As Microsoft adds more advanced agentic capabilities to Copilot, such as Cowork and Autopilot, usage-based pricing is becoming an increasingly important part of its business model.
But for customers used to predictable, fixed pricing, the addition of pay-as-you-go billing can be a significant change, creating new challenges around managing and predicting costs.
Questions about the consumption pricing model are now among the most common that Gartner receives from clients around Copilot, said Max Goss, research director at Gartner. There’s “some doubt about whether a full move to consumption-based pricing for everyday knowledge work is sustainable for organizations in the long run,” he said.
When Microsoft introduced Copilot three years ago, it charged businesses a flat monthly fee of $30 per user, with a $21 per user subscription later introduced for small and midsize businesses. Initial capabilities focused on helping employees find information, summarize meetings and documents, and draft content.
Microsoft has since expanded Copilot with AI agents that can perform more complex tasks. Some, such as Researcher and Analyst, are included in the existing subscription, while others, including certain agents built with Copilot Studio or accessed via Copilot Chat, required usage-based charges through Copilot Credits, priced at $0.01 per credit.
Now Microsoft is bringing consumption-based agents more directly into the core Copilot experience. At a recent event, it unveiled a new Home interface that places Cowork alongside the standard Chat option in the prompt box. Unlike Chat, Cowork incurs usage-based charges, with credit consumption varying depending on factors such as the model used and the complexity of the task, according to Microsoft’s pricing documentation.
Navigating Copilot’s expanding pricing options
As usage-based agents become more prominent alongside the subscription-based tools in Copilot, customers must ensure employees use the right tool for each task to avoid unnecessary charges.
“Bringing these together in a single interface is going to make this harder,” said Goss. Employees could, for example, use Cowork for a task such as summarizing calendars, “potentially burning tokens on a task Copilot can do for free.”
Anthropic has taken a similar approach with Claude, recently announcing that it’s merging its chat and Cowork tools. The main difference is that Claude Enterprise charges for usage with both features.
Microsoft said it’s building a tool to automatically determine whether a user’s query should be routed to Cowork or Chat. Such routing could help employees avoid using more expensive capabilities unnecessarily, although its effectiveness in controlling costs remains to be seen.
There’s also a risk that a vendor’s automatic routing could favor its more expensive capabilities, said Keith Kirkpatrick, research director at The Futurum Group. “When vendors decide whether a task should fall into one bucket versus another, there’s a risk that they’re overly weighted to moving the task to the bucket where they’re going to monetize it,” he said.
For now, Copilot customers will need to educate employees about when more advanced tools are actually necessary, Goss said.
To do so, businesses that enable tools such as Cowork should establish guidelines around when employees should use the agent tool and explain the potential costs, Kirkpatrick said.
“It’s incumbent upon the enterprise — whether it’s IT or department managers — to be really clear and say: ‘Here are the tasks for which it makes sense to use Chat and here are the scenarios where it makes sense to use Cowork’ and be very explicit about what typical consumption might be,” he said.
Usage-based pricing at scale
The growing use of consumption-based pricing reflects the higher costs of running agentic AI — a broader challenge for AI vendors.
“The AI of the future won’t just be ‘chat,’ it’s going to be ‘chat’ and ‘do’ combined into one interface,” said Goss. “We’re in this sticky situation now where the cost of the ‘do’ is too expensive, so the AI companies are passing that on to end users in the form of pay-as-you-go pricing.”
For Copilot customers, the challenge is managing those costs as usage expands across an organization’s workforce.
Microsoft has already adopted consumption-based pricing for developers using GitHub Copilot, but whether the same model is suitable for everyday knowledge work is unclear.
“It seems that consumption-based pricing for a specific task like development is plausible and sustainable, albeit with some caveats around what you use it for and what you don’t,” said Goss. “The big question — that is still an open question — is: is consumption-based pricing scalable for general purpose knowledge work?”
Each time a knowledge worker runs an agent while working on a spreadsheet, the cost to the organization could vary depending on the task, for instance. The employee’s level of AI proficiency could also have an effect, as repeated attempts to get the right result could increase consumption. Multiply that by 10,000 or 20,000 employees, Goss said, and the question becomes whether usage-based pricing is sufficiently predictable and sustainable for large organizations.
To help customers control usage, Microsoft has developed a cost management tool. This lets admins create organization-wide and individual spending limits, monitor credit balance, and set up notifications when users reach a threshold.
While these tools make it easier to monitor spending, predicting future consumption can be another problem, Kirkpatrick said. “The challenge is around forecasting,” he said, as businesses adopting AI agents generally lack data on how employees will use the technology. To help estimate costs, businesses should examine employee workflows and determine which are covered by subscriptions and where consumption-based pricing is required.
A recent study cited in the Wall Street Journal showed that just 11% of surveyed firms successfully predicted spending on agentic systems.
Managing usage-based costs may also require closer collaboration between finance, IT, and business units to manage costs — an approach known as FinOps.
“Successful organizations are the ones that have devolved this,” said Goss, with IT retaining overall oversight while business units have nominated “steward” role to manage their own credit consumption within a fixed pool.
It’s a new operational model for most organizations, however. “Going forward, IT procurement and finance teams will have to work closer together in more of a technical FinOps model. Most companies aren’t there yet, or are not familiar with how to do this at scale.”
Proceed cautiously with Copilot agents
As well as managing expenditure, Copilot customers should determine whether spending on agentic tools delivers productivity gains, analysts said.
Proving Copilot’s value has been a persistent challenge for customers since the AI assistant launched. A Gartner survey of 186 IT leaders published in June showed that 51% saw value as one of their top three barriers to successful Copilot deployments.
The survey also indicated that Copilot customers currently see greater benefits in more basic capabilities. Almost half (49%) reported high value from content generation and 44% from knowledge search, compared with just 20% for automating workflows and repetitive tasks using agentic tools such as Copilot Studio. (The survey was carried out prior to the launch of Cowork, which was made generally available in June.)
Gartner
The introduction of pay-as-you-go agents could bring back questions around return on investment that businesses faced when Copilot first launched. “If an organization just gives employees [access to the tools] and says, ‘go for it,’ they’re almost certainly going to see wastage. It’s going to be back to the early days of Copilot, when people were asking; ‘Is this valuable?’” said Goss.
Kirkpatrick recommends that organizations establish clear goals before deploying agentic tools. “Start by looking at what business outcomes you’re looking to drive,” he said. “If you just focus on the technology, you’re going to wind up trying things out … without any real understanding of whether or not there’s actually business value being delivered.”
Another factor in assessing the value of usage-based functionality in Copilot is what customers actually pay for it. Customers could also negotiate Copilot Credits as part of agreements for Microsoft’s new top-tier Microsoft 365 E7 subscription, which bundles M365 with AI tools such as Copilot and Agent 365 but doesn’t include Copilot Credits.
“Microsoft is almost certainly talking to you about E7. At the same time, they’re probably having a conversation about usage-based pricing. Perhaps try and tie the conversations together … if you are going to move to E7, what can you get?” Goss said. “Microsoft could technically assign you credits, even if it’s just a starter amount. I would be looking to tie those two things together.”
Gartner
Before rolling out Cowork and other consumption-based tools in Copilot, Kirkpatrick recommends launching pilot projects to better understand employee usage.
Goss suggests starting with small-scale Cowork trials involving experienced Copilot users, with strict limits on credit consumption, and guidance on when to use Cowork over traditional Copilot. But he notes the challenge of scale remains.
“It might be absolutely fine to pilot Cowork with fifty of your most advanced Copilot users, and you might see some brilliant returns,” he said. “But can you scale it to the rest of the organization? That’s the really interesting question here.”