Microsoft’s Copilot Cowork Puts Agents to Work, With a Meter Running
Microsoft has introduced Copilot Cowork, a new set of AI agent tools that can take on extended tasks and run them largely on their own, and it has paired the launch with a new way of charging: by usage rather than by seat. The company said the pricing change reflects the cost of the computing behind agents, which can run for hours on a single assignment. The tools join a market that Google and Amazon have entered with their own enterprise agents, and the pricing model is an early answer to the question of how companies will pay for work done by software rather than people.
The product is a departure from the way enterprise software has traditionally been sold. Older AI tools were priced per user, like the software they replaced, with a license covering everyone who might use the product. Agent software is different: a single agent can do the work of many people, running continuously on tasks that would occupy a team for days, and the value it delivers scales with what it does rather than who uses it. Microsoft’s decision to charge by usage aligns its revenue with the work performed, a model it says is fairer to customers and necessary to cover its own costs.
The cost structure explains the choice. Every time an agent processes a document, queries a model or takes an action, it consumes computing resources, and a task that runs for hours consumes a great deal. Microsoft executives said the cost of inference, the computing required to run AI models, has climbed sharply as models have grown more capable, and that per-seat pricing no longer covers the expense of agents that run constantly. Usage-based pricing passes those costs through, and it creates a direct link between what customers do and what they pay.
The company has begun showing what the agents can do. One customer used the tools to compare nearly four thousand documents in a matter of hours, a task that would have taken a team of people weeks, according to Microsoft. The agents can also pull together materials for a complex meeting, integrating email, internal files and calendars to prepare briefing documents and agendas before anyone arrives. Those examples are intended to demonstrate that agents are not just faster versions of search, but workers that can take an assignment and carry it through to a finished product.
The competitive field is moving in the same direction. Google and Amazon have introduced agent products for their enterprise platforms, with similar claims about autonomy and scale, and the three companies are now competing on capability, price and integration with the office software businesses each already dominates. The battle has shifted from who has the best model to who can package models into work that companies trust to run unattended, and the pricing models are part of that contest.
Trust is the obstacle the whole category faces. An agent that works for hours on its own is an agent that can go wrong for hours before anyone notices, and enterprise buyers have been cautious about letting software act with that level of independence. Microsoft has responded with safeguards, including human checkpoints and audit trails, and with billing that lets customers start small. But the caution is structural: the same autonomy that makes agents valuable is what makes them risky, and the industry is still learning how much independence enterprises will accept.
For Microsoft, the launch is also a strategic statement. The company has bet its AI strategy on embedding intelligence into the software people already use, and agents are the most ambitious version of that bet. If Copilot Cowork works as advertised, it turns Microsoft’s office suite from a collection of tools into a workforce, with implications for how companies think about headcount, productivity and software budgets. If it falls short, it will join a long list of AI products that promised more than they delivered.
The usage-based pricing will be watched closely by the rest of the industry. If customers accept it, it could become the standard for agent software, reshaping how enterprise AI is sold. If they resist it, Microsoft will have to find another way to cover the cost of agents that run for hours. Either way, the launch has put a price on a new kind of work, and the market’s reaction will define the economics of the agent economy. The launch also raises questions about the workforce implications that the industry has so far avoided answering directly. Microsoft executives were careful to describe the agents as assistants that multiply what employees can do, rather than replacements, and the marketing around the product emphasizes the hours it saves. But the arithmetic of an agent that does the work of a team is hard to ignore, and the pricing model, which charges for work done rather than people hired, makes the economics explicit. Customers adopting the tools will make their own calculations about headcount, and the answers will shape not only Microsoft’s revenue but the broader question of what AI does to employment in knowledge work. For now, the company is positioning the product as an addition to the workforce, and the market will test that framing in practice.
This article was prepared by Rhino Finance’s editorial team based on public reporting.
