---
title: "Should You Buy AI Per Seat, Per Token or Per Outcome?"
url: "https://goremy.ai/blog/articles/ai-per-seat-vs-per-token-vs-outcome-pricing"
author: Luis Chavez-Mattos
editor: Luis Chavez-Mattos
editor_title: Director, Product
published: "2026-10-26T22:00:08.348Z"
updated: "2026-10-26T22:00:11.776Z"
topic: AI Pricing & Usage Limits
description: "Seat, token and outcome pricing for enterprise AI compared, with Cedar, Sierra, forward-deployed engineering and OpenAI's Thrive deal, plus how to measure ROI."
---

# Should You Buy AI Per Seat, Per Token or Per Outcome?

> Seat, token and outcome pricing for enterprise AI compared, with Cedar, Sierra, forward-deployed engineering and OpenAI's Thrive deal, plus how to measure ROI.

Edited by Luis Chavez-Mattos, Director, Product · Published 2026-10-26

## Should a company buy AI per seat, per token, or per outcome?

It depends on how well you can already value the work the AI does. Per-token pricing suits experimentation and capped budgets. Per-seat pricing suits tools individuals use directly. Outcome pricing suits work with a clear, countable result, like a paid medical bill. But the meter matters less than the question behind it: do you know how to turn purchased intelligence into return? Most procurement contracts leave that burden on the buyer.

## TL;DR

- **Tokens meter compute, not value.** A token count tells you what you consumed, not whether any of it was useful.
- **Seats are a legacy SaaS unit.** They make sense for tools people operate, but they say little about the economic value of AI agents doing work.
- **Outcome pricing shifts risk to the vendor,** but it is hard to measure, and nobody expects every deal to move there soon.
- **The bigger gap is valuation.** A competitive price does not produce ROI if the organization has not learned which work to give the AI.
- **Forward-deployed engineering** is one way vendors help buyers turn capability into working systems, and it can grow into deeper partnerships like OpenAI's stake in Thrive Holdings.
- **Employee fluency and motivation** decide returns, because frontline people choose when to use AI and when to hand off to a human.

## Why is the seat vs token debate the wrong question?

Many enterprise AI deals stall on pricing. Procurement asks why a vendor offers seats, or why the meter is tokens. Those are reasonable questions, but they treat AI as a familiar product category.

One framing: companies have long had two economic inputs, people and capital. AI adds a third, intelligence. A token is a unit for metering what a model processes. A seat is a traditional software measure. Neither describes what a buyer actually wants, which is useful work done.

Token pricing has real benefits. It lets you forecast and cap spend, and it suits a company that wants to experiment without deep downside. The weakness is that a token does not carry a fixed amount of useful thinking. Whether it paid off depends on how your people used it. Under a standard contract, the buyer holds that assumption: you pay, and you had better capture enough value to justify the cost.

## How do the three pricing models compare?

| Model | What you pay for | Where the risk sits | Main weakness |
|---|---|---|---|
| Per seat | Access for each user | Buyer, if people don't use it well | Says little about value created, especially for agents that work without a user |
| Per token | Compute processed by the model | Buyer, who must turn usage into value | Predictable spend, but no link to results |
| Per outcome | A delivered business result | Shared, and more on the vendor | Hard to define and measure |

None of these is wrong. Each answers a different question. Seats and tokens answer "what did we consume?" Outcomes answer "what did we get?"

## What does outcome-based AI pricing look like in practice?

Cedar, a healthcare payments company, runs an AI agent called Cora. A patient can call about a medical bill at 2:00 in the morning. Cora verifies who they are, walks through the charges, confirms insurance, and lets them pay on the call.

The hospital does not care how many tokens that call used. It cares that the bill got paid. That is the logic behind outcome pricing: charge for the result the customer values.

Sierra, another agent vendor, is actively exploring per-outcome pricing. According to the source, it has found deals where it can price per outcome and both sides are comfortable with it. That is early evidence, not a market shift. Outcome pricing is hard to measure, and you need agreement on what counts as an outcome before a contract can use it.

## How do you measure the return on AI spend?

Start by admitting that a procurement process can only establish that you bought intelligence at a competitive price. It cannot show the investment case. That depends on the people using the AI.

A few practical checks follow from that:

1. **Name the work.** Decide which tasks the AI is assigned before you scale spend, rather than buying capability and hoping use cases appear.
2. **Define a good answer.** Someone must be able to tell when output is right. A model can look impressive in a demo while the organization is poorly equipped to use it.
3. **Change the surrounding process.** Dropping an agent into an unchanged workflow rarely captures value.
4. **Track the result, not the meter.** Tie spend to a business metric such as bills paid, issues resolved, or failures prevented.

The same logic applies to individual work. Take a report you produce regularly. You can ask an agent to finish it faster, or you can ask it to investigate the root cause of a customer problem and start fixing it. The second request treats AI as something to manage, not just use.

## Why does employee fluency matter more than the meter?

Traditional inputs are managed from above. A leader assigns people and budget, and the work follows. AI is different because everyone applies it. Frontline engineers and customer-facing staff decide when to use AI, when to dig deeper, and when to stop and talk to a human.

Consider a hypothetical equipment company, described as close to real cases. An engineer spots a recurring failure. An agent pulls maintenance records and compares past cases. But the engineer knows which operating conditions matter, which records are unreliable, and when an apparent pattern is really two faults described in the same language. Sending the agent off to generate more explanations would not fix anything. The engineer has to decide where the machine can handle the work and where a person must step in.

If enough cases follow that path, the team might see a repeatable opportunity. Customers could pay for ongoing advice that prevents failures. That is a new product, and it came from people who felt equipped and motivated to pursue customer value. Pricing the meter correctly would not have produced it. Companies that want this need staff who believe the company backs them and who have skin in the game.

## What is forward-deployed engineering?

[Forward-deployed engineers](https://goremy.ai/blog/articles/forward-deployed-engineering-openai-anthropic-ai-roi) work inside or closely alongside a customer's organization to turn technology into functioning systems. They help a buyer take intelligence it has purchased and convert it into value.

OpenAI's deployment company describes its engagements this way: engineers work with business leaders and frontline teams to identify opportunities, then design, build, test and deploy systems connected to the customer's data and processes. The work starts with choosing where intelligence applies and continues until it pays off in daily operations. Anthropic uses similar language for its forward-deployed work and for a joint venture announced with Blackstone, Hellman & Friedman and Goldman Sachs, aimed at businesses such as mid-size manufacturers and regional healthcare systems.

There is a balance to strike. Keeping forward-deployed engineering as a service forever means the buyer may not learn enough. Pushing it as a product too early can have the same problem.

## When does a pricing deal become a partnership?

Sometimes the opportunity is large enough to justify more than a contract. OpenAI took an ownership stake in Thrive Holdings and said it would embed research, product and engineering teams inside Thrive's companies. The initial focus is accounting and IT services.

The point is that an AI supplier may hold expertise that goes well beyond token pricing. Call it outcome pricing or a joint venture, but the structure changes the CFO's job. The question stops being how large a software discount procurement can win. It becomes how much more value a partnership could create, what the firm would retain, what it must commit, and whether the upside justifies the trade.

Not every company will do this, and the world will not pivot to outcome pricing overnight. But with intelligence as a new economic input, some experimentation with deal structure seems reasonable.

## Is capping AI spend the right call?

Constraining spend is understandable when returns are unproven. The argument in the source is that many companies are underinvested. The hard question is what to invest in, which is a question of ambition and agency: how to empower employees to make the right bets at scale and keep them aligned. A model cannot answer that for you.

## Frequently Asked Questions

### Is per-token pricing bad for enterprises?

No. Tokens give you a unit to forecast and cap spend, which helps while you experiment. The limit is that tokens do not guarantee useful output, so the buyer still has to turn usage into value.

### Why is outcome-based AI pricing not more common?

Outcomes are hard to define and measure. Both sides must agree on what counts as a result, and the vendor takes on more risk. Sierra is exploring per-outcome deals, but the source does not suggest this is the default yet.

### When does outcome pricing make the most sense?

It fits cases with a clear result the buyer already values, like Cedar's agent getting a patient's medical bill paid on a call. The hospital cares about the payment, not the compute behind it.

### What should a CFO ask besides price?

Ask how the company will allocate intelligence to specific work, how it will tell when answers are good, and what return it expects. For larger opportunities, ask what a deeper partnership could add, what the firm keeps, and what it must commit.

### Does forward-deployed engineering replace internal AI skills?

No. It helps a buyer turn purchased capability into working systems, but relying on it forever risks the buyer not learning enough. Internal employees still decide where AI applies day to day.