One AI agent can do the work of five to twenty human seats, which means per-seat contracts shrink even when platform usage and delivered value climb. That arithmetic is already landing in 2026 renewal cycles, and founders who wait for the customer to raise it will lose the deal.

Here's the mechanic breaking the model. A per-seat contract assumed one human, one login, one desk. An AI coding agent doesn't work that way. It runs constantly, spins up parallel sessions, and completes work that used to require three or four engineers logging into three or four seats. Cursor, GitHub Copilot, and Devin get deployed this way inside engineering teams right now. The seat count drops. The value delivered doesn't.

Per-seat pricing held for two decades because seats were a decent proxy for value. More salespeople on Salesforce meant more pipeline. More agents on Zendesk meant more tickets closed. Seat count and outcome moved together, so charging per seat was lazy but functional. AI agents snap that link.

GitHub reported in 2025 that teams using Copilot's agent mode completed pull requests with fewer named contributors touching the repo, because the agent handled boilerplate, tests, and first-pass reviews on its own. The team didn't shrink. The seat count did. If you're the vendor selling into that team, your contract just took a haircut with no connection to how much value you delivered — the customer runs more automated work through your platform than ever, and you bill them less because you counted logins instead of output.

What the leaders already did

Intercom moved its Fin AI agent to resolution-based pricing, charging per conversation the AI resolves rather than folding it into a seat tier. Zendesk followed with a similar structure, pricing its AI agents on resolutions instead of bundling AI into the per-agent license. Both companies said publicly that seat-based pricing couldn't capture the value once a bot handled a ticket end to end.

Sierra, the AI agent company founded by Bret Taylor, skipped seat pricing entirely: it charges per resolved customer issue, full stop. No seat to count, because no human seat does the work. That's the cleanest version of the new model, and it marks the ceiling — when the agent does the whole job, seat pricing isn't a compromise, it's just wrong.

HubSpot took a middle path, keeping core per-seat CRM pricing while layering AI credits on top that its Breeze agents consume as they work. It's a hedge, not a migration, and it tells you something: even a company built for twenty years on seat logic won't abandon it outright. It's running a second meter next to the first.

Don't just switch to usage-based pricing

The instinct, once you see the problem, is to junk seats entirely. Resist it. Usage-based versus per-seat isn't binary, and a pure consumption model creates its own failure modes: unpredictable bills scare procurement teams, revenue gets harder to forecast, and customers who liked a predictable monthly number start shopping around.

What works better is a blended structure with three parts.

First, keep a seat floor, but make it smaller and cheaper. Price the human seat for what a human does now: reviewing, approving, directing the agent. A $99 seat that used to be $199 isn't a discount — it's an honest reflection of a narrower job.

Second, meter the agent separately, and tie the meter to something the customer can see and verify: resolved tickets, merged pull requests, completed workflows.

The third part follows from the second: the mismatch surfaces first at renewal, so repricing has to happen before the customer brings it up.

The tradeoffs are real on both sides. Outcome-based pricing transfers volume risk to the vendor — if resolutions spike, your margin absorbs it. Usage meters shift that risk back to the buyer, which is exactly why procurement pushes back. The blended model exists to split the risk, at the cost of running two pricing engines and explaining both to every prospect.

Sierra's per-resolution model is the clearest signal of where the market is heading: as agents take over complete jobs, pricing will converge on what the work produces, not who logged in to do it.