For the better part of two decades, software pricing had a dependable default: charge per seat. A company bought a license for every employee who needed the product, the vendor’s revenue grew as the customer hired, and finance teams on both sides could forecast the bill with a spreadsheet and a headcount plan. Seat pricing was simple to explain, simple to buy, and closely tied to value, because the more people who relied on the software, the more it was worth to the business.
AI agents break that logic. An agent that resolves support tickets, reconciles invoices, qualifies inbound leads, or reviews contracts doesn’t log in, doesn’t need a license, and doesn’t occupy a chair. In many deployments it is doing the very work that used to justify the seats in the first place. That creates a problem most AI companies discover only after launch: the better the product performs, the fewer seats the customer needs, and the less a seat-based vendor earns.
This article walks through the three main ways to price an AI agent (seats, usage, and outcomes), explains where each one works and where it quietly breaks, and gives you a practical framework for choosing between them. Along the way you’ll find reference graphics you can save, share with your team, or bring into a pricing review.
The Seat Paradox: Why Traditional Pricing Works Against AI
To see the problem clearly, it helps to run the numbers on a single customer. Imagine a support organization handling 10,000 tickets a month with 20 human agents, each managing about 500 tickets. A vendor charging $150 per seat earns $3,000 a month from that team. Now the vendor adds an AI agent that resolves tickets on its own. As the AI handles a larger share of the volume, the customer needs fewer human agents, and the seat count starts to fall.
At a 50% resolution rate, the customer needs only 10 agents, and seat revenue drops to $1,500. At 80%, the team shrinks to four people and the vendor earns $600. The vendor has delivered enormous value, arguably far more than when it was a simple ticketing tool, yet its revenue from the account has collapsed by 80%. Every product improvement accelerates the decline.
Compare that with charging $1 for each ticket the AI resolves on its own. At a 50% resolution rate, the vendor earns $5,000. At 80%, it earns $8,000. The two revenue lines cross at around a 23% resolution rate, which is a level many AI support tools reach early in a deployment. Past that point, seat pricing isn’t just leaving money on the table. It is actively penalizing the vendor for building a better product.
This misalignment shows up in the customer relationship too. When a vendor’s revenue depends on headcount, its incentives quietly diverge from the customer’s goal of doing more with fewer people. Buyers notice. Procurement teams increasingly ask why they should pay per seat for a product whose main selling point is that it reduces the need for seats.
Three Ways to Price an AI Agent
There is no single correct model, and each of the three approaches below is in active use across the AI market. What distinguishes them is what the customer is actually paying for: access, activity, or results. That choice shapes your revenue growth, your margin risk, and the story your sales team tells.
Option 1: Seat-Based Pricing
Seat-based pricing treats the AI as a feature of a license that a person holds, such as “$50 per user per month.” Its strengths are real. Buyers understand it instantly, budgets are easy to forecast, and procurement already has a process for approving it. For products that bolt AI onto an existing workflow tool, keeping seats avoids forcing customers to learn a new way of buying.
The weaknesses emerge as the AI takes on more work. Revenue stays tied to headcount, which the AI is often designed to reduce. Seat pricing also ignores how heavily each user leans on the AI. A power user who runs hundreds of expensive model calls a day pays the same as a colleague who tries the feature once a week, which makes gross margin unpredictable when inference costs are significant. Finally, a seat price says nothing about the value the AI delivers, so it gives sales teams little to anchor a value conversation on.
Seat pricing works best when the AI is a copilot: it helps a person write, search, or decide faster, but the person still does the job. Coding assistants and writing tools are good examples. Once the AI is completing tasks end to end, seats begin to feel arbitrary to the buyer.
Option 2: Usage-Based Pricing
Usage-based pricing charges for what the AI does, measured in conversations, tasks, API calls, documents processed, or credits. A price of “$0.50 per conversation handled” is a typical example. This model scales naturally with adoption, since customers pay more as they rely on the product more, and it keeps revenue closely aligned with the vendor’s own compute costs, which protects margins. It also lowers the barrier to entry, because customers can start small and expand without renegotiating a contract.
The central weakness is that usage measures effort rather than results. A conversation that fails to solve the customer’s problem costs the same as one that succeeds. If a customer pays for 1,000 conversations and only 600 of them actually resolve the issue, the question “why am I paying for the other 400?” is inevitable, and it tends to arrive during renewal. Usage bills can also be hard to predict, which makes budget owners nervous, and abstract units such as credits or tokens force customers to translate your pricing into business value on their own. (We cover how to solve that translation problem in How to Communicate AI Credit Models to Customers.)
Usage pricing fits best when workloads vary widely between customers, when the product is an API or platform that others build on, or when success is hard to define but activity is easy to count. Our guide to transitioning to consumption pricing for AI covers the operational side of making this shift.
Option 3: Outcome-Based Pricing
Outcome-based pricing charges only when the AI delivers a defined result, such as “$1 per support ticket resolved without human involvement.” The model drew wide attention when Intercom priced its Fin AI agent per resolution rather than per seat, and it has since spread to other agent categories where the result is easy to count.
The appeal is obvious. Price is tied directly to value, so the sales conversation shifts from “what does this cost?” to “what is this worth?” The vendor carries the performance risk, which builds trust with skeptical buyers who have seen AI demos that didn’t hold up in production. And revenue grows as the product improves, which aligns the vendor’s roadmap with the customer’s goals.
The difficulty lies in the details. Both sides must agree precisely on what counts as an outcome, and that is harder than it sounds. Does a ticket count as resolved if the customer reopens it two days later? What if a human agent edited the AI’s response before it was sent? Attribution disputes can erode trust quickly if these rules aren’t settled upfront. Revenue can also be volatile, since it rises and falls with the customer’s own volume and with the AI’s success rate in any given month.
Outcome pricing is the strongest value story an AI company can tell, but only when the outcome is measurable, attributable, and agreed upon before the first invoice goes out.
Match the Model to How Much Work the AI Does
The single most useful question in AI pricing is how much of the work the AI actually does. AI products sit on a spectrum. At one end, the AI suggests: autocomplete, writing hints, and recommendations that a person accepts or ignores. Further along, it drafts: first versions of emails, code, or reports that a person reviews and edits. Further still, it acts with approval, preparing a refund or updating a record while a person signs off on each action. At the far end, it completes the work end to end with no human in the loop.
Pricing should move along that same spectrum. When a person does the work and the AI assists, the person remains the natural unit of value, and seats make sense. As the AI takes over more of the task, activity becomes a better proxy for value, and usage pricing fits. When the AI owns the task completely, the result itself is the product, and outcome pricing becomes both possible and persuasive.
Many companies have products at several points on this spectrum at once. A customer service platform might offer AI reply suggestions to human agents (a copilot feature) alongside a fully autonomous AI agent for simple tickets. In that case, it is reasonable to price each capability according to where it sits, rather than forcing one model across the whole product.
A Decision Framework: Which Model Fits Your AI Product?
Three questions will narrow the choice for most AI products.
First, does the AI assist a person, or complete a task on its own? If it assists, seat pricing remains viable, and you can add usage limits or overage fees if compute costs vary a lot between users. If it completes tasks, move on to the next question.
Second, can you and the customer agree on what success looks like and count it? “Ticket resolved,” “invoice matched,” “meeting booked,” and “claim processed” are countable events. “Better insights,” “faster decisions,” and “improved productivity” are not. If success is fuzzy, outcome pricing will generate more arguments than revenue, and usage pricing is the safer choice. Just make sure you translate usage units into outcomes on your pricing page so customers can connect cost to value.
Third, does the customer need a predictable monthly bill? If not, pure outcome pricing gives you the cleanest value alignment. If the customer’s finance team needs to budget with confidence, and most enterprise buyers do, a hybrid model will close deals faster.
A fourth consideration sits underneath all three questions: your cost structure. If one task might cost you $0.02 in inference and another $2.00, depending on document length or reasoning depth, any flat price exposes your margins to the customers with the heaviest workloads. Usage and outcome models let revenue move with cost, and even seat-based products benefit from fair-use limits in that situation.
The Hybrid Model Most Companies End Up Using
In practice, the most durable AI pricing structures combine a fixed component with a variable one. A typical version pairs a platform fee covering access, integrations, administration, and reporting with a per-outcome or per-usage fee that scales with the value delivered.
Consider an AI support agent priced at a $500 monthly platform fee plus $0.90 per resolved ticket. In a month where the AI resolves 1,800 tickets and escalates 700 to human agents, the customer’s invoice is $2,120: the $500 platform fee plus $1,620 in resolution fees. The 700 escalated tickets cost nothing, which reinforces the message that the customer only pays for results.
The structure works for both sides. The platform fee gives the vendor a revenue floor that covers fixed costs such as onboarding and support, and it gives the customer a baseline they can budget around. The outcome fee ties growth to delivered value, so the vendor earns more only when the AI performs. This is the same logic behind the hybrid subscription and usage models that have become common across SaaS, adapted to the way AI agents create value.
Hybrid models also give you more room to package. You can tier the platform fee by feature set, such as advanced analytics or additional integrations, while keeping the outcome fee constant. Or you can offer volume discounts on the outcome fee in exchange for a committed monthly minimum, which improves revenue predictability without abandoning value alignment.
Defining an Outcome You Can Actually Bill For
If you move toward outcome pricing, whether pure or hybrid, the definition of the outcome becomes the most important sentence in your contract. Vague definitions lead to billing disputes, and billing disputes lead to churn. Before you charge per outcome, check that your definition passes six tests.
It must be countable, meaning it is a discrete event your system logs, not a subjective improvement. It must be attributable, so you can show the AI caused it and you have a rule for tasks a human also touched. It must be agreed in writing before billing starts, including edge cases. It must be durable: set a window, for example treating a ticket reopened within 72 hours as unresolved and unbilled. It must be auditable, so customers can see every billed outcome in a report and dispute it through a clear process. And it should be capped, with a monthly ceiling or committed volume so an unusually good month doesn’t become a budget shock that sours the relationship.
Getting these rules right protects both sides. Customers gain confidence that they are paying only for real results, and vendors avoid the slow erosion of trust that comes from contested invoices.
Communicate the Price Against Human Cost
Whichever model you choose, buyers will judge your price against one benchmark whether you mention it or not: what it would cost to have a person do the same work. The most effective AI pricing pages make that comparison explicit instead of leaving it to the customer.
A price of “$1 per resolution” sounds like a line item. “$1 per resolution, compared with $6 to $12 when a human agent handles the ticket” sounds like savings. The price hasn’t changed. The reference point has. In the hybrid example above, the customer’s $2,120 invoice covers work that would cost roughly $14,400 if 1,800 tickets were handled by people at $8 each. That is a reduction of about 85%, and framing it that way turns a software expense into a cost-savings story that finance teams can take to leadership.
When customers compare your price to labor rather than to other software subscriptions, a four-figure monthly bill stops looking expensive and starts looking like the obvious choice. For a deeper look at how AI changes the way customers calculate value, see How Economic Value to the Customer Evolves in the Age of AI.
Final Thought
AI agents break the assumptions that software pricing was built on. Seats measure people, usage measures activity, and outcomes measure value. None of the three is universally right, but the direction of travel is clear: the further your AI moves from assisting work to completing it, the further your pricing should move toward outcomes.
The companies that price AI well won’t be the ones asking “how many users do you have?” They’ll be asking “how much work do you need done?” and building a price around the answer.
