and Claude Sonnet 5 are included with the seat license; Anthropic's Opus 5 is included with limits. It mitigates with spending policies and user-level caps. But the fundamental unit of billing is consumption, not completion.
The counterargument is that agentic work is genuinely variable in a way that seat-based software never was. A lawyer reviewing contracts for eight hours consumes a predictable amount of Word. An Autopilot agent watching Teams channels, following up on threads, running recurring tasks, and picking a project back up days later consumes an unpredictable amount of compute. Microsoft said capacity constraints have been a recurring challenge, flagged across several consecutive quarters as demand for AI infrastructure outpaces available supply. The credit model is a way to price that scarcity without raising seat prices or capping adoption outright.
The strategic bet is that enterprises will accept metered agents if the platform consolidates everything else. One destination. One admin center. One set of policies. One plugin registry for connecting capabilities. One runtime for running what employees build. The sprawl that CIOs were worried about—AI tools scattered across apps, duplicate licenses, shadow deployments—gets addressed by making Copilot the single front door. The cost gets addressed by making the expensive parts visible and controllable.
But visibility is not the same as predictability. The Make.com evidence shows that credit consumption is a function of architecture, not intent. A well-built workflow with webhooks and clean error handling can handle hundreds of leads on the cheapest paid tier. A poorly built one with polling and iterators can exhaust an entry-level quota on twenty orders a day. Microsoft's spending policies can cap the damage, but they cannot tell an admin whether the agent is built efficiently. They can only tell them when the credits are gone.
The deeper tension is between Microsoft's stated goal—making Copilot a unified platform where users describe what they want and the system routes the task—and the billing model that charges by the step. If Copilot gets good at routing, it will route to the expensive options when they are the right fit. If it routes poorly, users will burn credits on the wrong execution path. Either way, the meter runs. The seat license becomes a gateway drug. The credits become the habit.
Microsoft has not broken out AI-specific revenue figures separately, saying the contribution has spread across multiple business segments. That makes it difficult to judge whether the credit model is working. But the company's decision to consolidate the product and split the pricing is itself a signal. It suggests that enterprises were not buying Copilot as a per-seat product at scale. They were buying it as a capability, and capabilities get metered.
Last week we asked whether enterprise buyers would consolidate AI seats or unbundle them. This week the answer is both, with a credit meter attached to the unbundled part. The question for the next quarter is whether CIOs can build the governance to match. Spending policies and user-level caps are a start. But they are controls on consumption, not on value. The companies that figure out which agent runs are worth the credits—and which are just expensive polling loops—will be the ones who turn this pricing model into an advantage. Everyone else will be watching the Cost Management dashboard and wondering why the meter keeps running.