AI agent platforms price four ways: per seat, per task or action, per credit consumed, or per outcome achieved. Seat pricing punishes team growth and is misaligned with agent work; task pricing punishes exactly the high-volume workflows you most want to automate; credit pricing tracks work done and is usually the fairest for small teams; outcome pricing is the most aligned and the rarest. Model your real monthly volume against each before comparing headline prices, because the unit matters far more than the number.
Two platforms both say $50 a month. One costs you $50 and the other costs $900, because the unit underneath is different. The headline number in this category is close to meaningless without knowing what you are being charged for.
The four models
| Model | Charged for | At 10x volume | Aligned with your interest? |
|---|---|---|---|
| Per seat | Each human user | Unchanged | No: punishes team growth, not usage |
| Per task / action | Each step executed | 10x cost | Poorly: penalises the best workflows |
| Per credit | Work the agent performs | ~10x, but tracks value | Reasonably |
| Per outcome | Results delivered | Scales with results | Yes: rare because it's hard to define |
Per seat: the model that shouldn't exist here
Seat pricing is inherited from SaaS, where a seat approximates a person doing work. For agents that logic breaks: the point is that output no longer tracks headcount. Charging per human for software whose whole premise is doing work without humans is incoherent, and in practice it means you ration access: the ops person who would benefit most does not get a login.
It survives because it is predictable and finance teams like it. If you buy a seat-priced platform, budget for the fact that you will under-deploy it.
Per task: the trap
This is Zapier's model and it is fine at low volume. The problem is structural: your most valuable workflows are the highest-frequency ones, so the model bills you most for exactly what you most want to do.
The second problem is that a "task" is usually every step in a run, not every run. A seven-step workflow firing 500 times a month is 3,500 tasks. Teams routinely underestimate this by an order of magnitude when modelling.
Per credit: the current default
Credits are consumed as agents do work: reading a document, drafting a message, calling a tool. Cost tracks activity, an agent that ran and accomplished nothing costs almost nothing, and scaling is roughly linear.
The catch worth knowing: credits are not a standard unit. One platform's 20,000 credits and another's are not comparable in any way, because consumption rates differ by an order of magnitude. The only useful comparison is running the same real workflow on both for a month, unless the vendor defines the unit outright, which is still the exception rather than the rule.
Operater prices by action and defines the unit: one credit is one thing an agent does (one search, one draft, one message sent), and each of them appears in the activity log, so the bill is countable rather than inferred. Free is 150 actions a month with no card, then $39 for 400, $199 for 2,000 and $599 for 6,000, with unlimited seats, connections and spaces, and every agent available on every plan including the free one. The unlimited seats part is deliberate: under a consumption model there is no reason to charge for people, and rationing logins defeats the purpose.
Per outcome: the honest ideal
Charging per meeting booked, per ticket resolved, per qualified lead. Perfectly aligned, and rare, because defining an outcome cleanly is genuinely hard. Is a meeting that no-shows an outcome? A lead the agent qualified that sales disqualified?
Expect more of this over the next couple of years, mostly at the enterprise end where a single outcome is worth enough to make the accounting worthwhile.
How to model it in twenty minutes
- Write down your actual monthly volume for the workflow you care about: leads, messages, tickets. Not aspirational volume.
- Multiply by 10. This is the number that matters, because if the agent works you will use it more.
- Ask each vendor what that volume costs, in writing, with their consumption assumptions stated.
- Ask what happens when you exceed the plan: overage rate, hard stop, or forced upgrade. Hard stops mid-month are their own kind of expensive.
- Check whether seats are counted. If they are, add the people you would want to give access to but wouldn't.
One number worth keeping
Cost per completed outcome: a qualified lead, a resolved ticket, a published piece. It is the only figure comparable across pricing models, and it is the one that tells you whether the platform is working. If you cannot compute it after a month, the platform is not giving you enough visibility, which is its own answer.
For how this compares to the cost of doing the work with people, see AI agents vs hiring.
Key takeaways
- Per-seat pricing makes no sense for agents: the whole point is that output stops tracking headcount.
- Per-task pricing gets most expensive precisely where automation is most valuable.
- Credit pricing is usually fairest for small teams, but credits are not comparable across platforms.
- Outcome pricing is the most aligned model and rare because it's hard to define an outcome cleanly.
- Always model 10x your current volume. Every model looks affordable at low volume.