One AI agent does the work of ten people — how many seats do you sell? Pricing's third revolution.

Software pricing has been through two revolutions. First, perpetual licenses — buy the software once. Second, the SaaS era’s subscription-by-the-seat — pay per user, per month, the model this whole book has been dissecting. A third is now beginning, and AI is its trigger: the shift toward charging by usage and outcome. Seat pricing rested on a quiet assumption — software is used by people, so you bill by how many people use it. But when an AI agent does the work of ten employees, the “person” you were pricing against starts to dissolve. If one agent replaces ten seats, do you sell ten seats or one? The anchor is melting.

That’s why seat-based incumbents are scrambling to open a second, seat-independent revenue curve — Salesforce with Agentforce (billed by consumption/action), Microsoft exploring multiple ways to charge for Copilot. It’s a life-and-death transition: if the new usage/outcome revenue grows fast enough to hedge or exceed shrinking seats, they cross the revolution and may come out stronger; if it’s too slow, they fall into the gap between old seats shrinking and new usage not yet risen. Companies already built on usage — Snowflake (~$4.7B), Datadog (~$1B/quarter), Twilio (~$1.4B/quarter) — don’t have to rebuild the billing pipe, but they only benefit if AI actually drives up the workloads they meter.

The subtler casualty is a metric this book leaned on for sixteen chapters: NRR. Be precise here, because it’s easy to get wrong. NRR is, and always has been, a revenue/dollar-retention metric — how much recurring revenue the existing customer base contributes a year later. There is no standard “seat-basis NRR”; seat retention is a different thing that should be disclosed and named separately. AI doesn’t break NRR — it changes its driver structure. A customer who cuts labor seats and adds AI usage can push NRR up (if usage revenue more than replaces seat revenue) or down (if it doesn’t), and a high NRR might now come from healthy agent value capture or from volatile consumption. Same number, completely different textures underneath.

None of this means seats die tomorrow — plenty of software is still bought by the head, and will be for years. The point is narrower and sharper: the assumption that seats are the natural unit of software value is no longer safe, and any company whose growth story rests on selling more seats into a shrinking human workforce is telling you a story with an expiration date. Watch which side of the transition each company is actually building toward.

So “re-underwriting NRR” doesn’t mean abandoning it — it means reading it alongside seats, usage, price, mix, GRR, and margin, and checking whether the company quietly changed its basis. The seat- and NRR-centric toolkit we built for the subscription era now has to evolve for the agent era. That’s the real, uncomfortable methodological challenge AI poses to every SaaS analyst.

This is Chapter 17 of “Compounding in the Cloud.”