The Copilot discount is a loan

Here is a piece of arithmetic worth sitting with. A Microsoft 365 Copilot seat costs $30 a month, and that seat now includes access to frontier models — Claude Sonnet for everyday queries, Opus-class models for deep reasoning, and since March a full agentic product, Copilot Cowork, that runs multi-step tasks across your tenant. A single heavy agent run can chew through millions of tokens. Price that usage at any lab’s list API rates and a genuinely active user can consume more compute in a week than their seat costs in a month.

Somebody is covering that difference. Right now it is not the customer.

The shape of a land grab

The list price has technically held at $30 since launch, which lets everyone involved say there is no discounting. Around that number, though, the promotions keep multiplying: 15 per cent off for three-year commitments of 300 seats or more, 30 per cent off for the largest SMB deals this summer, an $18 Business tier running as a promotion through the end of the year. Behind it sits a $30 billion Azure compute commitment with Anthropic — the models being bundled into the seat are not getting cheaper to serve just because the seat price stayed flat.

None of this is a scandal. It is what a platform company does when a category is being decided: price for adoption, absorb the margin, win the default. Microsoft has done it before, with browsers, with Teams, and it usually works. Analysts covering the licensing are already predicting the likely endgame — a shift from flat per-user pricing to per-user-plus-usage tiers once the land grab phase ends.

Which means today’s price is best understood not as a price but as a loan. The difference between what the compute costs and what the seat costs is being extended to you, and loans get called.

Why companies are taking the deal

It would be easy to say organizations are buying Copilot naively. Mostly they are not. The reasons are rational, and worth stating plainly, because they explain what happens next.

Copilot is easy to buy: it attaches to an agreement you already have, from a vendor already through your procurement and security review. It is easy to govern: it inherits your tenant’s permissions, your data residency, your compliance posture, and your admin console, which for a regulated business is worth a great deal. And per seat, it is cheap — cheap enough that the business case barely needs to be made, which is precisely why, in most organizations, it never is.

That last point is the quiet problem. When a tool costs $30 a seat, “roll it out to everyone and see” feels like diligence. The evaluation that would normally accompany a serious tooling decision — which roles benefit, which workflows change, what finished work it produces — gets deferred, because at this price the question feels academic.

The question arrives at renewal

It stops being academic the moment the pricing normalizes. Whether that is a lapsed promotion, a usage tier on top of the seat, or simply a harder line at the next enterprise agreement, the effect is the same: someone in finance asks what ten thousand seats of Copilot actually produce, and the organization discovers it never built the means to answer.

We wrote recently about the gap between feeling productive and measuring it. This is where that gap gets expensive. An organization that cannot distinguish its fifty genuinely transformed workflows from its five thousand lightly-used licences will make the renewal decision on vibes — and vibes, under budget pressure, cut as easily as they buy. The likely outcome is not a dramatic cancellation but a blunt one: seats cut across the board, including the ones that were quietly paying for themselves many times over.

The practical advice is almost embarrassingly simple. Treat the discount window as a measurement window. While the seats are cheap, instrument them: which teams use the tool daily rather than weekly, which workflows produce finished work that would not exist otherwise, where the agent runs actually land. That evidence costs little to collect now and is nearly impossible to reconstruct at renewal time.

Subsidized pricing is a genuine gift to a buyer who uses the subsidized period to learn. It is a trap only for the buyer who mistakes the promotional price for the real one — and builds a dependency priced at the former, due at the latter.