Outcome-based pricing for AI agents: examples, pros, cons and when it backfires
Published November 17, 2026 · By Sergio Palmero · 12 min read
- Measurement & ROI
- Customer Support
- CX Operations

Quick summary
Of the four AI agent vendors compared here, Fin, Zendesk and Sierra bill per outcome, usually a resolved conversation; Salesforce Agentforce bills $2 per conversation (vendor pages, read 11 November 2026).
Fin, Intercom's AI agent, bills $0.99 per outcome and counts a resolution when "no further help is requested after Fin's last answer" (fin.ai, read 11 November 2026).
On the same conversations across all Zentesimal customers, a "not escalated" rule would count 86% as billable and a "customer confirmed" rule 15 to 20% (October 2026).
The bill grows as the AI improves: at Fin's list price, every extra 1,000 resolved conversations adds $990 a month, so a fixed budget needs a monthly cap (fin.ai, November 2026).
Outcome-based pricing charges for an AI agent only when it achieves an agreed result, usually a resolved conversation. It ties the bill to results and makes failed chats free.
It backfires when the vendor alone decides what counts as an outcome, when the buyer can't check the count, and when success grows the bill faster than the budget.
This guide covers what the model is, which vendors use it, how it compares with usage-based and seat-based pricing, and the contract terms that make it a good deal. For list prices across every model, see our guide to AI support agent pricing models.
What is outcome-based pricing?
Outcome-based pricing means the buyer pays each time the product achieves an agreed result, not per user or per use.
The idea is older than AI. Public services have used it for years. The Government Outcomes Lab, a research centre at the University of Oxford, defines it simply: "In an outcomes-based contract, payment is wholly or partly dependent on outcomes being achieved."
AI agents (software that answers customers and can act in other systems) made the model practical for software. Every conversation leaves a written record, and that record can be checked against a billing rule.
Definitions used in this guide
Term | What counts | Where it is decided |
|---|---|---|
Outcome | The result the vendor bills for: a resolved conversation, a saved cancellation, an upgrade | The contract |
Handled conversation | Any conversation the AI agent took part in, resolved or not | Usage logs |
Resolved conversation | The customer's problem was solved, shown by the exchange or confirmed by the customer | The vendor's billing rule |
Uncertain conversation | The record doesn't show whether the problem was solved | The vendor: billed or left out |
Escalated conversation | Handed to a human before it was solved | Usually not billed |
What are examples of outcome-based pricing for AI agents?
Fin, Zendesk AI agents and Sierra bill per outcome. Salesforce Agentforce bills per conversation or per action, which is usage-based pricing.
Outcome-based pricing examples for AI agents, November 2026
Vendor | Unit billed | Published price | What counts as the outcome | What isn't billed |
|---|---|---|---|---|
Fin (Intercom's AI agent) | Outcome | $0.99 per outcome; 50 a month minimum with another helpdesk | "No further help is requested after Fin's last answer", or a configured handoff procedure completed | Default escalations, failed procedures, more than one outcome per conversation |
Zendesk AI agents | Verified automated resolution | $1.50 per verified resolution (November 2026) | Resolved without a human, then verified by a large language model (LLM) that reads the conversation | Assisted escalations, and "contained" resolutions that fail the check |
Sierra (an AI agent vendor for large companies) | Outcome agreed per contract | Not published | A resolved support conversation, a saved cancellation, an upsell or a cross-sell | Unresolved or escalated conversations, "in most cases" |
Salesforce Agentforce, Salesforce's AI agent platform (for contrast) | Conversation or action | $2 per conversation, or $0.10 per action, paid from prepaid Flex Credits (May 2025) | No outcome: every conversation or action is billed | Nothing; usage-based |
Sources, read on 11 November 2026: Fin pricing, Fin outcomes, Zendesk automated resolution tiers, Sierra on outcome-based pricing, Agentforce pricing. Zendesk's price per verified resolution was checked on 12 November 2026.
How did Zendesk change its definition in 2026?
Since 18 May 2026, Zendesk sorts AI conversations into three tiers and bills only "verified" resolutions. An LLM reads each conversation first; the ones that fail the check become "contained" and don't use the resolution allowance (the block of resolutions included with the plan).
Email conversations are judged 72 hours after the first email. Messaging conversations are judged after 2 hours by default, adjustable up to 72.
Which outcomes go beyond support?
Sierra's list is where the model is heading: saved cancellations and upsells are revenue outcomes. They are harder to agree on than a resolved ticket, because some of those customers would have stayed or upgraded anyway.
How does outcome-based pricing compare with usage-based and seat-based pricing?
Seat-based pricing follows headcount, usage-based follows volume and outcome-based follows results. Each one puts the risk of a poor AI agent on a different side.
Seat-based, usage-based and outcome-based pricing compared
Seat-based | Usage-based (also called consumption-based) | Outcome-based | |
|---|---|---|---|
The bill grows with | Team size | Conversations, messages or actions | Results the vendor counts |
Who pays when the AI performs badly | The buyer | The buyer | The vendor |
Budget predictability | High | Medium: follows volume | Lower: follows volume and resolution rate |
The vendor earns more from | More users | More activity | More outcomes |
Example | Helpdesk seats for human agents | Agentforce per conversation; Braze, a customer engagement platform, priced on monthly active users (MAU) and Action Credits (prepaid credits per message channel) | Fin, Zendesk AI agents, Sierra |
Usage pricing draws its own complaint. Prospects running Braze have told us the bill grows with message volume whether a campaign converts or not.
What happens to the bill when the AI gets better?
Under outcome-based pricing, a better AI agent means a bigger bill. Under usage-based pricing, the bill stays flat whatever the AI resolves.
The example takes 10,000 AI conversations a month at Fin's list price per outcome and Agentforce's list price per conversation. It leaves out platform fees, seats and minimums.
Monthly bill for 10,000 conversations by resolution rate (illustrative)
Resolution rate | Outcome-based, $0.99 per outcome | Usage-based, $2 per conversation |
|---|---|---|
30% | $2,970 | $20,000 |
50% | $4,950 | $20,000 |
70% | $6,930 | $20,000 |
Outcome-based is cheaper at every rate in this example. It is also the only model where the bill moves with performance.
On the same volume, the outcome-based bill grows with every point of resolution rate; the usage-based bill doesn't move.
A finance team that budgeted a fixed amount reads that rise as an overrun. The support team reads it as success. The contract should settle which reading wins before the AI improves.
Decision matrix: which pricing model fits your situation
Your situation | Best fit | Why |
|---|---|---|
New AI agent, resolution rate unknown or below 30% | Outcome-based | Failed conversations cost nothing while the AI learns |
Stable, high resolution rate and a fixed annual budget | Outcome-based with a monthly cap, or usage-based with a commitment | The bill becomes predictable |
Many conversations are routing or greetings with no clear result | Usage-based | There is no outcome to count; Sierra says these "may align better with consumption-based pricing" |
No way to see which conversations were billed | Neither variable model until there is an audit trail | The vendor's count can't be checked |
AI copilot drafting replies for human agents | Seat-based | The value follows each person using it |
Revenue outcomes such as saved cancellations or upgrades | Outcome-based with an attribution window (the days after contact in which a result counts) and a control group (users who get no AI contact) | Otherwise the vendor is paid for results that would have happened anyway |
What are the advantages of outcome-based pricing?
The buyer pays only for results, failed conversations cost nothing, and the vendor's revenue grows only when its AI resolves more.
- Lower risk at the start. Resolution rates are lowest in the first weeks, while the AI agent's instructions and sources are still being tuned, and usage pricing bills every failed attempt.
- Incentives that point the same way. A vendor acting for a buyer is a classic principal-agent problem: their interests drift apart. Paying on results narrows the gap, because the vendor earns more only by resolving more.
- A simpler business case. Cost per resolved conversation compares directly with the cost of a ticket handled by a person. Our guide on measuring AI ROI in customer support covers that calculation.
When does outcome-based pricing backfire?
It backfires when the vendor alone defines the outcome, when the buyer can't audit the count, and when more resolutions push the bill past a fixed budget.
What happens when the vendor defines the outcome?
A loose definition can bill conversations where the customer simply stopped replying. The same conversations can then cost very different amounts.
Several Zentesimal customers have raised the same concern about Fin: which conversations counted as an outcome.
Fin's pricing counts a resolution when "no further help is requested after Fin's last answer". Intercom's help centre calls one case an assumed resolution: the customer "exits the conversation without requesting further assistance".
A customer who leaves without saying anything may have been helped. They may also have given up.

The same conversations under three billing rules
Billing rule | What it counts | Share of conversations |
|---|---|---|
Not escalated | Everything that didn't reach a human, uncertain included | 86% |
Resolved with confidence | Solved, shown by the exchange or confirmed by the customer | 72% |
Customer confirmed | Only conversations where the customer says it worked | 15 to 20% |
Method: Zentesimal's own classification of every conversation across all customers (more than 40,000 a month, October 2026), as escalated (about 14%), uncertain (about 14%) or resolved with confidence.
The gap between the first two rows is the uncertain conversations: a "not escalated" rule bills them, a "resolved with confidence" rule leaves them out.
Billing also changes what the number means. Goodhart's law puts it as "When a measure becomes a target, it ceases to be a good measure." Once a vendor is paid per resolution, its resolution rate stops being a neutral quality metric.
What happens when the buyer can't audit the count?
Without a list of billed conversations and the reason for each, a dispute is one claim against another.
Zendesk's LLM check is a step toward a verifiable count. What a buyer needs on top is the list itself: each billed conversation, its state, why it counted, and a way to contest it.
Public outcomes contracts know this risk well. The Oxford Government Outcomes Lab warns about providers picking the easiest cases to hit their targets, a behaviour it calls gaming.
In an AI agent, the equivalent is closing a conversation fast instead of asking the question that would solve it.
What happens to the budget when resolutions rise?
The bill rises with every extra resolution, so a fixed budget runs out exactly when the AI agent works best.
The resolution-rate table in the comparison section shows the size of the effect at a constant volume. Volume growth adds to it. A monthly cap and an alert threshold keep the bill inside the budget without switching the AI off.
What happens with revenue outcomes?
Paying per saved cancellation or upgrade also pays for the ones that would have happened without the AI.
The fix is a control group: a share of users who get no AI contact, so the difference between the two groups is what the AI caused.
Without a control group, the vendor's count includes customers who would have upgraded anyway.
What should an outcome-based contract define?
Nine terms decide whether outcome-based pricing is a good deal, and every vendor can answer them in writing before signing.
Checklist: outcome-based AI agent contract
- The outcome in one sentence. Which result is billed and how often, for example once per conversation.
- The evidence. Customer confirmation, an action completed in another system (refund issued, plan changed) or an LLM review.
- Uncertain conversations. Billed or left out, and who decides.
- The wait window. How long after the last message the outcome is judged, and whether a conversation reopened within 7 days is billed again.
- Handoffs. Whether a handoff to a human is ever billed. Fin bills a configured handoff procedure.
- The audit trail. A monthly export of every billed conversation with its state and the reason it counted.
- Disputes. How to contest a billed outcome, within how many days, and who reviews it.
- Caps and alerts. A monthly spending limit and an alert at, for example, 80% of it.
- Revenue outcomes. For saved cancellations or upgrades: an attribution window, such as 30 days, and a control group with no AI contact.
How Zentesimal charges for its AI agents
Zentesimal charges a fixed fee per plan, plus a fee per outcome: per real resolution for support agents and per goal achieved for Plays.
- Plan fee. A fixed standard fee per plan, listed on the pricing page.
- Support agents (AI agents that answer users who wrote in) are billed per resolution. A resolution counts when the result is clear from the exchange or the customer confirms it.
- Plays (AI agents that reach out first, for onboarding, upgrades or retention) are billed per goal achieved. The team sets the goal when it sets up the Play, such as an upgrade or a completed onboarding step, and can change it as goals are reached.
- Not billed: uncertain and escalated conversations. Uncertain ones stay visible for review.
Where the team checks what was billed
The Performance section shows each AI agent's successes against its total for the period.
For each AI agent and period, it also shows the success rate, hours saved, money saved and revenue generated. A success is the outcome Zentesimal bills: a resolution for a support agent, a goal achieved for a Play.

Agent traces show what the AI agent did in each conversation.
For revenue goals, Zentesimal runs experiments with versions of an AI agent against a control group that gets no contact.
Key takeaways
Takeaway | Next step |
|---|---|
Outcome-based pricing bills results, so the definition of a result decides the bill | A one-sentence outcome definition in the contract |
The billing rule changes the billable share of conversations more than the unit price does | The vendor's rule applied to a sample month of your own conversations |
The bill grows as the AI agent improves | A monthly cap and a budget tied to the expected resolution rate |
Revenue outcomes include results that would have happened anyway | A control group and an attribution window |
Conversations with no clear result fit usage pricing better | Usage pricing for routing and greetings |
FAQ
Is outcome-based pricing the same as value-based pricing?
No. Value-based pricing sets the price level from the value the buyer perceives. Outcome-based pricing sets when the buyer pays: only after an agreed result. A vendor can use both at once.
Is outcome-based pricing cheaper than per-seat pricing?
They pay for different things and usually sit on the same invoice. Seats pay for the people who use the helpdesk or a copilot. Outcomes pay for what the AI agent resolves on its own.
What is a hybrid pricing model for AI agents?
A hybrid model combines a fixed fee with a variable fee per use or per outcome. The fixed part makes the budget predictable; the variable part follows results. A fixed platform fee plus a fee per resolution is a common example.
Who decides whether a conversation counts as resolved?
The vendor's billing rule, written in the contract. Some vendors use an LLM review, some a rule on customer behaviour such as no further request, and some require evidence in the conversation or customer confirmation.
Is agentic AI pricing moving away from seats?
For agentic AI (AI agents that act on their own, not copilots), most published prices are per outcome, conversation or action. Seats remain for the human team and for copilots that draft replies for a person.
Related articles
- AI support agent pricing: per resolution, credits or seats?
AI support agents are priced per resolution, per conversation or action, or per seat. 2026 list prices, what each model rewards, and what to ask vendors.
- Beyond "Not Escalated": How Not to Measure AI Support Success
Containment rate looks like success but hides silent churn. Here's the four state model that catches it before customers walk away.
- How to Measure AI ROI in Customer Support
Cost per ticket is only one lever. Here’s how CX teams measure AI agent ROI across OPEX, revenue, and the hours the team gets back.
- Best agentic AI tools for customer service in 2027: from answering tickets to selling and retaining
Agentic AI tools for customer service, grouped by job: what each can act on, how its AI is billed, plus a decision matrix and a checklist before it acts.
- How to run lifecycle experiments across the customer journey with AI agents
How to run lifecycle experiments across the customer journey: hypotheses from your baseline, tests on demo accounts and AI agents running each Play.