AI agents are billed three ways. Per-seat is a flat monthly fee per user — Microsoft 365 Copilot is $30 per user per month. Per-outcome charges only when the agent resolves something — Intercom's Fin is $0.99 a resolution. Per-token meters the model's input and output. Match the model to your volume, not the vendor's pitch.
Why the model matters more than the sticker price
We already broke down how much an AI agent costs — the sticker range runs from a few dollars a month to six figures. This is the other half of that question. Once you're renting a prebuilt agent, the pricing model decides whether your bill tracks the value you get or just the volume you push through it. The same customer-service agent can be sold as a flat monthly fee, a fee per solved ticket, or a fee per word processed — and those three bills diverge fast as the agent gets busy. Here are the three models, with real vendor numbers, and who each one actually favors.
Per-seat: predictable, but you pay for access, not results
The oldest software model, applied to agents: a flat fee per named user. Microsoft 365 Copilot lists $30 per user per month (annual) as an add-on on top of a qualifying Microsoft 365 license. Salesforce Agentforce also offers a per-user edition starting around $125 per user per month. On the consumer end, Issue #001's Gmail agent runs on nothing but Claude Pro at $20/month — a per-seat plan by another name.
Per-seat is the easy row to budget: one predictable line item, no metering surprises, zero engineering. It favors heavy daily users and teams that want a fixed number. It penalizes the opposite case — a 50-seat license where only twelve people ever open the agent means you paid for 38 seats of nothing. The bill tracks access, not work done.
Per-outcome: the vendor wins only when the agent works
The fastest-moving model in 2026 is outcome-based: you pay only when the agent completes a defined job. Intercom's Fin charges $0.99 per resolution — a conversation the agent closes without a human stepping in. HubSpot moved its Customer Agent the same direction: now you pay per resolved conversation, at $0.50 (down from a per-conversation $1.00, effective April 2026), and escalations to a human are not billed. Salesforce Agentforce offers a $2 per conversation option alongside its credit model.
This is the one model where the vendor's revenue only grows when the agent genuinely works — the incentive points the same way you do. It suits customer-facing volume where a "resolution" is clean to define, and it maps directly onto the sort-vs-resolve line in support triage: you pay for resolutions, not for the sorting. The catch is in the fine print — read the vendor's definition of a resolution carefully (Fin, for instance, prices a matched sales lead at a much higher $9.99), because the definition is where the meter really lives.
Per-token or per-action: you pay for raw work
If you build the agent yourself against a model API, the meter is consumption. Anthropic's pricing is per million tokens: Claude Haiku at $1 in / $5 out, Sonnet at $2 / $10, Opus at $5 / $25 — output runs about 5× input, and batch processing plus prompt caching cut the bill by 50% and up to 90% respectively. No-code platforms meter a coarser unit: Salesforce's Flex Credits price actions at $500 per 100,000 credits, roughly $0.10 per action (voice actions cost more).
Per-token is the cheapest floor for a light agent — Issue #001's API-path Gmail triage ran about $3/month — and it gives you full control. But you pay for every token whether the run succeeded or not, and a chatty, badly-scoped agent can burn tokens on work that produced nothing useful. Consumption pricing rewards efficient prompting; it punishes sprawl.
Which model fits your volume
There's no cheapest model in the abstract — only the cheapest for your pattern. Building it yourself at low volume, or want maximum control? Per-token. Want one predictable number across a team of steady daily users? Per-seat. Running customer-facing volume where you can define a clean "resolution"? Per-outcome aligns the vendor's incentive with yours. Whichever row you land in, the real number is the running cost including human review, so measure it against a baseline rather than a launch-day quote — the same discipline we use for every real build we document.
FAQ
Do AI agents charge per seat, per resolution, or per token? All three, and the same agent can be sold under more than one. Per-seat is a flat fee per user (Microsoft 365 Copilot, $30/user/month). Per-resolution charges only on a completed job (Intercom's Fin, $0.99). Per-token meters the model's input and output (Anthropic's pricing). Pick by your volume pattern, not the label.
Which AI agent pricing model is cheapest? It depends entirely on volume. A light, self-built agent is cheapest per-token — a few dollars a month. A team of steady daily users often comes out ahead on a fixed per-seat plan. High-volume customer support is usually cheapest per-resolution, because you stop paying for the conversations a human ends up handling.
What counts as a "resolution" in outcome-based pricing? The vendor defines it, so read the definition. HubSpot counts a conversation as resolved when the agent handled it and it wasn't passed to a human within 72 hours. Intercom's Fin counts a resolution when the customer doesn't ask for more help after the agent's answer. Different definitions change the bill even at the same headline price.
Is per-token pricing only for developers? Mostly, yes — you pay per token when you build against a model API directly. Most no-code platforms hide the tokens behind a coarser unit (an action, a task, a workflow run, or a credit), which is easier to reason about but usually a markup over raw model cost. Issue #001 shows both paths side by side.
Why are vendors moving to outcome-based pricing? Because it's the model where the vendor only gets paid when the agent delivers, which is an easy story to sell a skeptical buyer. It also ties the vendor's revenue to the agent actually working — the opposite of a per-seat license that bills whether anyone uses it. The risk sits in the definition of the billable outcome, so that's the clause to negotiate.
We document real agents professionals actually run — with the real pricing, not demo-day math. Subscribe free for a new build each week.