Outcome pricing arrived in support, and only in support

In June I wrote that almost no one can charge for outcomes: two products on the whole AI Credit Index panel billed a result rather than a seat, a credit or an action, and both were in customer support. Edition 2 adds 24 products, five of them in support, and the count of outcome-priced products has doubled to four. Every one of the four is still in support. Nothing outside the category has moved to the result, and the two vendors that reached furthest have both pulled back: Salesforce from a $2 conversation to a $0.10 action with seat options beside it, eesel from pure per-task billing to prepaid credit batches.

That is a stronger finding than the June one. It is not that outcome pricing is rare. It is that outcome pricing works in exactly one place, for reasons that are about the work rather than the vendor, and the eighteen-month record of the products that tried to leave that place is the evidence.

The four that bill the result#

Fin is the reference case: $0.99 per resolved conversation since 2023, nothing charged when it fails to resolve, standing beside Intercom’s seat plans at $39, $99 and $139 rather than replacing them. It has not moved on the continuum since January 2025. What has moved is around it: in September 2026 Intercom added a Fin Pro add-on at $99 a month for 2,000 credits, with extra credits at $0.01, a second meter denominated in credits beside the per-outcome price.

Zendesk converted its installed base to “automated resolutions” before the study window began and has held there. The rates were never on the public page; in September 2026 they appeared in the page’s own data, $1.50 committed and $2.00 pay-as-you-go, with plans including 5, 10 or 15 resolutions a month, from $25 a seat on Support Team to $149 on Suite Professional, and Enterprise priced on request.

Help Scout is the migration. On 15 January 2025 its page sold contact-based plans, Standard at $50 and Plus at $75 a month for 100 contacts with unlimited users, and said in so many words: “Help Scout has no per-seat or AI usage fees. We earn revenue when you help customers.” An AI-resolved visitor simply counted as a contact. That model was still on the page on 6 November 2025. By 14 December it had returned to per-user plans, now $30, $54 and $90 month to month, with “AI Answers Add-on $0.75 per resolution”, a user-set monthly cap and a three-month trial. Nothing included. The resolution is the whole AI price.

Gorgias was already outcome-priced in December 2024, selling monthly allotments of “interactions that get completely resolved by Automate features, without agent involvement”, from 30 for $30 with $2.00 overage up to 5,000 for $2,500 with $1.00 overage. On 28 May 2025 it changed what an interaction costs the customer: each one “also counts as a helpdesk ticket”, where before a ticket drew a ticket fee or an automation fee, never both. By December 2025 the add-on was “$1.00 per resolved conversation” on Starter and $0.90 on the plans above, and the October 2026 page puts AI Agent on every plan at $1.00 per automated interaction monthly and $0.90 annual, though it also carries conflicting rates from $0.85 to $1.50. The rung never moved; the unit’s price did.

Four vendors, four slightly different units, a range of $0.75 to $2.00. Three of them have changed the economics of the unit since January 2025 without leaving it. That is a category with a working price, being tuned.

The two that reached and retreated#

Agentforce launched at Dreamforce 2024 charging $2 per conversation, a coarse, outcome-flavoured unit. On 15 May 2025 Salesforce overhauled the model: Flex Credits at $0.10 per action became the recommended meter and seat-based alternatives were added. The priced unit moved from a $2 conversation to a $0.10 action. By September 2026 the pricing page lists a “Help Agent Resolutions $2” SKU again, beside the Flex Credits, with 30 credits per voice action made official and “no overage penalty … contracted rate billed monthly in arrears”. The outcome price survives as one SKU in a credit system, which is a different thing from being the price.

eesel AI is the cleaner experiment because it ran both ways inside the window. In January 2025 it sold capped plans: $49 for 500 interactions, $299 for 1,000, $799 for 3,000, with the AI stopping at the cap until you bought more. In April 2026 it replaced them with per-task pay-as-you-go: $0.40 per support ticket or chat, $4.00 per blog post, no platform fee, no seats, no minimum, $50 of free usage. That is as close to pure usage pricing as the panel has seen. Between 20 September and 3 October 2026 it pulled back: Teammate plans from $299 for 500 credits, one credit per ticket or chat, in steps to 5,000, with agents pausing at exhaustion unless opt-in overage at $0.80 a credit is switched on. Six months from a hard cap to per-task pricing, one rung short of outcomes, and back to a prepaid batch with overage.

Salesforce’s change followed a backlash against the conversation price; eesel did not say why. The shape says enough: a per-task or per-conversation price with no floor gives the customer a bill that tracks volume exactly and gives the vendor revenue that tracks it exactly too, in both directions. The products that stayed at the outcome, Fin and Zendesk and Gorgias, all keep a seat or a plan underneath it. eesel, which made usage the whole price, came back for a floor; Salesforce kept a usage meter but put seat-based options beside it.

Why support, and only support#

The June essay gave the general reason: to charge for an outcome you need a unit that is countable, attributable and roughly equal in value each time. What you charge for matters more than what you charge, and most AI work fails the test, because the vendor cannot tell whether the output was any good.

Support passes it in a way nothing else on the panel does. A conversation either ends without a human or it does not; the system can see which. The customer’s own helpdesk already counts tickets, so the vendor’s count is auditable against a number the buyer trusts. And a resolution has a known alternative cost, the loaded minutes of a human agent, so a dollar a resolution can be sold against a figure the buyer already has. Coding agents, media tools and workflow builders have none of those three. A generated clip has no “resolved” state. A workflow run that completed may have done the wrong thing. There is no independent count to reconcile against.

So the finding is not that support vendors are braver. It is that the unit exists there and not elsewhere, and the vendors elsewhere know it, which is why every one of the 19 new products outside support arrived with a credit or a seat and none with a result.

What the four are doing to the unit#

If the category is settled on the resolution, the action has moved to the definition of one.

Gorgias’s May 2025 change, under which a resolved interaction also consumes a helpdesk ticket, raised the effective price of a resolution without touching the dollar figure. Zendesk’s pay-as-you-go rate is a third above its committed rate. Help Scout went from a resolution counting as a contact, which was free at the margin inside your tier, to $0.75 each in arrears. Fin’s new credit add-on sits beside the $0.99 and bills things that are not resolutions. Freshdesk, which prices sessions rather than resolutions, repriced the session five-fold in late 2025, from $100 per 1,000 to $49 per 100, and then redefined an email session as a 72-hour window in mid-2026.

This is the same pattern the rest of the panel shows with credits, applied to outcomes: hold the headline unit price, change what the unit contains. A buyer comparing “$0.99 per resolution” with “$0.75 per resolution” is comparing two numbers whose definitions differ, and the definitions moved more this year than the prices.

What to do with this#

If you buy support AI, the price per resolution is the start of the comparison, not the end. Ask what counts as resolved, whether a resolution also consumes a ticket or a seat, what the committed rate is against pay-as-you-go, and what the plan underneath costs. The four definitions above are on each product’s Index page with the archived page they came from.

If you sell AI anywhere else, the eesel and Agentforce records are the most useful eighteen months of evidence on the panel. Salesforce priced the conversation and retreated to a finer per-action meter with seat options beside it; eesel priced pure usage with no floor and returned to a prepaid unit with overage. If you are going to charge for outcomes, keep a plan underneath, and be sure you can count the outcome against a number the customer already trusts.

If you are building the category, the resolution is the unit, and the next three years of pricing change in support will happen in its definition. Track the definition, not the dollar.

Every figure here comes from the pages captured in the Edition 2 window or the archived pages cited in each product’s trajectory review, and the eesel and Agentforce moves are dated in the change log.