The per-seat license is one of the great business models in history: simple to understand, easy to forecast, and it grew automatically as the customer's headcount grew. For twenty years, nearly every enterprise software company — Salesforce, ServiceNow, Workday, Slack — ran on some version of it. AI introduces an awkward question: if the software resolves the support ticket itself, what exactly is the seat?
The question is no longer hypothetical. Vendors shipping AI agents that handle customer service, IT requests and back-office processing are discovering that their most successful features can reduce the number their revenue multiplies against.
Why it matters
The direct effect lands on customers' budgets and vendors' revenue models simultaneously. A company that automates a third of its support volume may need a third fewer agents — and under per-seat pricing, the vendor's revenue from that account falls by a third as reward for the software succeeding. Every major SaaS vendor now faces some version of this arithmetic.
The indirect effect is on how software gets bought. Per-seat pricing made procurement simple; outcome-based pricing requires agreeing on what an outcome is, measuring it, and auditing it. That shifts power toward buyers who can define value precisely, and toward vendors whose products produce measurable results rather than ambient productivity.
How it works
The replacement models fall into four families. Consumption pricing charges for usage — API calls, compute, tokens — and dominates the infrastructure layer, where AWS, Azure and the model providers have always billed this way. Per-action pricing charges per completed unit of work: a resolved ticket, a processed invoice, a booked meeting. Outcome pricing goes further and charges for the business result, typically shared savings. Hybrid models keep a platform fee for the software and meter only the AI labor on top.
Each model has a failure mode. Consumption pricing makes costs unpredictable for the buyer. Per-action pricing invites disputes over what counts as resolved — the vendor claims a resolution, the customer's reopened-ticket rate disagrees. Outcome pricing requires attribution that most organizations cannot measure cleanly. And hybrids, the current market favorite, preserve predictability at the cost of complexity: the contract now has two meters instead of one.
Evidence
The public moves are accumulating. Salesforce introduced per-conversation pricing for its Agentforce autonomous agents, charging per resolved interaction rather than per user. Intercom's Fin AI agent has charged per resolution from launch, with the company publishing its resolution-rate methodology. Microsoft has layered consumption-based Copilot agents on top of its per-user Copilot licenses — the hybrid model in its purest form. On the infrastructure side, every major model API has always been metered per token.
The financial signal appears in how vendors talk to investors: pricing-model discussion has become a standard feature of SaaS earnings calls, with analysts probing whether AI revenue complements or cannibalizes the seat base. Industry surveys of software CFOs consistently rank pricing-model redesign among their top concerns, which is a remarkable thing to say about the business model that funded two decades of growth.
The competing read
One camp argues the seat survives because humans remain accountable: even with agents doing the work, someone supervises, and that someone needs the software — so the seat count shrinks less than feared while the price per seat rises to absorb the AI value. The history of technology transitions supports a version of this: spreadsheets did not reduce finance headcount; they raised expectations per person.
The other camp notes that the seat survived previous transitions because the human was still the unit of work. If agents genuinely become the unit of work, pricing tied to humans is pricing tied to the wrong thing, and the only question is which replacement wins. The honest reading is that both are happening in different segments simultaneously — which is exactly why the transition is so messy.
What happens next
Expect two to three years of experimentation before consolidation. The near-term tells: whether per-resolution pricing survives contact with procurement departments at scale, whether vendors start publishing resolution-rate methodologies as a competitive weapon, and whether a standard emerges for auditing agent work the way uptime SLAs standardized cloud reliability. Buyers should insist on one principle through the churn: the meter should measure something they can verify independently.