How to Price a Vertical AI Agent Product

How to price an AI agent product comes down to one rule: charge in the unit your buyer already counts.

If they count claims, charge per claim. If they count tickets, charge per ticket. If they count staff, you may be stuck with seats. Then protect your margin, because your costs move even when your price does not.

This post is for domain experts who know a workflow cold and are about to put a price on an agent that does it.

Pick the unit buyers already measure

Start with the buyer’s spreadsheet, not yours. Ask: how do you track the cost of this work today?

A claims shop counts cost per claim.
A recruiting firm counts cost per screened candidate.
A support team counts cost per ticket or cost per contact.
A tax practice counts hours per return.
A freight broker counts cost per load booked.

Whatever they count is your candidate price unit. The buyer can already compare your price to their current cost, and finance does not need to learn a new metric.

Mark Stiving of Impact Pricing makes a similar point: outcome pricing works when you charge on KPIs the buyer already uses to track progress. He also warns that outcome-based pricing is far less common in practice than people think, because measurement, attribution, and seller control over the result are hard.

Three tests for a good unit:

The buyer can verify it. They can see the count in their own system, not just in your invoice.
It tracks value. More units means more value to them.
You control enough of it. If a human on their side decides the result, you cannot bill on it.

Seats fail the second test for agents. Decagon makes this argument: AI agents are not tools for humans, they do whole workflows themselves, so counting seats misses their impact.

Per seat, per task, and per outcome compared

Here are the three main units, plus the hybrid most people end up with.

Model	What the buyer pays for	Strength	Weakness	Fits when
Per seat	Access for each human user	Predictable for both sides	Disconnected from work done; the agent shrinks the seat count	Agent assists named users
Per task or conversation	Each unit of work attempted	Simple, scales with usage, no fights over success	Buyer pays for failures	Quality is high and consistent
Per outcome	Each verified result	Strong alignment, easy to sell internally	Definition disputes, attribution, vendor bears failure cost	Outcome is binary and visible
Hybrid	Base fee plus per-task or per-outcome	Predictable floor with upside	More to explain	Most early-stage products

A few notes on each.

Per seat. It is the model buyers know. The problem is built in. Sierra points out that buyers ask a fair question about agent products: “How much will my seat-based license bill shrink?” If your agent works, fewer humans need seats. Your revenue falls as your product succeeds. Use seats only when the agent is a tool for named people, such as a copilot.

Per task or conversation. Decagon says most of its customers gravitate to per-conversation pricing. Its reasoning: costs scale with usage, invoices are not a surprise, and there are no renegotiations over what a “successful outcome” means. Flat per-task pricing has its own flaw. Salesforce described its own $2 per conversation price as the same “regardless of the complexity of the task”. In May 2025 it added Flex Credits priced per action, 20 credits or $0.10 each. A password reset and a hard dispute should not cost the buyer the same, and they do not cost you the same.

Per outcome. This is the model the market likes to talk about. Sierra charges only when a task is completed, such as a resolved conversation, a saved cancellation, or an upsell. Intercom’s pricing page lists its Fin agent at $0.99 per outcome, charged at most once per conversation, with no charge if the customer asks for a human. (That page shows no date, so check the current terms.) For deeper coverage, read outcome-based pricing for AI agents. Outcome pricing fits when the result is binary, visible to both sides, and mostly caused by your agent.

Hybrid. Bessemer’s playbook says hybrid models, a base subscription plus usage or outcome tiers, win when you are uncertain. They give customers predictability and let you capture upside as they scale. Bessemer calls it the effective middle ground for early-stage startups. That matches what I’d do for a first product.

If you sit on the buyer’s side of the table, the AI agent pricing models buyer guide shows what they are comparing you against.

Margin and model cost

Agent products are not classic SaaS. Bessemer’s playbook says AI companies see 50 to 60 percent gross margins versus 80 to 90 percent for SaaS, and that pricing must account for inference costs. It also quotes the fal.ai co-founder: running the same model might get cheaper, but everyone wants the best model, and those cost more to run.

So three things move your margin that your price does not control:

Model price. It can change when you switch models or a provider changes rates.
Completion rate. Each failed attempt still cost you tokens and tool calls.
Task difficulty. Hard cases use more steps and more context.

Ways to protect yourself:

Price per outcome on cost per completed outcome, not cost per attempt. Failed attempts are your cost.
Add a usage floor or minimum. It covers fixed costs when usage is low.
Tier by complexity. A simple task and a complex one get different prices, as Salesforce moved toward.
Set a fair-use cap on the flat part. A monthly base fee includes a set number of tasks. Overage is charged per unit.
Route to cheaper models for easy work. Save the best model for hard cases.
Write model-cost pass-through language into larger contracts. If your costs rise by more than a set percent, you may reprice with notice.
Track cost per outcome weekly. You cannot manage margin you do not watch.

Worked example with made-up numbers

Say a former insurance claims adjuster builds an agent for first-notice-of-loss intake. Every number below is made up to show the math.

The buyer’s current cost per intake, handled by staff: $6.00. This is the anchor.
Your cost per attempt (model, tools, hosting share): $0.30.
Completion rate (agent finishes without a human): 70%.

You choose a price of $1.50 per completed intake. That is 25 percent of the buyer’s current cost. Their savings per completed intake are $4.50.

Per 100 attempts:

Revenue: 70 completed × $1.50 = $105
Cost: 100 × $0.30 = $30
Gross margin: ($105 − $30) ÷ $105 = 71%

Now stress it.

Scenario	Revenue per 100 attempts	Cost per 100 attempts	Gross margin
Base case	$105	$30	71%
Completion falls to 50%	$75	$30	60%
Model cost rises 50% ($0.45)	$105	$45	57%
Both happen	$75	$45	40%

The last row is the one to fear. If you want to keep a 70 percent margin when cost is $0.45 per attempt and completion is 70 percent, your floor price is about $2.14 per completed intake. That is the number to use when you write a minimum or a repricing clause.

A cleaner structure for this product might be a platform fee that covers setup and a minimum volume, plus $1.50 per completed intake above that. The fee protects the downside. The per-outcome part keeps the buyer’s incentives aligned.

Pilot pricing

Do not give pilots away free. A free pilot attracts tire-kickers and tells the buyer your work is not worth much.

My recommendation: charge a fixed fee for a short, scoped pilot.

Scope it tight. One workflow, one team, a fixed number of weeks.
Cover your costs plus a margin. At minimum, setup time and model spend.
Write the success criteria into the order form. The metric, the target, and how it is measured.
Credit part of the fee toward the first contract. This turns the pilot into a down payment and makes the buyer want to convert.
Define the decision date. The buyer decides by a set day: convert, extend, or stop.

If the buyer will not pay anything, ask why. If budget is the issue, a smaller paid pilot works. If the issue is no internal owner, a free pilot will not fix that either. For why pilots stall, the TMA piece on why AI pilots fail is worth a read.

Contracts and minimums

Put these in the paper, not just the pricing page.

Unit definition. What exactly counts as a billable unit, what does not, and the measurement window. Write it twice: the general definition and the customer-specific one.
Source of truth. Which system’s data decides the count.
Dispute process. How the buyer challenges a charge, and how fast you reply.
Minimum commitment. A monthly or annual floor. Intercom, for example, notes minimum commitments apply when its agent is used with another helpdesk.
Overage rate. What happens above the included volume.
Term and renewal. Annual terms help your cash flow. Month-to-month helps their trust. Pick one on purpose.
Price change terms. When and how you can reprice, and with what notice.
Fair-use and abuse limits.
Data and security terms. Buyers in regulated fields will ask.

If you are working with a technical partner, the split of this revenue matters too. See revenue share in AI development partnerships. And your edge as a domain expert is often distribution, which I cover in the domain expert distribution advantage.

What buyers push back on

Expect these. Have an answer ready for each.

“What counts as resolved?” Decagon puts it bluntly: you never want to be in a situation where you are arguing over what a “resolution” is. Its example: an upset user leaves and gets counted as a resolution. Answer with a written definition and an audit trail they can inspect.
“Do I pay when it hands off to a person?” Sierra lists paying for escalations among the things no one wants. Say no, or price the handoff separately and say so up front.
“What will my invoice be next month?” Offer a cap, a committed volume, or a forecast with a range.
“Why should I pay more for the hard ones?” Because they cost more and are worth more. Show the tiers.
“Your seat count should drop. Why charge by seat?” Fair. Drop seats unless the agent is a copilot.
“Prove the savings first.” This is what the paid pilot is for.
“What if the model gets cheaper?” Then you may pass on part of the saving at renewal. Say so up front, in return for a pass-through clause when costs rise.
“Can I switch back?” Offer data export and clear exit terms.

Bottom Line

Price in the unit your buyer already counts. Start hybrid: a base fee plus a per-task or per-outcome rate.

Compute your cost per completed outcome, then stress it against a worse completion rate and a higher model cost.

Charge for the pilot, write the unit definition into the contract, and keep a repricing clause.

FAQ

How should I price an AI agent product? Start with the unit your buyer already measures, such as cost per claim or per ticket. Price below their current cost per unit, and cover your cost per completed outcome with a healthy margin. For a first product, a base fee plus a per-task or per-outcome rate is the safest start.

Per seat, task, or outcome? Use seats only if the agent is a tool for named users. Use per-task pricing when quality is steady and you want simple invoices. Use per-outcome pricing when the result is binary, visible to both sides, and mostly caused by your agent.

How do I protect margin? Price on cost per completed outcome, not per attempt. Add minimums, tier by complexity, route easy work to cheaper models, and put a repricing clause in larger contracts. Track cost per outcome every week and stress-test it against lower completion rates and higher model costs.

How do I price a pilot? Charge a fixed fee for a tight scope and a fixed number of weeks. Put the success metric and decision date in writing. Credit part of the fee toward the first contract so the pilot works as a down payment.

What do buyers push back on? The definition of a billable outcome, paying for escalations, unpredictable invoices, and paying the same for easy and hard tasks. Answer with written definitions, audit trails, caps, and complexity tiers.

Three Ways to Work With TMA

Need an agent built? We deploy production AI agents in your infrastructure. Working pilot. Real data. Measurable ROI. → Schedule Demo

Want to co-build a product? We’re not a dev agency. We’re co-builders. Shared cost. Shared upside. → Partner with Us

Want to join the Guild? Ship pilots, earn bounties, share profit. Community + equity + path to exit. → Become an AI Architect