What happened

Microsoft reorganized Copilot into three sections, according to a post by Jared Spataro, chief marketing officer for AI at Work, on the company’s official blog on September 25.
Home brings together chat and Cowork, for delegated tasks, and integrates Word, Excel and PowerPoint inside Copilot. Code lets people without technical training build apps and dashboards in natural language, in an isolated environment inside each organization’s account.
Autopilot is a persistent agent: it follows conversations, picks up recurring tasks and runs complete processes without anyone calling it. It enters private preview at the end of the month.
The per-user license stays for everyday use, at a fixed cost. Cowork, Code, Autopilot and frontier models such as Astra and Fable move to usage-based billing. A selector called Auto chooses the model for each request based on accuracy, speed and cost.

Why it matters

For a Chilean company on Microsoft 365, the AI budget stops being a fixed line in dollars and comes to depend on how much the agents work. An agent that works when the office is empty also consumes when nobody is watching.
The Auto selector decides which model is used for each request. That means part of the cost is set by a vendor’s algorithm, not by whoever signs the purchase order. The spending control tool Microsoft is announcing in parallel comes through the same door.
Code puts people who used to ask for a spreadsheet to work building apps. The isolated environment reduces the possible damage, but it doesn’t replace someone reviewing what data each app touches.

2 bills. The per-person license covers chat and Office; work done with agents is billed separately, by consumption.

Variable billing is becoming the industry norm. Google had already made Gemini’s pricing more flexible because the AI bill had become a problem, and agents with access to email and documents bring their own risk, as the case of ChatGPT Work and prompt injection showed.

Autopilot: private preview from the end of September.
Home and Code: gradual rollout over the coming weeks through the Frontier program.
The announcement doesn’t publish a usage price table. No date for general availability.

Three days ago, Microsoft researchers measured that the best frontier agent gets 59.7% right on judgment tests. Now the company is selling one that works on its own, at night too, and is billed according to what it consumes.

Edited by Rodrigo Cornejo. How we select and verify: who writes these notes.