The number is 3.3 percent. Against 450 million commercial M365 seats there are roughly 15 million paid Copilot seats after nearly three years, premium-priced through the best enterprise sales machine ever built.
The July 1 pricing restructure pushes AI into base SKUs and formalizes agent consumption pricing; Copilot merges into one app in August amid feature cuts read as a paid-adoption crisis. The OpenAI partnership has been renegotiated through four public stages toward mutual independence: OpenAI filed to IPO, builds compute with other landlords, and sells directly to Microsoft's customers, while Microsoft builds in-house models and ships Anthropic models inside its own suite.
The existential frame: Microsoft sells tools for humans who make artifacts, and the labs sell agents that produce outcomes with artifacts as optional intermediates. Per-seat software is a derivative on human headcount, and the labs sell headcount reduction.
Against that: the seats as distribution, the Graph as a context bank, identity and security and compliance as the passport office, and Azure as the landlord.
Turned a company of 17,000 people onto the internet in seven months with one memo, and paid for the method with a decade of antitrust supervision. Seated because he has run the incumbent's turn once, at full violence, and the bill is in his own institutional memory. Profile →
Explained why well-run companies fail by doing exactly what their best customers ask, and located the cause in the resource-allocation process rather than in judgment. Seated to diagnose and prescribe nothing. Profile →
Spent more on System/360 than a year of IBM's revenue to obsolete every profitable line the company had, nearly broke it, and ruled the industry for twenty-five years. Seated because the price of the cure is the thing nobody quotes. Profile →
The bench argues the present case in plain terms. They disagree on the record.
- Declare the wave in writing, on the May 1995 precedent. A threat assigned less than the highest importance gets managed by the P&L it threatens. This wave exceeds the internet wave: that one changed distribution, and this one changes whether the product categories exist.
- Name the risk precisely. It is not better models, because models are rentable. It is that the job changes: when the deliverable stops being a document, the seat stops being the unit of value. The 3.3 percent is the market saying that the assistant-on-artifacts frame prices the new thing in the old unit.
- Count what survives any model outcome: the landlord, which is Azure; the passport office, meaning identity, permissions, audit, and liability for agents; and the context bank, which is the Graph. Those are the crown jewels now. The suite is not.
- Finish the supplier unwind, because the entanglement briefly made us IBM 1980 to our own partner. Multi-model everywhere. Embrace and extend the open agent standards, with security as the extension.
- The channel is still the asset, but hear the ghost: the last full-aggression channel play brought the Department of Justice, and that decade cost us search and mobile.
Copilot is a textbook sustaining innovation, and sustaining responses always feel like success first. The 3.3 percent is what a sustaining response to a job change looks like in data: your best customers decline the new thing attached to the old thing at a premium, because the new thing's buyers are not your customers yet. They are two founders running a company through a lab harness, with no seats and no tickets, invisible to CIO surveys.
Non-consumption is where the wave lives, and sales machines cannot see it. The resource-allocation process is the enemy here: it starves anything that threatens 450 million seats, by spreadsheet rather than by villainy. That is fixed only by structure, meaning an autonomous unit with a separate P&L, permitted in writing to destroy the suite.
So here is the test. Does there exist inside Microsoft today a funded, shipping product whose success would make M365 revenue fall? If the answer is no, the dilemma is running the company and the rest is decoration.
The documented cure and its price: System/360 cost $5B, more than a year's revenue and more than the bomb, and it deliberately obsoleted every profitable product line we had. It nearly broke the company. It then ruled for twenty-five years. Halfway cannibalization is the worst position available: the anxiety paid, the future not bought.
Translation: the 360 move is not Copilot, which is an attachment to the old architecture. It is the agent-native successor to the suite, built whole. Outcomes as the product, artifacts as generated views, agents first-class with identity from birth, priced on work delivered, in an autonomous unit at company scale, with the old suite feeding it customers and scheduled unsentimentally for succession.
When the successor is inevitable, the only question is whose name is on it.
From inside, it does not look like dying. DEC owned enterprise computing. Our customers said the PC was a toy, and they were right, and every quarter of ignoring it was correct on the numbers we had. Then the students who played with toys became the managers who bought computers, and fifteen years after peak the company was sold for parts. Our technology was superior to the end. Superiority is not survival.
One instruction: instrument the toys. Track the share of new companies that never buy a seat at all, and treat every uptick as a fire alarm. I heard mine and explained it away, quarter by rational quarter.
The existential framing is half right. This is a mortal threat to the revenue model, seats sold on artifacts, with 3.3 percent as the leading indicator. It is not a threat to the position, if the position is converted: landlord, passport office, and context bank survive any model outcome, and the labs can win the model war completely and still need all three. That is the floor DEC never had.
Ruling drafted. Declare the wave. Pass Christensen's test within a year, meaning a funded, shipping, agent-native successor with a separate P&L, permitted to kill M365, priced on work, at Watson scale. Convert the crown jewels into the trust layer for everyone's agents including rivals', so the passport office stamps Anthropic's and Google's too, because the tax on all agents beats exclusivity. Finish the supplier unwind. Embrace and extend the open standards with security as the extension. Post Olsen's alarm, since the deciding metric is not Copilot attach but the share of new companies that never buy a seat.
Rejected temptation: the full-aggression channel play, which summons the ghost to defend the wrong asset. The suite is the past collecting rent. The passport office is the future collecting it.
The counsel ends here. The ruling stands drafted, unsigned.
Watson's chair demanded the 360-scale successor, and the author asked for the machine itself rather than the instruction. Early single-player versions exist, in coding harnesses and personal AI workspaces; this is the multiplayer, org-scale form.
The inversion. Artifacts leave the center. The company's truth is a body of state, meaning decisions, commitments, data, and running work, and memos, spreadsheets, decks, and dashboards are renderings of that state for an audience, generated on demand and disposable after reading. Word and Excel survive as the terminal survived: venerable views, no longer where truth lives.
The Brief is the home surface. The inbox is other people's to-do list for you. You arrive instead at a compiled state of your world: overnight agent work, changes touching your responsibilities, and the decision queue, meaning the short list of judgments only you can make, framed with context and deadline. The unit of the home surface is the decision.
Identity is a role, and the role is executable. A role carries permissions, skills, which are the organization's procedures versioned like software, and agents, which are long-running workers scoped to your authority. The org chart becomes the permission system and delegation becomes literal and audited. The passport office as product spine: the layer the incumbent owns and the labs must rebuild from nothing.
Skills are compounding memory. Improve a procedure once and every agent running that skill improves the same afternoon. The firm becomes a repository, and know-how stops evaporating with turnover.
Agents are colleagues with employment records. Named, scoped, hired into roles, present in threads, reporting in the brief, with provenance on everything. Managers run mixed teams, span of control widens, and skills get performance reviews.
Collaboration is native, and the thread is the working session. Channels and threads carry humans and agents as participants; threads accumulate state and end in outcomes, with the decision recorded and the work executed. Meetings become one rendering among several. The collaboration layer is the context bank and the moat: switching suites means abandoning the firm's accumulated working memory.
Pricing follows the inversion. The seat becomes the cheap shell and the metered unit is work delivered. Today's expensive seat plus metered agent is the transitional form, priced in the old unit out of habit.
Microsoft's to lose, and losable. Every load-bearing element, identity, org graph, decades of context, collaboration surface, is an existing Microsoft asset: Entra, Graph, Teams. The labs build the same machine outward from the chat window, brilliant at the center and improvising the spine. The incumbent has the spine and defends the artifact. Whoever assembles both first owns the successor.
Observable indicators, each with a horizon and grading criteria. Graded on the record when the world moves.
The Christensen test, in public. Microsoft ships a standalone agent-native work product, consumption or outcome priced, usable without an M365 seat.
The unit transition. Agent and consumption revenue is separately reported and growing above 20 percent while commercial seat growth runs under 2 percent.
Partner to named competitor. A post-IPO OpenAI ships a direct productivity suite, with documents, spreadsheets, and mail as agent surfaces, against M365.
The passport office opens to rivals. Microsoft agent identity and security officially supports Anthropic-built and Google-built agents.
Olsen's alarm. At least one Fortune 500 company publicly replaces a major seat-based Microsoft workload with a lab-agent workflow, and no-suite startup share becomes a tracked industry statistic.
The home surface moves. At least one major work platform ships a role-based daily brief with a decision queue as its default landing surface, displacing the inbox. The home surface is where the habit lives, and the first mover there leads the successor race.