AI agents are priced in five overlapping structures: per-seat (by user), usage-based (by run, token or call), outcome-based (by result delivered), credit-based (prepaid units), and hybrid combinations of the above. In 2026, hybrid and usage-based pricing dominate the market; outcome-based pricing, despite getting the most attention, is used by fewer than 1 in 25 AI agent companies.
Key takeaways
- Hybrid pricing is now near-universal: 95% of AI agent companies combine a base fee with a variable component, up from 92.4% a year earlier (Orb, 2026).
- Usage-based pricing climbed from 83.3% to 91.3% of agent vendors in the same period, while outcome-based pricing actually fell, from 4.5% to 3.8% (Orb, 2026).
- Per-seat pricing is shrinking across software broadly, not just agents: a 240-company survey found seat-based as the primary model dropped to 15% and hybrid rose to 41% in one year (Growth Unhinged, 2025).
- Uncapped usage pricing is a real budget risk: Gartner expects at least half of GenAI projects to overrun their budgeted costs through 2028, with inference running costs making up roughly 70% of a model’s lifetime cost (Gartner via Campus Technology, 2026).
- The model you pick should match how your task actually behaves — steady and predictable favors a flat or per-seat fee; spiky and occasional favors usage or outcome pricing.
What are the main AI agent pricing models?
There are five structures in active use, and most vendors now combine at least two of them rather than picking one.
The five models, side by side
| Model | What triggers a charge | Fits best when |
|---|---|---|
| Per-seat | A person logs in or has access | Usage is steady and the agent sits inside a tool your team already opens daily |
| Usage-based | A run, token, API call or compute-minute | Volume is unpredictable and you want the bill to track actual load |
| Outcome-based | A defined result: a resolved ticket, a booked meeting, a recovered dollar | The result is easy to define and verify, and worth paying a premium for |
| Credit-based | Prepaid credits drawn down per action | The underlying usage units (tokens, API calls) aren’t meaningful to a buyer on their own |
| Hybrid | A base fee plus one of the above on top | You want predictable minimum revenue and a price that still scales with heavy use |
How does usage-based pricing actually work?
A usage-based agent charges for a countable unit of work — most often a token, an API call, a full agent run, or a block of compute time — so the bill moves with how hard the agent actually works on a given task.
The mechanics matter more here than in simpler software because one task can trigger many billable units. An agent that reasons through several steps, calls a tool, checks its own output and retries before finishing one job can rack up ten or more model calls behind a single user request — all before anything gets billed as “one use.” That gap between what a customer thinks they’re paying for and what’s actually metered underneath is the single biggest source of AI pricing complaints in 2026, and it’s a large part of why usage-based pricing now gets bundled inside a hybrid plan rather than sold on its own: a base fee absorbs the unpredictability, and the usage meter only kicks in past an included allowance.
What is outcome-based pricing, and why is it rare?
Outcome-based pricing charges only when the agent delivers a defined, verifiable result — and it stays rare because that definition is harder to pin down than it sounds.
Zendesk’s own AI agent pricing is a working example of how narrow that definition has to get. Since May 2026, Zendesk bills $1.50 per resolution on a committed pack (or $2.00 pay-as-you-go) — but only for a “Verified Resolution,” where a separate model reads the full transcript afterward and confirms the customer’s issue was actually solved, not just that the conversation ended (eesel.ai, 2026). Building and trusting that verification step is the real cost of outcome-based pricing, and it’s why only 3.8% of AI agent companies currently price this way, down slightly from 4.5% the year before (Orb, 2026). It tends to show up where the result is unambiguous and valuable enough to justify the engineering: a recovered debt, a booked meeting, a closed ticket — not a vague “productivity” gain nobody can audit after the fact.
Is per-seat pricing dead for AI agents?
No, but it’s losing ground fast, and for a specific reason: an agent that works around the clock doesn’t map cleanly to “one seat, one person.”
A seat-based price assumes usage caps out at how much one human can click through in a workday. An agent that runs continuously, or scales to handle ten times the ticket volume without adding a person, breaks that assumption — which is why seat-based pricing fell to just 15% of companies as a primary model in a 2025 survey of 240 software and AI businesses, down from a larger share the year before, while hybrid pricing rose to 41% over the same stretch (Growth Unhinged, 2025). Per-seat pricing still works well for one case: an agent that genuinely augments one person’s existing workflow (drafting, summarizing, searching) rather than replacing a whole queue of work on its own.
How do you pick the right pricing model for your task?
Match the model to how the work actually behaves, not to which model sounds the most modern.
Four questions to answer before you sign anything
- Is the volume steady or spiky? Steady, predictable volume suits a flat fee or per-seat price; occasional or unpredictable spikes suit usage-based pricing so you aren’t paying for idle capacity.
- Can you define success in one sentence? If the result is as clean as “ticket resolved” or “meeting booked,” outcome-based pricing is worth asking a vendor about. If success is fuzzier than that, it isn’t — you’ll end up disputing what counts.
- Do you understand the billing units? If a vendor can’t explain, in plain language, what one “run” or “credit” actually covers, get that in writing before you commit, not after the first invoice.
- What’s your worst-case monthly bill? Ask for a cap or an alert threshold on any usage-based component. Gartner’s own research ties most GenAI budget overruns to exactly this gap between expected and actual usage, not to the headline price (Gartner via Campus Technology, 2026).
If the answer to all four questions is “the task is scoped and the volume is predictable,” a flat fee — our own AI Agent Kit, priced once instead of metered — is often simpler than negotiating any of the five models above. For a broader look at what these agents can take on day to day, see our piece on what AI agents actually do for a small business, and if build-vs-buy cost is the open question, our AI agent cost breakdown covers that side specifically.
Frequently asked questions
What are the five main AI agent pricing models?
Per-seat (by user), usage-based (by run, token or call), outcome-based (by verified result), credit-based (prepaid units), and hybrid combinations of the above. Hybrid is now the most common approach.
What’s the difference between usage-based and credit-based pricing?
Usage-based pricing bills the raw unit directly — a token, a call, a compute-minute. Credit-based pricing converts that same usage into a simpler prepaid “credit” so a buyer doesn’t have to reason about tokens or API calls to understand their bill.
Why is outcome-based pricing so rare for AI agents?
Because proving an outcome happened is harder than metering activity. A vendor needs a reliable way to confirm the result was genuine — Zendesk’s model, for example, uses a separate check on the transcript — and building that verification step is expensive, which is why only 3.8% of agent companies price this way.
Is per-seat pricing dead for AI agents?
Not dead, but declining. It still fits agents that augment one person’s existing work. It fits poorly once an agent replaces a whole queue of work that no longer scales with headcount.
What pricing model should a small business pick?
Start with how predictable the task volume is. Steady, bounded volume favors a flat fee or per-seat price. Spiky or occasional volume favors usage-based pricing with a cap. Ask for the billing-unit definition in writing either way.
Does TheMarkElite’s AI Agent Kit use one of these models?
No — it deliberately sits outside all five. It’s €49, paid once, with no seat count, no usage meter and no outcome clause to negotiate, for a team that already knows the task and wants to skip pricing it with a formula.
Where does a flat-fee kit fit in all this?
Nowhere on the five-model list above, and that’s the point, not an omission. Our AI Agent Kit is €49, once — not a sixth pricing model so much as a deliberate opt-out of all five, built for a team that has already answered the four questions above and wants one line on an invoice instead of a formula to recalculate as usage changes.
Last updated: November 15, 2026. Written by the TheMarkElite team.