Microsoft holds 450 million commercial seats, the identity and security layer of corporate computing, and the second-largest cloud. Its revenue model is per-seat pricing on tools for humans who make artifacts. The frontier labs sell agents that produce outcomes, for which artifacts are optional.
After nearly three years of premium-priced Copilot attached to the suite, roughly 15 million seats pay: 3.3 percent. The July pricing restructure pushes AI into base SKUs and meters agents by consumption. The question is whether the company converts, and history has run this experiment several times.
This is the second sitting of Issue No. 4, and the first issue written precedents-first: three precedents told in full as the body, then analysts arguing from them. The first sitting's council, exhibits, and predictions stand.
The evidence of the issue. Each account closes with what it contributes to the argument; the counselors below argue from these cases and nothing else.
By 1961 IBM dominated computing and was strangling on it: six incompatible computer lines, each with its own architecture, peripherals, software, engineering fiefdom, P&L, and executives whose careers were the line. The terror underneath the dominance: a customer who outgrew a small IBM machine had to rewrite everything to move up, and if he had to rewrite anyway, a competitor's machine cost no more to move to. Honeywell noticed, and marketed a machine deliberately compatible with IBM's best-selling 1401, sold with software that translated 1401 programs automatically. They called it the Liberator. The most profitable product in IBM's history had become the beachhead for raiding it.
In 1990 Britannica peaked at about $650M in revenue, the most prestigious reference product on earth, growing. Thirty-two leather volumes at $1,500 to $2,000, and underneath the product, a sales force: more than 2,000 door-to-door reps in North America earning several hundred dollars per set, selling aspiration to parents at kitchen tables. The company knew digital was coming, and not vaguely: it owned Compton's, which shipped one of the first CD-ROM encyclopedias in 1989, and Microsoft approached it to license Britannica content for a home CD product. Britannica declined, because the brand could not be cheapened. Microsoft licensed down-market Funk & Wagnalls instead and shipped Encarta in 1993 at about $100, then effectively free, bundled with new PCs.
Microsoft has done the incumbent's turn once, at full violence, and both the method and the bill are in its own institutional memory. Netscape shipped Navigator in late 1994 and it became the most rapidly adopted product in software history, carrying a mortal idea: browser plus network as the platform, Windows demoted to a device driver. Gates read his own skeptical middle management and went over their heads. The Internet Tidal Wave, May 26, 1995, assigned the internet the highest level of importance, called it the most important development since the IBM PC, and named Netscape. In December, on Pearl Harbor Day, chosen for the symbolism, he turned the battleship in public.
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. Reads all three precedents through one variable: where the resources actually flowed. Profile →
Wrote the standing account of what it costs to introduce a new order, and what happens to the prophet who arrives unarmed. Seated to price the human bill of each precedent rather than to recommend a strategy. Profile →
Financed and reorganized the industries of an age, and never forgave the one sin his world does not forgive: running out of cash with commitments outstanding. Seated to say what a bet-the-company decision actually costs, and whether this client has one available. Profile →
Arguing from the precedents above and nothing else. They disagree on the record.
Read all three through one variable, where the resources actually flowed, and they become a single controlled experiment. Britannica's strategy documents saw everything; its resource allocation was governed by the comp plan, so the future was starved by the very process that made the company excellent. IBM's allocation was seized, physically, by two men with authority to override every line P&L, and that seizure, and nothing about vision, is why the 360 shipped. 1995 is the intermediate case.
So the diagnostic is embarrassingly concrete. First: does the successor platform have a Learson, one named executive with authority to hurt M365's P&L and the chief executive's public backing when the barons appeal? Second: has the sales force been re-quotaed onto consumption revenue, or does the 2026 comp plan still pay commission on the leather volumes? Until those two answers change, the July pricing move reads as Britannica.
Announcements are strategy documents. Comp plans are the truth.
Nothing is harder than introducing a new order, because the innovator makes enemies of everyone who prospers under the old and gains only lukewarm defenders. The three precedents are three prices paid, or refused. Watson introduced it fully: armed a hard man against his own barons, killed a product line over its leaders' objections, and fed the crisis his own brother.
Gates introduced it by charisma and siege, and note whom he made war on: the outsider, never his own barons, and the state punished the outward violence with a decade of supervision. Britannica's gentlemen refused to injure anyone, and so injured everyone, terminally, which is the fate I promised the unarmed prophet.
So the question for the present prince is not strategic but personal: which of his own is he prepared to hurt? Visible injury to the old order's beneficiaries, done early, once, from strength. If no baron has bled by this time next year, the gentleness is the answer, and Britannica is the prophecy.
Watson's $5B was more than a year and a half of revenue, financed with debt and equity while the software burned, and the company was genuinely at risk of the one thing my world does not forgive, running out of cash with commitments outstanding. That is what bet-the-company meant. Britannica's owners had no such balance sheet, and cowardice had arithmetic behind it.
Now mark Redmond's position, because it changes the moral character of the decision entirely. The company generates on the order of $100B of operating cash a year and already commits capital of that scale to data centers without discomfort. A Watson-scale successor platform, funded five years, priced to lose money while it conquers, would not move its credit rating. There is no solvency bet available to them; the capital markets have removed the excuse.
Which leaves only the price money cannot pay: the barons, the brother, the decade of nerve. When capital is infinite, the only scarce asset is the will to spend the people. The market will misprice the announcement in either direction; it priced the 360 wrong for three years. Ignore it.
The decisive evidence is not in any product announcement but in two administrative documents nobody puts in a keynote: the org chart and the comp plan. As of this sitting the visible record, 3.3 percent attach answered by base-SKU bundling and suite price increases, reads closer to Britannica than to the 360.
The bet, specified. The 360 was four commitments: one architecture replacing six, forward compatibility as the moat, 1401 emulation carrying the installed base across the bridge, and a whole-line announcement that ended the old catalog's future in a morning. The translation follows.
One architecture: the work graph. Convert the Graph from a record of artifacts, meaning files, mail, and meetings, into the operating record of work: commitments, decisions, processes, running agent work, with humans and agents as first-class principals. The instruction set is the triplet only this company owns at scale, accumulated Graph context, identity extended to agents with scoped authority and audit, and a runtime where the organization's skills version and compound. The moat afterward is the state: a customer that leaves abandons its firm memory.
Office becomes the 1401 emulator. The suite is not killed; it is demoted to renderings and editors of graph state, fully supported, explicitly the bridge. This is how the 360 obsoleted the cash cow without abandoning the cash cow's customers, and it is the answer to the 450-million-seat question.
Price the work, not the human. A cheap universal shell seat, metered agent work on top, and enterprise agreements rebuilt as consumption commitments. And the inverse-Britannica lever, the most concrete sentence in the plan: sales compensation credits consumption revenue against quota at a premium to seat renewals. Pay the force more to cannibalize than to protect. Britannica died of a comp plan.
Pass the Gates test on day one. The runtime runs rival and open-weight models natively, with in-house models as the price floor. The bet is never on out-frontiering the labs; it is that graph, identity, runtime, and comp plan are the durable assets while models commoditize into the supply chain.
Announce whole; ship to non-consumption first. Tidal Wave discipline, so no customer buys the old architecture believing it has a future. First shipments standalone and self-serve with no M365 prerequisite, to companies that were never going to buy a seat; then mid-market; then enterprise across the emulator.
The enforcement clauses. The 8000-series clause: the per-app Copilot add-on SKUs are killed, publicly, as proof the new order is real. The Learson clause: one named executive, separate P&L, permitted in writing to shrink M365, reporting to the chief executive alone. The size: on the order of $75B across four years, product plus pricing subsidy plus comp transition, which is less than one year of operating cash flow. Watson staked solvency; this stakes margin points and a stock multiple. In one sentence: bet the company that the record of work is worth more than the tools of work.
The council's counsel. The precedents yield a rule: rent components, never architecture. The Mosaic license worked because the browser was, at that moment, a component. The harness is the successor platform itself, the thing whose state, permissions, skills, and ecosystem become the moat, and a licensee learns none of the craft while the landlord learns all of it.
The prescription: recreate the harness in-house under the one culture in the company that has shipped harness-craft, which is GitHub's and never Office's; open the ecosystem to anyone's connectors, apps, and skills, converting the Graph from a wall into gravity; and sublimate Office into the emulator position. These are one condition, not two, because every open connector is a vote against suite exclusivity, and a closed harness over a demoted Office is only a smaller castle. Teams proves the build capability and contains the poison to avoid: it won by bundling into the suite, the new thing deployed to defend the old unit, the exact deployment the harness must invert. The enforcement clauses stand: the Learson, the separate P&L, and compensation that pays more for cannibalizing than protecting.
The counsel ends here. The ruling belongs to the author.
The council's job ended with the precedents and the grading. This part is mine, and it departs from the counsel in one important way.
On the move itself I agree with the bench: don't license Cowork or Codex. Licensing was an admission they couldn't build their own harness for work, and what they did with the license proves the point better than any argument. The real move is to recreate it, open the ecosystem to other people's connectors and apps, and sublimate Office 365 beneath it. The castle can't be defended. They need to storm it themselves.
But here is where I part from the council's prescription: I'm suspicious Microsoft can build it. The strongest evidence is the cage itself. Handed a complete harness, the organization's first act was to strip out dispatch, connectors, plugins, and scheduled tasks, wire the remainder into its own products, and price it into E7. That isn't a strategy error someone can memo away. It's a reflex, and organizations are their reflexes. Microsoft wins by distribution and fast-following; Teams proves that muscle and nothing in the modern record proves the other one. Recreating a harness that wins on product, deployed in a way that hurts their own suite, requires exactly the muscle they have used least.
So my ruling is a forecast, not a prescription: Britannica by default. Not because the strategy is unknowable, since the council just wrote it down in five points, but because the organization that would have to execute it keeps telling us who it is.
What would change my mind: the harness shipped whole, standalone, no seat prerequisite; a Learson named with public authority over the M365 P&L; sales comp inverted to pay for cannibalization; the build housed under GitHub. Any two of those and I'll re-rule. Until then, the signals table above is my scoreboard and the caged Cowork is exhibit one.
The ledger's first grading, entered the day the issue shipped. A signal is graded once, on the record, with the evidence that graded it.
July 26, 2026 · two Britannica signals fire. Microsoft licensed a full agent harness from Anthropic, the Cowork integration, and deployed it stripped: dispatch, connectors, plugins, and scheduled tasks removed, the chat surface retained and wired to Microsoft products, the whole gated inside the Frontier program on the premium E7 tier.
The strip list is a diagnostic map. What was removed corresponds one-to-one with threatened moats: connectors bypass the Graph's exclusivity as context source; plugins and skills seed an ecosystem whose standard Microsoft does not author; scheduled tasks and dispatch are agents working without a human seated, the direct solvent of per-seat economics. The amputations trace the castle walls, and the company's internal fear structure is legible from outside in the feature list alone.
The Compton's clause. Precedent II already contains this move: Britannica owned Compton's, shipped one of the first CD encyclopedias in 1989, and caged it to protect the sets. Acquiring or licensing the future in order to domesticate it now has a name in this ledger. The mirror is Precedent III: in 1995 the same company licensed the future from Spyglass and shipped it whole, free, into the widest channel on earth. Licensing what you lack is the 1995 play only if you ship it at full strength. Call it the Spyglass test; the Cowork deployment fails it.
The steelman, defeated by the tell. A staged rollout for agent-safety reasons is the benign reading, and agents with standing authority genuinely warrant gates. But safety rollouts gate by admin controls and tenant policy; this gates by price tier. E7 placement is a monetization decision: the licensed harness's assigned job is justifying the premium seat. That is the unit of account deciding, the Britannica mechanism operating in the open.
Reading updated: the preponderance moves further toward Britannica. What would reverse it: the harness shipped whole, with dispatch, connectors, plugins, and scheduled tasks intact, standalone, without an M365 prerequisite, priced on work, and Office attaching to it as the rendering layer rather than it attaching to Office.
Observable indicators, each tagged with which precedent it reads on. Graded on the record when the world moves.
The Learson clause. A named executive is given the successor platform with authority over, and hostile incentives to, the M365 P&L, publicly backed.
The comp plan. Enterprise sales compensation is rebuilt around consumption and agent revenue rather than seat renewals.
The declaration. A Tidal-Wave-style public declaration: the wave named, the competitor named, the reorganization dated.
The deeper bundle. Deeper bundling of AI into existing suite SKUs, with per-seat price increases and no standalone successor product.
The $995 CD. The agent-native product ships gated to M365 seats, priced to protect E5 economics.
The bill. Aggressive default bundling of agents across the estate draws formal regulatory attention in the United States or the European Union.