Outcome-Based AI Pricing: Only 4 of 12 Published Meters Bill on a Defined Result

Intercom's $0.99 per resolution made "pay only for results" the default pitch for AI support agents. Read the billing documentation behind seven vendors' price cards and the pitch thins out fast. Of 12 published meters, only 4 bill on a result defined in public, under rules that leave a failed conversation free. Outcome-based AI pricing is real, and it is much rarer than the marketing suggests.

The short answer. Outcome-based AI pricing means you pay only when the agent achieves a defined result, such as a resolved ticket, and nothing when it fails or escalates. Intercom Fin, Zendesk, HubSpot and Gorgias publish meters that meet that test. Salesforce and Freshworks mostly bill activity, and Sierra negotiates its outcome definitions privately.

Outcome-based AI pricing has a test, and most meters fail it

Every vendor in this survey uses the language of results somewhere on its site. That is not the same as billing on results. To separate the two, I applied three questions to every meter, using only what the vendor itself publishes.

First, is the billing trigger a customer result, or is it an action the agent took? A resolved ticket is a result. An answered message, a session or a recommended lead is an activity, however useful. Second, is the trigger defined in public documentation that a buyer can read before signing? Third, does a failed or escalated conversation cost nothing?

A meter has to pass all three to count. The bar is deliberately low. The interesting gap is not between generous and stingy definitions. It is between meters with a definition you can read and meters where the word "outcome" appears only in the sales deck.

The earlier pieces on this site priced a 5,000-ticket month on Fin, Zendesk and HubSpot and listed the contract clauses an outcome deal needs. This post asks which meters on the market are outcome meters at all.

Twelve outcome pricing meters, one scorecard

The sample is 7 vendors and the 12 billing meters they publish for AI agents in customer service and sales. Ada and Decagon were also checked and are left out of the count because neither publishes a rate card or a billing definition I could read. Every row below cites the vendor's own page, read on 2 November 2026.

Prices are US dollar list rates from each vendor's pricing page or help centre. The HubSpot 72-hour rule was read in HubSpot's knowledge base on 25 October 2026; that article returned a 404 on 2 November 2026. Gorgias states that automation prices depend on subscription tier and does not print them in its billing documentation. Agentforce action price is computed as 20 credits at $500 per 100,000.

Verified or inferred results: the four that pass

Zendesk, Fin, HubSpot's Customer Agent and Gorgias pass all three questions. Each bills a customer result, each publishes the rule, and each excludes escalations. That is 4 of 12 meters, and every one of them sits in customer support, where a result is comparatively easy to observe.

Sierra's meter is a result but fails the visibility question. Salesforce's Help Agent Resolution has a results-sounding name and fails the definition question outright, because the pricing page gives a price and nothing else.

Completions and activity: the eight that do not

The three completion meters bill a step on the way to value. A Fin qualification at $9.99 bills a routed prospect, not a closed deal. HubSpot moved its Prospecting Agent from a monthly charge per enrolled contact to $1.00 per lead recommended for outreach. Both are fairer than the old seat or contact charge. Neither is a result your finance team would recognise.

The three activity meters are honest about being activity. They are included because they are what the outcome meters compete against, and because two of them come from Salesforce, which agreed to buy Fin in June 2026, as covered in the analysis of Fin's resolution definition.

Intercom Fin: the clearest exclusions, a silence-based trigger

Fin has the most complete public rulebook in the sample. The Intercom help centre article on outcomes, dated 23 September 2026, counts a resolution when the customer either confirms the answer or "exits the conversation without requesting further assistance".

The exclusion list is long and specific. An explicit request for a human is not billed at any point. Neither is a failed procedure, a default escalation, a workspace-rule escalation, spam, an abandoned conversation or an outbound message nobody answers.

Two further rules protect the buyer more than most. Fin charges at most once per conversation however many questions are asked. And if a customer returns to a resolved conversation, the resolution is deducted, "even across billing periods". No other vendor in this sample publishes a clawback that explicit.

Two pages, one definition, one missing number

The weakness is the trigger. The current Intercom article describes an assumed resolution without a time window. The number lives on a second page, the Fin help centre, which sets it at 24 hours of disengagement after Fin's last answer.

That is a minor documentation gap with a real consequence. A customer who reads a wrong answer and leaves bills exactly like one who was helped. The point here is comparative: Fin's exclusions are the best in the sample, and its trigger is ordinary.

Fin also sells completions alongside the resolution meter. The Fin pricing page lists resolutions, procedure handoffs, disqualifications and self-serve routings at $0.99, and qualifications at $9.99. A procedure handoff can bill a ticket that a human finished. If you design many procedures that end at a person, your outcome count will run ahead of your resolution count.

Zendesk: the only outcome meter that checks the work

Zendesk rebuilt its definition on 18 May 2026, and it is now the strictest in the sample by some distance. Its help centre article on resolution tiers sorts every AI conversation into three groups.

An assisted escalation is one where a human finished the job. A contained resolution is one where the customer did not ask for more help. A verified resolution is one where "a large language model (LLM) that evaluates the text of the conversation" confirms the request "was satisfactorily resolved". Only the verified tier draws down your allowance. The other two are explicitly free.

A contained resolution is exactly what Fin, HubSpot and Gorgias bill: a conversation where the customer went quiet. Zendesk looks at the same silence and declines to charge for it unless a second model agrees the problem was solved. It is the most defensible definition in the sample, and the reason Zendesk's higher rate is not automatically the worse deal.

The pricing page undersells its own rule

The Zendesk pricing page lists $1.50 per committed resolution and $2.00 pay as you go, with 5 included resolutions per agent each month on Team plans and 10 on Suite Professional. Its FAQ describes the meter as paying "only for customer requests that were successfully resolved by the AI agent, without any escalation to a human agent".

That wording describes a contained resolution, which the help centre says is not billed. So the sales page promises less protection than the billing documentation delivers. The help centre is the operative text. Two gaps remain. The model doing the verifying belongs to the vendor, and Zendesk does not publish how often it overturns a contained conversation.

Email conversations close for evaluation 72 hours after the first email. Messaging closes after 2 hours by default, adjustable up to 72. Voice closes on hang-up.

HubSpot and Gorgias: 72 hours of quiet, then a charge

HubSpot and Gorgias both pass the scorecard test, and both use the longest silence windows in the sample. A customer who comes back after two days is caught by a 72-hour window and missed by a 24-hour one.

HubSpot's resolution meter passes, its lead meter does not

HubSpot's April 2026 announcement moved Customer Agent from $1.00 per conversation to $0.50 per resolved conversation from 14 April, billed as 50 credits. It says customers "only pay when the agents complete the task it's been assigned". The announcement itself does not define resolved.

The definition sat in HubSpot's knowledge base, which on 25 October 2026 counted a resolution when there was no qualifying handoff within 72 hours of the last visitor reply. It also said manual assignment by your team is not a qualifying handoff. That page returned a 404 when I checked on 2 November. The rule may have moved rather than changed, but a buyer cannot cite a page that does not load. The mechanics of the credit switch are in the breakdown of HubSpot's credit repricing.

Gorgias can bill one ticket twice

Gorgias has the cleanest definition of the inferred group. Its billing documentation treats an interaction as automated when "the customer does not require the involvement of a human agent within 72 hours". A ticket handed to a human pays only the ticket fee, and a shopper who ignores the agent's first search-assist message costs nothing.

The catch is in the same document. An automation fee and a ticket fee can both apply to the same ticket when the AI resolves it entirely. So a successful AI resolution costs you a helpdesk ticket plus an outcome fee. Accounts created before 28 May 2025 on legacy pricing are counted one way or the other, not both. If you are on newer pricing, model your outcome cost with the ticket fee added back in.

Sierra and Salesforce: outcomes you cannot read in advance

The two most ambitious outcome claims in the market come with the least public detail.

Sierra prices saves and upsells, privately

Sierra's own explanation of its model, written by Elliot Greenwald in December 2024, lists billable outcomes as "a resolved support conversation, a saved cancellation, an upsell, a cross-sell". It says unresolved and escalated conversations carry no charge "in most cases", and that it provides "clear, agreed-upon criteria for each outcome upfront".

Those criteria are agreed in each contract, and no rate is published. That means Sierra fails the visibility question, and that the hardest definitions in the sample are invisible. A resolved ticket is observable. A saved cancellation needs a rule about how long the customer must stay, and an upsell needs an attribution rule against every other channel that touched the buyer.

If you negotiate a saves meter, write the retention window and the attribution rule into the order form, and treat any answer of "standard terms" as a request to see them.

Agentforce sells a resolution it does not define

The Agentforce pricing page offers three units. Flex Credits cost $500 per 100,000, and a standard action uses 20 credits, which is $0.10 an action. Conversations cost $2 each. Each action is "metered individually", whether or not it helped.

The third line is the odd one. The comparison table lists "Help Agent Resolutions" at $2, with no definition anywhere on the page. A buyer cannot tell whether that resolution is verified, inferred from silence or simply a conversation that ended. The history of the conversation meter is covered in the conversation pricing analysis.

Salesforce now owns, or is about to own, the market's best-documented resolution meter in Fin. Whether Fin's exclusions migrate into Agentforce, or Agentforce's undefined resolution absorbs Fin, will set the default for a large share of enterprise buyers.

When a flat session costs less than a per-resolution price

A per-resolution price feels safer because failures are free. It is not always cheaper. The comparison that matters is cost per conversation handled, which is the resolution price multiplied by your resolution rate.

Freshworks makes a useful benchmark because its meter is pure activity. A Freddy AI Agent session costs $49 per 100, or $0.49, and covers every interaction with one user in 24 hours. It bills whether or not the issue was solved.

The arithmetic is short. Zendesk at $1.50 equals a $0.49 session at a resolution rate of 32.7%. Fin at $0.99 equals it at 49.5%. HubSpot at $0.50 stays below the session price until resolution reaches 98%. Against Agentforce's $2 conversation, every per-resolution meter in the sample is cheaper at every resolution rate.

This is a mechanism, not a verdict. A Freshworks session that fails still leaves a human to finish the work, and that human costs more than any line on this chart. If your agent already resolves most conversations, the "only pay for success" meter is the more expensive one per conversation, and you are paying a premium for insurance against failures you no longer have. The broader trade between these models is set out in the analysis of what replaces the seat.

Where this scorecard is weakest

The scorecard reads public pages. That is its method and its main limitation, and three objections deserve a direct answer.

Contract terms beat public pages

Enterprise buyers rarely sign a pricing page. Sierra's criteria, Ada's and Decagon's rates, and probably Salesforce's resolution definition all exist in order forms I have not seen. A vendor can score badly here and write an excellent outcome clause for you. The scorecard measures what a buyer can know before the first sales call, which matters for a shortlist and much less for a signed deal.

The sample is also small and dated. Seven vendors and 12 meters, read on one day, in English. Billing documentation in this category has changed several times in 2026 alone, including Zendesk's tiers in May and HubSpot's switch in April. Treat every row as true on 2 November 2026 and recheck it before you rely on it.

Silence timers are a defensible proxy

The scorecard rewards Zendesk for verifying with a model and marks Fin, HubSpot and Gorgias down for relying on silence. The strongest counter is that silence is usually right. Satisfied customers rarely say thanks, and requiring confirmation would undercount real resolutions badly.

A model check is also not neutral. The verifying model belongs to Zendesk, its error rate is unpublished, and a buyer cannot audit an individual verdict. Nobody has compared Zendesk's verified count with Fin's assumed count on one queue. I would not claim the stricter definition produces a materially smaller bill until someone does.

Finally, the grouping involves judgement. Treating a Fin qualification or a HubSpot recommended lead as a completion rather than a result is my call. A sales team that values a routed lead in its own right would score both meters as outcomes, and the headline would read 6 of 12.

Frequently asked questions

How to read the next outcome-based price card

Put three questions to any vendor quote this week, in writing, and score the answer against the table above.

What exactly triggers the charge, and is that a customer result or an agent activity? Where is that definition published, with a date? What does a failed, escalated or reopened conversation cost? A vendor who answers all three with a link has an outcome meter. A vendor who answers with a case study has a usage meter and a good story.

Then run the crossover sum on your own numbers. Take your piloted resolution rate, not the vendor average, multiply it by each per-resolution price, and compare the result with any flat per-session option. If the per-resolution meter is the more expensive one, ask for a cap or a committed rate before you sign. The contract language to hold that line is in the outcome pricing clause list.

Related reading