Seats, Tokens and Microsoft
Thirty years of per-user pricing meet an agent that does the work. Who has to change, and what I’d do as a buyer
Microsoft sold Office by the seat for thirty years.
Now an agent builds the spreadsheet, and a person only checks it.
I came across an investor thesis for shorting Microsoft. Its claim is simple: the model enterprise software has run on since the 1990s, one license per employee, has started to break.
I don’t share the conclusion. The question behind it is still the right one, and it applies to every company that buys or sells software.
Here is how the seat worked.
Word, Excel and PowerPoint were sold per user. A company hired 100 people and bought 100 licenses. CRM, HR, collaboration and security tools followed the same rule. Headcount was the meter.
It made sense while people did the work inside the software.
Then the work started to move.
On November 10, Anthropic launched Claude Sheets, Whiteboards and Data on all its paid plans. Sheets builds and edits spreadsheets with you. Whiteboards is a shared board where a team brainstorms with Claude. Data turns connected company data into dashboards and data stories.
That covers much of what people open Office for every day.
If an agent does the work and one person reviews it, every vendor faces a new question: what is a seat worth when fewer people touch the product?
Microsoft holds strong cards.
Azure revenue grew 48% last quarter, to about $33 billion.
At our enterprise clients, Microsoft’s position is very solid. In many companies, Copilot is the only window into AI an ordinary employee has.
It also has three pressure points.
First, capacity is shared.
The same GPUs serve Microsoft’s own products, model training and Azure customers. In January, CFO Amy Hood said that if all the newly added GPUs had gone to Azure, its growth figure would have been over 40, against the 39% reported.
Second, the backlog is concentrated.
In the same quarter, Microsoft said about 45% of its commercial contract backlog came from OpenAI. If OpenAI runs into money problems, Microsoft will feel it.
Third, it plays defense and offense at the same time.
Amazon has no Office to protect. AWS sells compute, and its Trainium chips can lower the cost of inference.
Google has a frontier model and TPUs, which it now sells to outside customers.
Microsoft has neither a model of comparable level nor its own chips proven at scale.
So I’d separate two things.
Demand for software will stay. Companies will keep automating processes and protecting data, possibly with bigger budgets.
What weakens is the link between headcount and the vendor’s revenue. The bill will follow finished work and compute.
This goes far beyond Microsoft. Workday, collaboration tools and most of enterprise software sit on the same model.
Salesforce saw it early.
It struck a deal with Anthropic earlier this year. It charges for Agentforce work in Flex Credits per action. And this week at Microsoft Ignite, the two companies announced Copilotforce, which brings Salesforce into Microsoft Copilot through MCP.
Accenture is rolling it out internally to 55,000 Salesforce users.
Going to where the user already works is the right instinct when the interface itself is moving.
Here’s the part I find most interesting.
The agents that weaken per-seat pricing will become the largest buyers of compute. It’s the Jevons paradox I wrote about in August: when something gets cheaper to use, people use far more of it.
For Microsoft, Office and Azure now sit on opposite sides of the same deal. Office may lose seats, and Azure earns on the compute the agents run on.
If you buy enterprise software, three things I’d do now:
Ask every vendor what your bill looks like if half your users stop logging in and agents do their part of the work.
Track cost per finished task, such as an invoice processed or a request resolved, next to cost per seat.
Before signing a three-year seat-based renewal, keep the option to move part of the spend to usage.
So I wouldn’t rush to bet against Microsoft.
Its agent tools, like Copilot Studio, are already billed by usage. The main Copilot product is still sold the old way, as one more per-user license on top of Office.
My forecast: within two or three years, most enterprise software contracts will carry a usage component next to the seat.
The vendors that move their pricing to the work being done will keep the customers that the seat count used to guarantee.

Discussion about this post