AI Agent Pricing in 2026: Per Seat, Per Task, or Per Skill?

AI agents are sold three ways in 2026: per seat, per task or per skill. Per seat is predictable but punishes teams who add people faster than they add work, per task is fair in theory but makes your invoice a function of a number you cannot forecast, and per skill ties cost to the number of jobs the agent does, which is the only one of the three that a buyer can plan a budget around.

Why is AI agent pricing so hard to compare?

Because the three models are not priced on the same axis, so the headline number is not comparable. A 20 dollar per seat product and a 500 dollar per month product are not 25 times apart in cost, and which one is cheaper depends entirely on facts about your team that the pricing page does not know. The comparison has to be normalized to cost per outcome before it means anything.

How does per seat AI pricing work, and who does it suit?

Per seat charges by the number of people with access. It is inherited from the SaaS era because it was easy to meter and easy to forecast. It suits products where every user gets value personally, such as a writing assistant. It breaks for agents, because the value of an agent is the work it completes, and work does not scale with headcount.

The pathological case is a ten person team where two people ever invoke the agent, paying five times what they use. The opposite case is a large team where the agent quietly becomes essential and the bill grows with hiring rather than with usage.

How does per task or usage based AI pricing work?

Per task charges by units consumed, whether that is a task, a run, a credit or a token. It is the most honest reflection of underlying cost, since inference genuinely is consumption. Its real problem is forecasting. Buyers cannot estimate task counts before they have run the system, and the number moves as the agent gets used more, which means the invoice grows exactly when the product is succeeding.

There is a second order effect worth naming: teams start rationing usage to protect the budget, which caps the value of the thing they bought. Watch for the credit repackaging pattern too, where credits do not map cleanly to anything a buyer recognizes, making the effective rate hard to derive.

How does per skill pricing work?

Per skill charges by the number of distinct, scoped jobs the agent is contracted to perform. A skill is a named job, such as triage the shared inbox or build the weekly report. Cost is therefore a function of scope, which is the one variable the buyer actually controls and can decide in advance. Unlimited invocation within scope means the team is never punished for using what they bought.

The honest limitation: it requires the vendor to define scope precisely, and a vendor who defines skills vaguely can inflate the count. The buyer protection is to insist that every skill is named in writing before signing.

Steward is a worked example, using only figures published on its pricing page: 500 dollars per month for one skill, 1500 for up to three, 2500 for up to five.

Which pricing model is cheapest for a small team?

For a team under fifty with a handful of recurring jobs, per skill is usually lowest total cost because the bill does not move when headcount or usage moves. Per seat wins when access is broad and usage is light. Per task wins only when volume is genuinely low and genuinely stable, which is rarer than buyers expect.

What questions should you ask before signing any AI agent contract?

• What exactly is metered, and can I see a sample invoice?
• What happens when I exceed the plan - overage or hard stop?
• Is there a minimum term?
• Who owns the configuration if I leave?
• What is explicitly out of scope?
• Is there a price change clause, and how much notice does it require?

Closing

Pricing model is a proxy for who absorbs uncertainty. Per task puts forecast risk on the buyer, per seat puts it on the buyer’s org chart, per skill puts it on the vendor, because the vendor commits to an outcome for a fixed fee. That is the real thing being priced.

Related reading: see how Steward’s pricing works in practice, how it compares to other approaches, and whether a managed AI agent fits a small business.