Why Per-Token and Per-Seat AI Pricing Is Breaking Down
Enterprises are rethinking AI pricing beyond tokens and seats toward outcome-based deals and partnerships. Here's what's driving the shift.
Why is enterprise AI pricing suddenly a problem?
Enterprise AI pricing is breaking down because the two standard metering systems, per-token and per-seat, measure compute and headcount, not the value intelligence actually produces. Procurement teams know how to buy software; they don’t yet know how to buy intelligence. A hospital doesn’t care how many tokens an AI agent used to resolve a billing call. It cares whether the bill got paid. That gap between what’s metered and what matters is pushing vendors and buyers toward outcome-based pricing and deeper partnership structures instead.
TL;DR
- Token and seat pricing were built for compute and software licenses, not for valuing intelligence, which is why procurement teams keep getting stuck negotiating the wrong thing.
- Buyers carry the burden of proving ROI under current contracts: a company can pay a competitive price for AI and still have no idea how to turn it into returns.
- Healthcare payments company Cedar runs an AI agent called Cora that handles patient billing calls end to end, an example of intelligence creating value a token count can’t capture.
- Forward deployed engineering, where vendor engineers embed with a customer’s team to turn AI into working systems, is becoming a recognized path from raw model access to real returns.
- OpenAI and Anthropic are both moving past software licensing, with OpenAI taking an ownership stake in Thrive Holdings and Anthropic forming a joint venture with Blackstone, Hellman & Friedman, and Goldman Sachs.
- Sierra is actively exploring per-outcome pricing, charging based on delivered business results rather than usage volume, in deals where both buyer and seller are comfortable with that structure.
- Individual employees are becoming “managers of intelligence,” deciding in real time when to deploy AI, when to stop and bring in a human, and how that judgment turns into enterprise value.
Seven tools to build an app. Or just Remy.
Editor, preview, AI agents, deploy — all in one tab. Nothing to install.
What’s wrong with per-token and per-seat pricing?
Tokens measure how much text a model processes. Seats measure how many people have access to a tool. Neither measures whether the output was useful. That mismatch puts the entire burden of proving value on the buyer: a company can purchase intelligence at a perfectly competitive rate and still have no mechanism for knowing if it generated a return.
This isn’t an argument to scrap token pricing. It’s a reasonable way to budget and cap exposure, especially for companies still experimenting and wary of open-ended spend. The real issue is that token and seat metrics assume a fixed, knowable amount of value per unit purchased, and intelligence doesn’t work that way. Whether a given token of model output is worth anything depends entirely on what the organization does with it, what process it’s embedded in, and whether anyone downstream knows how to judge the answer. A model can look impressive in a demo while the surrounding organization remains completely unequipped to extract value from it.
That’s why so many companies report disappointing AI returns despite real capability gains. The contract structure never required anyone to figure out the hard part: discovering which work to hand off, how to change surrounding processes, and how to tell a good answer from a bad one.
How does outcome-based pricing actually work?
Outcome-based pricing ties payment to a result rather than to usage volume. Instead of billing per API call or per seat, a vendor charges based on a bill getting resolved, a support ticket getting closed, or some other defined business outcome.
Sierra, an AI agent company, is actively exploring this model and finding deals where both sides are comfortable pricing on a per-outcome basis. Cedar, a healthcare payments company, illustrates why this appeals to buyers: its AI agent Cora takes patient billing calls at any hour, verifies identity, walks through charges, confirms insurance, and gets the bill handled on the call. The hospital’s interest isn’t in metering that interaction. It’s in the bill getting paid. Pricing structured around that outcome aligns the vendor’s incentive with the result the customer actually wants.
Outcome pricing is harder to measure than tokens or seats, and it won’t replace usage-based billing everywhere. But as a new economic input (intelligence, alongside the traditional inputs of labor and capital) enters the picture, it makes sense that pricing experiments would follow.
What is forward deployed engineering and why does it matter?
Forward deployed engineering describes vendor engineers who work inside or alongside a customer’s organization to turn purchased AI capability into functioning systems, rather than just shipping a model or API key and walking away.
- ✕a coding agent
- ✕no-code
- ✕vibe coding
- ✕a faster Cursor
The one that tells the coding agents what to build.
OpenAI’s deployment arm describes its engagements this way: engineers work with business leaders and frontline teams to identify where AI can apply, then design, build, test, and deploy systems connected to the customer’s actual data and processes. Anthropic uses similar language around its forward deployed engineering work and its joint venture with Blackstone, Hellman & Friedman, and Goldman Sachs, which targets mid-size manufacturers and regional healthcare systems among other customers.
The appeal is that it closes the gap between buying intelligence and knowing how to use it. A company that purchases general-purpose model access still has to discover which tasks are worth automating, how to restructure workflows around it, and how to evaluate output quality. Forward deployed engineering puts people in the room who’ve done that discovery before. The tradeoff is that pure service delivery forever means the customer never builds internal capability, and rushing straight to a packaged product risks skipping the learning that makes the product valuable in the first place. The useful middle ground treats forward deployment as a bridge: build enough shared understanding that outcome-based or partnership pricing becomes viable later.
Why are AI labs taking equity stakes and forming joint ventures?
When an AI vendor’s expertise in a domain is deep enough, licensing fees stop reflecting the real value exchanged, and ownership stakes or joint ventures become the more honest structure.
OpenAI took an ownership stake in Thrive Holdings and is embedding research, product, and engineering teams directly inside Thrive’s portfolio companies, with initial focus on accounting and IT services. Anthropic’s joint venture with Blackstone, Hellman & Friedman, and Goldman Sachs follows a similar logic: rather than selling API access to financial and industrial firms, the lab becomes a structural partner in how AI gets applied across those businesses.
This matters for how CFOs should think about AI spend. A licensing negotiation is a procurement exercise: get the best per-token or per-seat rate. A partnership negotiation is a different question entirely: how much additional value could this relationship create, what would the company have to give up to get there, and is the upside worth the commitment on both sides. That’s not a discount conversation. It’s closer to an investment decision.
Is outcome-based AI pricing worth pursuing now?
For most companies, full outcome-based pricing isn’t a near-term necessity, but understanding why it’s emerging is. Outcome pricing and partnership structures work best when the value at stake is large, measurable, and recurring, like a healthcare system’s billing resolution rate or a manufacturer’s equipment uptime. For smaller or more exploratory AI use cases, usage-based pricing still offers useful budget predictability.
The bigger shift worth paying attention to isn’t which pricing model wins. It’s the underlying point that buying intelligence doesn’t automatically produce returns. Returns depend on people inside the organization, from individual contributors to executives, learning to act as managers of intelligence: deciding what work to hand to AI, what to keep with a human, and when a repeatable pattern is valuable enough to turn into a new product or service line. No contract structure substitutes for that organizational work.
Frequently Asked Questions
What’s the difference between per-token, per-seat, and outcome-based AI pricing?
Plans first. Then code.
Remy writes the spec, manages the build, and ships the app.
Per-token pricing charges based on the volume of data a model processes. Per-seat pricing charges based on how many users have access to a tool, following traditional SaaS licensing. Outcome-based pricing charges based on a defined business result, like a resolved customer issue, regardless of how much compute or how many users were involved in getting there.
Why are companies struggling to get ROI from AI even when it works well in demos?
A model can perform impressively in isolation while the organization around it hasn’t figured out which tasks to assign it, how to restructure workflows, or how to judge whether its output is actually good. That know-how gap, not the model’s capability, is usually what limits returns.
What is forward deployed engineering?
It’s an engagement model where a vendor’s engineers work inside or closely alongside a customer’s team to apply AI to that customer’s specific data and processes, covering everything from identifying opportunities to building and deploying the resulting systems. OpenAI and Anthropic both use this approach with enterprise and institutional customers.
Are OpenAI and Anthropic actually taking ownership stakes in companies?
Yes. OpenAI took an ownership stake in Thrive Holdings and is embedding teams in its portfolio companies, starting in accounting and IT services. Anthropic formed a joint venture with Blackstone, Hellman & Friedman, and Goldman Sachs focused on applying AI in businesses like mid-size manufacturers and regional healthcare systems.
Does this mean token and seat pricing are going away?
No. Token and seat pricing remain useful for budgeting and for lower-stakes or exploratory use cases. Outcome-based pricing and partnership structures are emerging alongside them for situations where the value created is large and measurable enough to justify a different kind of deal.
