I Read 20 AI Agent Pricing Pages. Nine Define 'Resolution'. All Nine Count Silence as a Win.

TL;DR: A resolution counts as whatever your vendor's definition says it counts as, and most vendors do not publish one. I checked 20 named AI support agent vendors against their own public pricing, docs and terms on 2026-08-15: 9 publish a definition of the unit they bill for, and all 9 of those definitions treat a customer who stops replying as a billable success. Intercom's is the plainest, counting an outcome when customers "don't ask for more help after Fin responds." At $0.99 per outcome, if 30% of billed outcomes are people who gave up rather than people who were helped (an illustrative rate, not a measured one), your real cost per actual resolution is $1.41.

Outcome pricing is the default pitch for support agents now. Pay per resolution, not per seat, and the vendor carries the risk of the agent being bad. The mechanism only works if both sides agree on what a resolution is, which is also why the category's most-quoted proof point, Klarna's assistant handling 80% of chats, is a number only Klarna can define.

What did I count, and how?

Twenty named vendors selling agent work priced on an outcome, resolution, conversation or session unit. For each I read only public pages: the pricing page, the public help center or docs, and the public terms. No sales calls, no trials, no signups. Every figure below was fetched on 2026-08-15.

Vendor	Public rate	Defined?	Silence counts?	Handoff bills?	Credit-back?	Overage policy?
Intercom (Fin)	$0.99/outcome	Yes	Yes	Yes	Yes	Yes
Zendesk	None	Yes	Yes	No	No	Yes
Salesforce Agentforce	$2/conversation	No	n/a	Not stated	No	Yes
Ada	None	Yes	Yes	No	No	No
Gorgias	$1.50/automated interaction	Yes	Yes	No	No	Yes
Freshworks (Freddy)	$49/100 sessions	Yes	Yes	Yes	No	Yes
Decagon	None	No	n/a	Not stated	No	No
Sierra	None	No	n/a	Not stated	No	No
Forethought	None	No	n/a	Not stated	No	Yes
Kustomer	None	No	n/a	Not stated	No	No
Tidio (Lyro)	Tiered/conversation	Yes	Yes	Yes	No	No
HubSpot (Breeze)	50 credits/resolution	Yes	Yes	No	No	Yes
Help Scout	$0.75/resolution	Yes	Yes	No	No	Yes
Yellow.ai	$0.99/resolution	No	n/a	Not stated	No	Yes
Netomi	None	No	n/a	Not stated	No	No
Cognigy (NiCE)	None	Yes	Yes	Yes	No	Yes
Maven AGI	None	No	n/a	Not stated	No	No
Lorikeet	0.95 credits/resolution	No	n/a	Not stated	Yes	No
Siena AI	$0.90/ticket	No	n/a	Not stated	No	No
Crescendo	From $2.99/resolution	No	n/a	Not stated	No	Yes

Eleven publish a rate. Nine publish a definition. Only six publish both, so for 14 of 20 you cannot compute a price per unit of anything from public information.

Which vendors actually define the word?

Intercom's pricing page is the cleanest specimen because it is short: "An outcome is counted when: A customer confirms their issue is resolved, or They don't ask for more help after Fin responds, or Fin completes a workflow (Procedure), including handoffs." Clause two is the whole post. A customer who reads the answer, decides it is useless and closes the tab has not asked for more help.

Help Scout writes the same rule in the negative: "A resolution is counted only if the customer receives an AI response and doesn't use escalation, search the knowledge base, ask more questions, or indicate that they need more help." Gorgias uses a clock: "An interaction is considered automated if the customer does not require the involvement of a human agent within 72 hours of the interaction." HubSpot uses the same 72 hours, evaluated after "the last visitor response."

Zendesk's legacy definition says it in a bullet: a resolution is counted after 72 hours of inactivity when "the end user provided positive feedback ('Yes, problem solved') or provided no feedback." Zendesk's current AI agents add the only real guard in the sample, an LLM that reads the transcript, and conversations failing verification "do not consume an automated resolution."

Ada's service specific terms define an Automated Resolution as "a Conversation between Customer and End User that is not Escalated," qualified by relevance, accuracy and safety. Escalation is the only disqualifier, so abandonment sits inside the definition. Ada also publishes the sample's oddest clause: "Automated Resolutions will be calculated based on the average rate, using a sample size, at which Conversations result in Automated Resolutions." You are billed from an extrapolation.

Three of the nine are not resolution units at all. Freshworks bills a session, "a 72-hour window starting from the customer's first email." Cognigy bills a conversation capped at 50 user inputs in 24 hours. Tidio bills any interaction "that has at least one reply from our AI Agent." None of the three claims the agent solved anything, which is at least honest.

Why not just use the comparison table one vendor already published?

Because the cross-vendor price table at fin.ai is published by Fin, which is Intercom's own property, and it is a specimen of the problem rather than a solution to it. It lines up seven vendors in one price column as if the units were comparable, when three of the seven publish no rate at all and its figures for those are labelled "reported." Its Freshdesk row reads $0.10 per session; Freshworks' own page sells extra AI Agent sessions at $49 per 100, which is $0.49.

What does silence do to the price you were quoted?

The sticker is a price per billed unit. What you care about is the price per unit that worked. If a fraction f of your billed resolutions were customers who left rather than customers who were helped, your effective cost is sticker divided by (1 minus f).

Take Intercom's public $0.99 and 10,000 billed outcomes in a month, so $9,900 on the invoice. Assume f = 0.30. That 30% is an illustrative assumption, not a measurement: nobody in this sample publishes a false-resolution rate, and I have not run one. Then 7,000 outcomes were real, and $9,900 divided by 7,000 is $1.41 per actual resolution, 43% above the number on the page. At f = 0.10 it is $1.10. At 0.20, $1.24. At 0.40, $1.65. Run the same formula on Crescendo's $2.99 and a 30% rate takes you to $4.27.

This is unresolvable from outside because the one measurement that would settle it, the share of billed resolutions where the customer went away unhappy, is the number none of the 20 publishes. Same shape as the AI note-taker vendors who sell action items and publish no accuracy figure for them, and the same reason a per-credit sticker tells you almost nothing about a monthly bill.

Isn't outcome pricing still better than per seat?

Yes, often. A per-seat license for software whose output varies wildly is its own bad deal, and outcome pricing does move real risk. Five of the nine definitional vendors (Zendesk, Ada, Gorgias, HubSpot, Help Scout) explicitly do not bill when the conversation reaches a human, which is money the vendor forgoes. Two go further. Intercom reverses the charge if the customer returns to the same conversation "even across billing periods," the best buyer-protective clause in the sample. Lorikeet publishes the only unconditional one: "If you're unhappy with how Lorikeet handled a ticket, you don't pay for that ticket."

That is two vendors out of 20 with any documented credit-back at all. None of this requires bad faith from anyone. The structural fact is that the unit is defined by the party being paid for it, and definitions move: Intercom widened "resolution" into "outcome," and made completed handoffs billable, in a blog post dated March 12, 2026. The Wayback index holds 73 distinct versions of intercom.com/pricing and 199 of zendesk.com/pricing for 2026 alone. Every figure here carries a fetch date for that reason, and so should yours.

What should you get in writing before you sign?

Three things, in the order form, not the marketing page. First, the definition itself, including the silence clause, plus what happens when the customer reopens or files a fresh ticket on the same issue inside a window you name. Second, a dispute mechanism: how many billed units you may challenge, on what evidence, and by when. Eighteen of 20 vendors document nothing here, so it is a term you have to ask for. Third, the raw event export, one row per conversation with the resolution flag and timestamps, so you can compute your own false-resolution rate instead of reading the vendor's dashboard. Without that third one you are buying a number the seller both defines and measures, which is where buyers of AI infrastructure keep ending up.