Why Seat-Based Pricing Breaks the Moment You Buy an AI Agent

A seat is a beautifully dumb unit. You hire a person, you buy them a login, the vendor bills you. The math is so clean it built the entire SaaS economy: count heads, multiply, send invoice. For thirty years the seat tracked value because the seat tracked the work. One human, one chair, roughly one human's worth of output flowing through it.

Then the agent shows up and sits in the chair without sitting in the chair.

That's the whole problem in one image. An AI agent doesn't consume software the way a person does. It can do the work of three reps at 3 a.m. or do nothing for a week, and either way it occupies exactly one "seat" on the org chart of nobody. The unit you've priced around has quietly stopped measuring anything. Chargebee calls the underlying failure a "three-body problem," and it's the cleanest way to see why the old invoice falls apart.

The three-body problem, in plain terms

Three forces pull on agent pricing at once, and none of them holds still long enough to anchor a per-seat number (Chargebee, 2026).

First, the product changes fast. An agent that resolved 40% of tickets this quarter resolves 60% next quarter because someone improved the knowledge base, so "what you're paying for" moves under your feet.

Second, consumption varies wildly per user. A seat assumes everyone consumes about the same. Agents don't. One workflow fires a single model call, the next fires forty, and the bill behind the scenes swings with it.

Third, the infrastructure cost is volatile. Every agent action burns tokens, and token cost is a live wire, not a fixed line item.

Stack those three and the seat collapses. Agents break a goal into many steps, so pricing tied to inputs (seats, logins) stops tracking either the value delivered or the cost incurred. You end up charging for a chair nobody sits in, in front of a desk whose workload you can't predict.

Three models that actually fit the chair-less worker

The market didn't wait for a committee. Three pricing structures are already live, and each one solves a different version of the problem.

Outcome-based: pay per result. This is the model everyone points at, because the canonical example is so tidy. Intercom's Fin bills $0.99 per resolved issue, charged once per conversation no matter how many steps it took to get there (Intercom, 2026). The revenue line moves in lockstep with customer success, and it reads like a receipt your CFO can actually parse. The catch hides in the word "resolved." Fin counts a resolution only when the customer confirms the answer helped, or exits without asking again, and only when Fin actually answered a real query. Greetings don't count (Intercom, 2026). Define that boundary loosely and the metric becomes a thing both sides argue about on every renewal call.

Usage-based: pay per task or per credit. Here the vendor abstracts a mess of heterogeneous costs (model calls, retrieval lookups, tool calls) into one intuitive unit. N8N charges per 10,000 workflow executions. Clay runs a credit "burn table" where different actions cost different amounts of credit (Chargebee, 2026). The elegance is that revenue tracks your actual cost of goods automatically, so the vendor never sells a dollar of compute for ninety cents. The risk is sticker shock. A usage spike you didn't see coming arrives as a bill you didn't budget for.

Hybrid: a base fee plus usage on top. The fixed floor (a platform fee, a minimum commit, a bundle of included credits) covers the lights, and variable tiers ride above it. Relevance AI and Lovable both fold a credit allowance into a recurring plan (Chargebee, 2026). This kills the "blank check" anxiety that scares procurement off pure usage, and it turns overages into a clean upsell conversation instead of a surprise. Botch the packaging, though, and you've built a new way to annoy customers: the ones who blow past their limit and the ones who never touch it both feel cheated.

How to tell which one you're being sold

If you're the operator evaluating vendors, the pricing model is a forecast of your own bill, so read it that way. Three questions sort it.

How clearly can you attribute the output to a measurable outcome? If "the agent resolved this" is a clean, agreed-on fact, outcome pricing rewards both sides. If the win is fuzzy or shared with a human, $0.99-per-result becomes a fight over what counts.

How autonomous is the agent? A fully autonomous agent that closes the loop without a human maps neatly to outcome pricing. One that hands off to a person mid-task muddies whose result it was. Chargebee's rule of thumb: clear attribution plus high autonomy favors outcome pricing; lumpy, unpredictable workloads favor usage or hybrid (Chargebee, 2026).

How predictable is the workload? Steady compute per run is friendly to flat or outcome pricing. Spiky, unpredictable demand is exactly where usage and hybrid earn their keep, because they let the bill breathe with the work instead of guessing wrong twice a year.

The part nobody can lock in: the price keeps moving

Here's the line that should make every buyer and every vendor uneasy. Per a16z, the cost of an LLM at equivalent performance drops roughly 10x per year (a16z, cited in Chargebee 2026). Whatever an agent costs to run today, the floor under that number is falling fast. A price set against this year's infrastructure is mispriced against next year's. The cautionary tale is already written: Cursor had to drag users off unlimited plans onto usage caps after pricing below its true infra cost (Chargebee, 2026). Set it once and forget it, and the market resets it for you.

The seat survived for thirty years because the chair always held a person, and a person's output was knowable. The agent broke that contract the day it learned to work the night shift for free. Price the result, price the task, or price the floor plus the spillover, but stop pricing the chair. Nobody's sitting in it.