Microsoft Priced the Agent Seat at $15 a Month. Meta's Ex-Director Says That's the Wrong Unit

A widely shared headline — 'Meta AI director Rohit Patel leaves to take on Microsoft Office and Google Workspace' — is easy to misread, and several people read it as a defection to one of the two rivals. Patel never joined either rival: he spent eight years at Meta Platforms, most recently as a director in Superintelligence Labs working on its models, and left to build a competitor to both suites called Ndoc.

What Patel carried out of Meta is a research record built around measurement. His benchmarks include a modality maturity index of 893 self-contained questions across five modalities, a general intelligence benchmark of 820 expert-authored problems, a 6,500-triplet multimodal set spanning 13 domains, and a wearables set on which open and proprietary models scored between 24% and 52%. Ndoc turns that instinct into a product: links every number in a document to the script that produced it, so a reader can ask the agent which line holds the formula.

Microsoft put agents on the price list. Agent 365 became generally available on May 1, 2026 at $15 per user per month for anyone who manages or sponsors agents, and the new Microsoft 365 E7 bundle sells the combination for $99 per seat. Alphabet went the other way: Gemini now sits inside every Google Workspace plan, and list prices for Starter, Standard and Plus rise to $7, $14 and $22 per user per month from February 10, 2027.

The attach rate is the weak point, and it shows where the money is not yet. Microsoft said paid Microsoft 365 commercial seats passed 450 million in January 2026, growing 6% a year, and that Microsoft 365 Copilot had passed 30 million paid seats by June 30, with net seat additions more than doubling quarter over quarter. That is roughly 7% of the base — leaving more than 90% of the installed seat base unconverted.

That gap explains why the incumbents keep repricing rather than discounting. A per-seat agent licence raises revenue per user on a base that is already paid for; per-token pricing lowers it. Salesforce shows how far the second model goes — an action on Flex Credits costs $0.10, a conversation costs $2, and the user licence is $5 a month on top of credits. A single $30 Copilot seat covers 300 agent actions, which makes the seat an expensive way to buy inference.

Multiply Agent 365's $15 list price by the 450 million commercial seats Microsoft disclosed, then by twelve months, and the agent line alone turns 450 million seats into an $81.0 billion revenue line. For scale, Alphabet's entire Google Cloud segment did $58.7 billion of revenue in FY2025, up from $43.2 billion in FY2024.

That is the sharpest number in the story. Meta's most-cited bull case for Muse lands on $10.8 billion a year, which is also roughly what 30 million Copilot seats are worth at Microsoft's list price. On those figures, Meta's consumer agent is valued as a smaller version of an enterprise product that already ships, and the enterprise product already exists.

Meta Platforms's Muse sells a consumer subscription with no workplace distribution: it launched on September 8, 2026 as a personal agent running on a dedicated secure machine with a browser that can fill forms, negotiate and keep working after the app closes. Meta's own announcement describes a free tier plus subscription plans for people who want more, with distribution on iOS, Android and muse.ai and no enterprise offering anywhere in the release.

Reuters reported on September 22, 2026 that the paid Muse tiers are priced at $20 and $100 a month, that Muse logged 2.8 million downloads in its first 12 days on iOS and Android in the United States and Canada according to Apptopia, and that JPMorgan Chase's analysts argued it could become the most widely used consumer AI application since ChatGPT. The financial mirror of that gap is stark: Meta Platforms did $64.6 billion of revenue in Q3 2026, up 26% year over year, spent $20.05 billion on research and development and $34.91 billion on capital expenditures including finance-lease principal payments. Operating cash flow was $32.79 billion, so free cash flow turned negative at $1.16 billion in the quarter, with trailing capital expenditure of $104.4 billion against $133.1 billion of operating cash flow.

The uncomfortable symmetry is that the asset Meta lacks is the one it does not have to build. Meta Platforms's apps reached 3.64 billion daily active people in September 2026, an enormous distribution asset — but a consumer login is not a workplace seat, and no amount of consumer reach substitutes for the licence an agent needs at work. Meanwhile the one product that could bridge the two, an agent that can be trusted with a file, is the one thing the departure took out of the building.

Upstream, whoever wins, the tokens still have to be computed and stored — and Nvidia captures that work whichever surface wins. Data Center revenue reached $89.0 billion in the June 2026 quarter, up 117% year over year, on group revenue of $96.2 billion. Micron did $54.23 billion of revenue in its September quarter against $11.32 billion a year earlier, and $133.19 billion for the full fiscal year, at a $47.0 billion quarterly gross profit.

The mechanism is the point. An open document format lowers the cost of changing the model behind the document, so value migrates away from the model vendor and toward the compute vendor and the surface that meters usage. Salesforce has already made that move, from per-user seat licensing to $0.10 per action, and its agent revenue is compounding at more than 240% while Microsoft and Alphabet move the other way by adding per-seat AI charges to a base they already own.

The clearest cross-market proof that unbundling can work is a listed office suite already doing it. Kingsoft Office grew H1 2026 revenue 24.7% to RMB 3.31 billion while moving subscriptions toward AI-usage-based pricing inside the product rather than a separate agent licence. That is the shape Ndoc is betting the American market on, which means the model is not theoretical.

Over the next two quarters, three things decide whether this is a story or a trend. Ndoc's beta is promised within weeks and a full launch in 2027; an open-spec format only matters if a platform adopts it. Microsoft's next quarterly report, covering the September quarter, will show whether the agent licence is being bought standalone or only inside the $99 bundle. And Meta Platforms has to show a workplace surface, a distribution partnership, or an explicit statement that it is not building one — its fourth-quarter results, due in late January 2027, are the natural window.

Microsoft has 450 million commercial seats, an agent licence at $15 and a Copilot add-on at $30, and it grew Microsoft Cloud 27% to $214.4 billion in FY2026. Alphabet has the seat, the token business and the operating leverage, with Google Cloud operating income tripling year over year to $8.81 billion in the June quarter. Meta Platforms has the models and 3.64 billion daily active people, but no workplace seat, and has just lost the director who was building the agent layer. If surfaces win, the winners are the seat holders and the token meters, and Meta is on neither side of that trade.