---
title: "How to Price an AI Agent: Four Models"
description: "How to price an AI agent: per seat, per action, per workflow or per outcome. When a result is verifiable enough to bill, and why flat seats break on heavy use."
canonical_url: "https://builderscamp.com/guides/other/how-to-price-an-ai-agent"
date_published: "2026-09-26"
date_modified: "2026-09-26"
author: "Andre Albuquerque"
publisher: "Builders Camp"
guide_class: "other"
---

# How to Price an AI Agent

**TL;DR:** Price an AI agent in the unit the customer can check: per outcome when the result is observable and attributable, per workflow when the job is standard, per action when usage is spiky, and per seat only when work per user is flat. Most teams end on a hybrid, a predictable base fee plus a metered or outcome component, because a flat seat loses money on your heaviest users.

## What are the four ways to price an AI agent?

The four models are per agent, per action, per workflow and per outcome, and the choice is harder than it looks: when Manny Medina launched Paid, 130 of 175 signups (75 percent) said "I am not sure how to price my AI features," as he wrote in [a guest post on Kyle Poyar's Growth Unhinged](https://www.growthunhinged.com/p/ai-agent-pricing-framework). His framework, drawn from 60+ AI agent companies, names the four models and where each one breaks. The table below restates it in plain terms.

| Model | What the customer pays for | Budget it draws from | Where it breaks |
|---|---|---|---|
| Per agent (seat) | A fixed monthly fee per agent or user | Headcount or software seats | Heavy users cost far more to serve than they pay |
| Per action | Each discrete step the agent takes | Usage or outsourcing spend | Easy to compare and undercut; prices only fall |
| Per workflow | Each completed multi-step job | Process or operations budget | A long, messy run can cost more than its flat price |
| Per outcome | A defined result, such as a resolved ticket | The budget for the result itself | Needs a result both sides can see and attribute |

The four are not a ladder from worse to better. Each one moves a different risk between you and the buyer: per seat puts the usage risk on you, per action puts it on the buyer, per workflow splits it, and per outcome puts the performance risk on you. Pick the model by deciding which risk you can carry.

## Why do flat seat prices break on heavy users?

A seat fixes revenue per user, while an agent's cost to serve grows with how much work each user sends it. For a human-driven SaaS tool the gap between a light and a heavy user is small. For an agent it can be enormous, because one user can ask for a long, multi-step task that burns many times the tokens of a simple lookup.

Cursor's June 2025 pricing change is the clearest public example. Its Pro plan had included 500 requests a month; it moved to $20 of included frontier model usage at API pricing. In the follow-up post, [Clarifying our pricing](https://cursor.com/blog/june-2025-pricing), Michael Truell wrote that new models "can spend more tokens per request on longer-horizon tasks" and that "the hardest requests cost an order of magnitude more than simple ones." The same post said the pricing changes "were not communicated clearly" and offered refunds for unexpected usage charges between June 16 and July 4, 2025.

Two lessons come out of that episode, and they point in different directions. The economics were right: a request count stopped measuring cost once requests varied tenfold in size. The rollout was wrong: heavy users found out from their bills. If you move off seats, publish what typical and heavy users will pay before the switch, not after.

## When is an outcome verifiable enough to charge for?

An outcome is billable when it passes four tests: you can observe it in data both sides trust, you can attribute it to the agent, it settles inside a billing period, and the agent cannot reach it by a shortcut that costs the customer more than the fee. Fail any one and you will spend your margin on disputes.

Intercom's Fin is a working example of a narrow, observable definition. [Intercom's Fin pricing](https://fin.ai/pricing) charges $0.99 per outcome and defines a resolution as "No further help is requested after Fin's last answer," billing at most one outcome per conversation even when Fin takes several actions. That definition is observable and quick to settle. It is also a proxy: a customer who gives up and leaves requests no further help either, so a vendor using this kind of definition should watch complaint and churn data alongside the resolution count.

The shortcut test catches the most expensive mistakes. Imagine a travel-booking agent billed per "rebooked trip" that learns it can close almost any cancellation by issuing an unrestricted voucher. Every one of those closes is a real rebooking in the logs, and every one costs the customer more than the agent's fee. Write the exclusions into the outcome definition up front: an outcome reached by a refund, a credit or a discount above a set threshold does not count.

Attribution is the other trap. Medina's framework warns that outcome pricing "requires confidence in your agent's ability to consistently deliver those outcomes," and that outcomes "may also be subject to attribution, like with AI SDRs," where a meeting booked might owe as much to the salesperson as to the agent. If you cannot show the agent caused the result, through a holdout group or a pilot comparison, bill per workflow and report the outcomes instead.

## Seat or resolution pricing: which earns more?

The numbers below are illustrative, invented for this guide to show the mechanics, and do not describe any real vendor. Assume a customer support agent that costs $0.15 to serve per conversation, resolved or not, sold either at $120 per human support seat a month or at $0.90 per resolution. Each customer has 10 support seats.

| Customer (illustrative) | Conversations a month | Resolution rate | Seat revenue | Resolution revenue | Cost to serve | Seat margin | Resolution margin |
|---|---|---|---|---|---|---|---|
| Typical | 4,000 | 50 percent | $1,200 | $1,800 | $600 | $600 | $1,200 |
| Heavy | 30,000 | 50 percent | $1,200 | $13,500 | $4,500 | minus $3,300 | $9,000 |
| Poor fit | 4,000 | 20 percent | $1,200 | $720 | $600 | $600 | $120 |

Read the table by row. On the typical customer both models make money. On the heavy customer, seat pricing loses $3,300 a month, the same shape of problem Cursor described, while resolution pricing grows with the work. On the poor-fit customer, where the agent rarely resolves anything, the seat price keeps earning and resolution pricing barely covers cost.

That last row is the part vendors skip. Outcome pricing moves the risk of a weak agent from the buyer to you, which is exactly why buyers like it. Only offer it on the segments where your eval data says the agent resolves reliably, and keep a seat or workflow price for the rest.

## What does Salesforce's Agentforce pricing history show?

Salesforce has changed its agent pricing unit in public. Its [May 15, 2025 press release](https://www.salesforce.com/news/press-releases/2025/05/15/agentforce-flexible-pricing-news/) said "thousands of organizations" had used conversational pricing "at $2 per conversation," and introduced Flex Credits at $500 per 100,000 credits, with one Agentforce action consuming 20 credits, or $0.10 per action. The same release added a Flex Agreement that lets customers convert user licences into credits and back.

The move from a conversation to an action is a move to a finer unit: a short conversation and a long one no longer cost the same. The licence swap is a hedge in the other direction, keeping a seat-shaped option open for buyers who want one. In the release, Sheryl Kingstone of 451 Research said: "While the majority of businesses still prefer to purchase technology using a license model, there is a rising desire for pricing that's aligned to outcomes." That sentence is the tension every agent pricing decision has to settle.

## What is the strongest argument against outcome pricing?

Predictability. A finance team can approve a seat budget in one meeting; a bill that scales with resolutions, actions or workflows needs a forecast, a cap and someone watching usage. Kingstone's line above says most buyers still prefer licences, and Medina's framework adds that outcome contracts can multiply into bespoke agreements, one definition per customer.

The honest answer is to keep a predictable layer and put the variable part on top of it. A platform or seat fee with included usage, a metered or outcome charge above the allowance, and a monthly cap the buyer sets gives finance a number to approve and keeps your revenue tied to the work. Intercom's published pricing works this way, with per seat helpdesk plans and a per outcome charge for Fin on top.

## How do you choose a pricing model for your agent?

Start from what the customer pays today for the same work, not from your model costs. If the agent replaces outsourced tickets, the buyer already thinks in a price per ticket. If it replaces part of a role, they think in salary. If it speeds up a tool they already license, they think in seats. Pricing in the unit the buyer already uses shortens the sale.

Then run three checks before you publish a price:

- **Cost at the heavy tail.** Model your cost to serve for your top 10 percent of users, not the average. If a flat price loses money there, meter above an included allowance.
- **Outcome verifiability.** Put your proposed outcome through the four tests above. If it fails attribution or the shortcut test, bill per workflow and report outcomes on the invoice.
- **Change cost.** Write down how you would move a heavy user to a new price without a surprise bill. If you cannot, the model is not ready to ship.

## Where do you learn to make these calls with evidence?

Pricing an agent sits on top of two other product decisions: what the agent is allowed to do, covered in [how to design an AI agent](https://builderscamp.com/guides/tools/how-to-design-an-ai-agent), and how you know it did the job, covered in [how to write evals for AI products](https://builderscamp.com/guides/tools/how-to-write-evals-for-ai-products). The eval set is what lets you promise an outcome with a straight face. For the general framework underneath all of this, see [pricing strategy](https://builderscamp.com/guides/glossary/pricing-strategy) and [unit economics](https://builderscamp.com/guides/glossary/unit-economics), and for how the product type shapes all of it, [AI-native vs AI-enabled products](https://builderscamp.com/guides/other/ai-native-vs-ai-infused-products).

Builders Camp's AI Product Management bootcamp runs over 2 weeks and 4 live sessions, and covers capturing the value: how AI products sell finished work and price it against the labour budget rather than the software budget.

[See the AI Product Management bootcamp](https://builderscamp.com/bootcamps/ai-product-management?utm_source=guide&utm_medium=organic&utm_campaign=how-to-price-an-ai-agent)

## Frequently asked questions

### What are the four ways to price an AI agent?

Per agent (a fixed monthly fee, like a seat or a digital employee), per action (metered usage), per workflow (a flat price per completed multi-step job), and per outcome (paid only when a defined result happens). Manny Medina's framework in Kyle Poyar's Growth Unhinged, built from 60+ AI agent companies, uses these four.

### When should you charge per outcome?

Only when the outcome is observable in data you both trust, attributable to the agent rather than to a person or a discount, settled within a billing period, and hard to game. If any of those fails, charge per workflow or per action and report outcomes alongside the bill instead of billing on them.

### Why do flat seat prices break for AI agents?

A seat fixes revenue per user while the cost of serving that user scales with how much work the agent does. Cursor said in July 2025 that its hardest requests cost an order of magnitude more than simple ones, which is why it moved Pro from 500 requests a month to $20 of included usage at API prices.

### Is per action pricing a race to the bottom?

It can be. Medina's framework rates per action pricing as the lowest in competitive differentiation, because a buyer can compare your unit price directly with a rival's and model prices keep falling. It still fits spiky, unpredictable workloads where the buyer wants to pay only for use.

### What is a good first price for a new AI agent?

Start from the budget line you replace, not your cost. Work out what the customer pays today for the same work, in people, outsourcing or software, then price well below that while keeping a margin above your cost to serve a heavy user, not an average one.

### Should you mix pricing models?

Usually yes. Intercom pairs per seat helpdesk plans with a $0.99 charge per Fin outcome, and Salesforce's May 2025 Flex Agreement lets customers convert user licences into Agentforce credits and back. A hybrid gives the buyer a predictable base and keeps your revenue tied to the work the agent does.

### How should you announce a pricing change?

Before it takes effect, with worked examples of what typical and heavy users will pay. Cursor's July 2025 post apologised that its June changes were not communicated clearly and offered refunds for unexpected charges between June 16 and July 4, which is the cost of skipping that step.

## Sources

- [Manny Medina in Kyle Poyar's Growth Unhinged: A new framework for AI agent pricing](https://www.growthunhinged.com/p/ai-agent-pricing-framework)
- [Cursor: Clarifying our pricing (Michael Truell, July 4, 2025)](https://cursor.com/blog/june-2025-pricing)
- [Salesforce: New flexible Agentforce pricing (press release, May 15, 2025)](https://www.salesforce.com/news/press-releases/2025/05/15/agentforce-flexible-pricing-news/)
- [Intercom: Fin pricing](https://fin.ai/pricing)

## How this guide was made

Researched from Builders Camp's bootcamp, track and masterclass material and the sources listed on this page, drafted with AI, and fact-checked against every source cited.