Per resolution, per outcome, per action: what AI support pricing actually charges for

The finding first. Intercom's pricing page lists Fin at $0.99 per outcome, not per resolution. Its help centre defines an outcome as one of two things: a resolution, or a procedure hand-off, which is a conversation Fin passes to your team because a procedure you configured told it to. Both cost $0.99. A hand-off that Fin makes under your workspace escalation rules instead costs nothing. The customer reaches the same person on your team either way. What decides the bill is which screen the instruction was written on.

This is not brand new, which makes it more surprising that so few buyers know it. Intercom's help article has described procedure hand-offs as billable outcomes since at least 31 March 2026, according to the Internet Archive. In January the same article was titled "Fin AI Agent resolutions" and did not mention hand-offs at all.

Meanwhile Helply, whose pricing page on 7 September said it charged for outcomes, has replaced that page with credits per AI action, and added an answer explaining why it thinks per-resolution pricing is the wrong unit. The two moves point in opposite directions. Between them they make an abstract question concrete: when you buy AI support, what exactly are you paying for?

Four ways to sell AI support, and who pays for a failure

Every pricing model answers one question, usually without saying so: when the AI gets a conversation wrong, who pays?

Per seat

You pay the same whatever the AI does. A failure costs nothing extra on the invoice and all of it in your team's time. Zendesk is partly here: every agent seat carries a monthly AI allowance of $2 to $10, depending on plan.

Per action or usage

You pay each time the AI does something, whether or not it helped. The vendor carries none of the failure risk, but the unit is objective: an action ran or it did not. eesel, Helply, Salesforce's Agentforce and Crisp's Hugo sell this way.

Per resolution

You pay only when the AI succeeds, so the vendor carries the failure risk. In return the vendor decides what success means, and that decision is where the risk can quietly come back to you.

Per outcome

Per resolution, plus other events the vendor counts as having delivered value. Intercom is the example, and the other event is a hand-off.

None of these is dishonest in itself. What matters is whether the unit on the invoice matches the unit in the marketing, and whether you can check it.

What seven vendors charge for

From each vendor's own pages, read on 23 September 2026. Where a vendor does not address something, the cell says so rather than guessing. Follow the links in the sources rather than trusting the table: one of these vendors replaced its whole pricing model this month.

Two things stand out. Zendesk is the only per-resolution vendor here that does not bill silence on its own. Since 18 May 2026 a conversation has to pass a language model's reading of the transcript to count as a verified resolution. That deserves more credit than it gets. It is still the vendor's own model marking the vendor's own work, and Zendesk does not publish what a verified resolution costs.

Both Zendesk and Intercom also let you open every conversation behind the count, from the allowance dashboard and Fin's performance funnel respectively. That is what makes the method at the end of this piece possible, and not every vendor offers it.

The second is that the per-action vendors are the easiest to predict and the hardest to argue with. They do not claim to sell outcomes, so there is no gap between the claim and the charge.

The case for per-action pricing, made fairly

Helply's argument, from its own pricing page, runs like this: per-resolution pricing charges nothing for the AI that helps your team all day, and then sells that help back as a separate per-agent add-on. Intercom's page makes the point for it. Copilot, Intercom's AI assistant for your agents, is $29 per teammate per month after ten free conversations, on top of the seat price and $0.99 per outcome.

Per-action pricing is also objective. eesel says it plainly: AI resources are used on every task, so every task is billed, whatever the result. Helply publishes its credit weights and commits to keeping existing customers on their current table until renewal if the weights change. Nobody has to interpret a customer's silence.

The weakness is that you carry every failure. An AI that spends twelve credits investigating its way to the wrong answer costs more than one that spends one credit getting it right, and the vendor has no financial reason to make its AI do less. The better per-action vendors soften this. Helply does not charge for failed actions, retries or drafts your team rejects, and its spending cap defaults to the plan price. eesel's default cap is $250 a month, and its agents pause when they reach it.

The case for per-resolution pricing, and where it leaks

Per-resolution pricing moves the failure risk to the vendor. If the AI is poor, the vendor does not get paid. That is the strongest alignment of incentives on offer, and the reason buyers like it.

The leak is the definition. A vendor that bills per resolution also decides what a resolution is, and the cheapest signal available is silence. Fin and Help Scout both count a conversation in which the customer did not ask for more help. Intercom's help centre addresses the hardest case directly: a customer who is frustrated with Fin's answer and simply leaves is counted as an assumed resolution, and billed at $0.99. Help Scout is stricter about people, since any help from a human removes the charge, but its definition still rests on the customer not asking again. We set out the mechanism, and what it does to a headline rate, in an earlier piece on Fin.

What that costs in practice was described by an Intercom customer, posting as bosbeest, on Intercom's own forum on 24 April 2025. During a server outage, the team replied straight away to every affected customer rather than let Fin walk them through troubleshooting steps that could not work. Because none of those customers had asked for a person first, each reply counted as a Fin resolution: Each of these replies cost us $1. On 18 May 2025 an Intercom support engineer in the same thread agreed that stepping in should be seen as good judgment, not something to be penalised. Sixteen months later, the help centre's answer on a teammate joining is unchanged. We found no public complaint yet about procedure hand-offs specifically, which may say more about how few customers have noticed than about how many are affected.

When a hand-off becomes an outcome

Intercom's reasoning deserves stating properly. In its own example, a customer asks for a refund. Fin checks eligibility against your shipping and billing systems, then hands the conversation to your team for final approval. Fin has done real work that a person would otherwise have done, and charging for it is defensible. But notice what kind of argument it is: payment for work done, whatever the result. That is the per-action case, priced at the per-resolution rate.

1. Fin hands the conversation to your team
2. Did the customer ask for a person? Yes Not billed
3. Did Fin escalate by its default behaviour, or by your workspace escalation rules? Yes Not billed
4. Did the procedure fail before it finished? Yes Not billed
5. Billed as a procedure hand-off, $0.99 The hand-off was the configured end of a procedure: guidance written inside the procedure, or a hand-off step. The customer reaches a person on your team, exactly as they would have by any of the free routes above.

Two consequences follow. First, the bill and the resolution rate now measure different things. Intercom says Fin's resolution rate is unaffected by the change, which means procedure hand-offs are billed but not counted as resolutions. You can no longer divide a Fin invoice by $0.99 to find out how many conversations Fin resolved.

Second, where you write a hand-off instruction now changes your bill without changing anything the customer sees. If you run Fin, list the procedures that end in a hand-off and ask of each one whether the work Fin does before handing over is worth $0.99. Some will be: a checked refund eligibility is. A procedure that asks two questions and passes the customer on probably is not, and the same hand-off written as a workspace escalation rule costs nothing.

How to compare prices across different units

List prices in different units cannot be compared. $0.40 a task, $0.99 an outcome and four credits an action are not the same kind of number. One number can be compared across all of them: what you paid per conversation that was actually resolved.

A worked example

This is an illustration, not data. The prices are each vendor's published ones, read on 23 September 2026. The conversation counts are invented to show the arithmetic, and deliberately identical for every vendor. Suppose the AI takes part in 1,000 conversations in a month. On a careful reading, 350 were really resolved. Another 100 ended with the customer going quiet unresolved, and 50 were rescued by your team before the customer asked. Fin also runs a procedure that ends in a hand-off 50 times.

Three things can be read off it. Fin's list price is 32% above Help Scout's, but per real resolution it costs 61% more, because its definition bills the rescues and the procedure hand-offs as well. Zendesk and Agentforce are missing because neither publishes the number the arithmetic needs: a price per verified resolution, or the actions in a typical conversation.

Second, a per-action bill does not move when the AI gets worse, but its cost per real resolution does. If only 200 of the 1,000 had really been resolved, eesel's $400 would be $2.00 per resolution.

Third, compare whole bills, not units. Hugo's usage is by far the cheapest line here, and it runs on a Crisp plan whose price is not in the table. So do Intercom's and Help Scout's seats. Helply's figure already includes its help desk.

How to work out your own

It takes an afternoon and nothing but your own data.

1. Add up one month of everything you paid for AI support: the usage bill, plus any seat price, add-on or allowance that exists only because of the AI. Copilot seats count. So does the part of a plan price that buys credits.
2. Export every conversation the AI took part in that month, not only the ones you were billed for.
3. Take a random sample of 40. Do not choose them by eye.
4. Read each one. Mark it resolved only if the customer's problem was dealt with and nobody on your team took part. A customer who went quiet is unknown, not resolved.
5. Multiply the share you marked resolved by the total number of AI conversations. That is your estimate of real resolutions.
6. Divide the month's spend by that estimate. That is your cost per real resolution, and it is comparable across every vendor in the table, whatever unit it bills in.

A sample of 40 tells you whether you are near the rate your vendor reports or a long way from it. It is accurate to within roughly 15 points either way, which is enough to make a decision and not enough to quote to a decimal place.

Before you sign with anyone, ask four questions and get the answers in writing. What is the unit? Is a customer going quiet billed? Which hand-offs are billed? Can I see the list of conversations behind every line of my bill? A vendor that answers all four on a public page has done most of your checking for you.

Where we stand

Varley sells AI support, so it competes with every product above, and this section is the pitch. Everything before it can be checked against the vendors' own pages without taking our word for anything.

Varley uses two units and says which is which. A conversation it finishes on its own, end to end, is billed as a resolution at $0.50, under a definition that is published, dated and versioned. Silence never counts, and a hand-off to your team is always free, whatever caused it and however much work came before it. Work Varley does alongside your team, such as drafting a reply or answering a question, is billed per action on a published price list, with nothing per seat. We think Helply is right that help for your agents should not be sold back per seat, and that Zendesk is right that silence is not evidence. We have tried to price accordingly.

The weakness is the one every per-action vendor shares: you pay for Varley's assistance whether or not a draft was any good. Actions your team rejects are free, and so are failed or cancelled answers. Everything else on the list is billed, and the list is marked as early-access prices, dated.

Right of reply

None of the companies named here was approached before publication. Every claim is taken from their own public pages, read on the date given in the sources. If anything on this page is wrong or has changed, write to support@dragapp.com and the correction will be published here with a date, including if it is inconvenient for us.

Changes to this piece

23 Sep 2026

First published.

Sources

• Fin AI Agent outcomes, Intercom help centre, updated 23 September 2026. Read 23 September 2026.
• Fin pricing: Outcomes, Fin help centre. Read 23 September 2026.
• The same Intercom article in the Internet Archive: 6 January 2026, titled "Fin AI Agent resolutions", and 31 March 2026, the earliest capture describing procedure hand-offs. Read 23 September 2026.
• Fin's flawed assumed resolved and pricing design, Intercom Community. Thread started 22 January 2025; the customer's reply quoted is dated 24 April 2025 and the Intercom support engineer's 18 May 2025. Read 23 September 2026.
• Intercom pricing, including the "What counts as an outcome?" panel and Copilot. Read 23 September 2026.
• About automated resolution tiers, Zendesk help, edited 24 August 2026. Read 23 September 2026.
• Managing resolution allowances, Zendesk help, edited 8 September 2026. Read 23 September 2026.
• AI Resolutions pricing, Help Scout docs. Read 23 September 2026.
• Hugo pricing and billing explained, Crisp help centre. Read 23 September 2026.
• Helply pricing, including its FAQ. Read 7 and 23 September 2026.
• eesel AI pricing, including its FAQ. Read 23 September 2026.
• Agentforce pricing, Salesforce. Read 23 September 2026.