Support AI bills per resolution, never per return

In mid-October a customer support lead put a quarterly report in front of me. Two numbers sat side by side: the AI resolution rate had climbed from 61 percent at the start of the year to 76 percent, and labor cost was ten percent higher than the same quarter a year earlier. He said both numbers were on one page and he had no way to explain them to his boss.

The numbers do not contradict each other. They measure two different stretches of time. Resolution rate measures the instant a conversation ends. Labor cost measures a month of human work. The stretch in between is unmeasured, and that is exactly where the money goes back.

The moment the 76 percent is measured at

Start with the definition. Fin documents resolution rate as the share of conversations closed without human involvement, measured at the moment the conversation ends. The vendor claims 76 percent across more than 12,000 customers, with strong implementations above 80 percent and a few reaching 93, at roughly two million resolutions a week. Salesforce repeated the same 76 percent in its September 10 acquisition announcement. It remains a vendor claim, not an independent benchmark.

Pricing is 0.99 dollars per resolution, charged only when the AI closes the conversation itself. That is a clever design: the vendor gets paid only when its own side of the work lands.

A second document from the same company is worth remembering. It states that under per-conversation pricing, a 65 percent resolution rate means 35 percent of AI spend goes to interactions that were never resolved. The sentence exists to sell outcome pricing, and along the way it admits something useful: the vendor understands that the billing unit decides who carries the cost of failure.

The trouble is that outcome pricing only aligns the vendor side. Two costs remain with the buyer. One is the human handling for everything the AI could not close, which most models do capture. The other is the second contact from a customer whose issue was marked resolved and came back a few days later. That one usually goes uncounted, because in no system is it labeled a cost. It is labeled a new ticket.

What Klarna reported, and what it admitted a year later

In February 2024 Klarna published numbers the whole industry quoted: 2.3 million conversations in the first month, 67 percent of all support chats, work equivalent to about 700 full-time agents, average resolution time down from 11 minutes to under two, 25 percent fewer repeat inquiries, and a projected 40 million dollars of profit improvement for the year. I have these through press summaries rather than the original release, so treat them as secondhand.

The 700-agent figure deserves its own look. 2.3 million conversations at 11 minutes each is 25.3 million minutes, about 421,000 hours a month. At 160 productive hours per agent, that is roughly 2,630 agent-equivalents, and 67 percent automation covers about 1,760 of them. That is two and a half times the announced 700. The gap comes from the denominator: conversation time, scheduled headcount, or only the share of chats an AI can fully absorb. I cannot say which basis is right. What matters is that one event produced three different equivalencies, and the chosen one traveled through the industry for a year as a result.

In May 2025 Bloomberg reported the same chief executive saying the cost-led approach had gone too far, and that customers should always have the option of reaching a person. Follow-up coverage was blunter on the mechanics: headcount had fallen from about 5,500 to roughly 3,000, and the company began hiring human agents again.

Put the two statements together and the interesting part is not that the AI failed. Both statements rest on the same metric set. The first used two-minute resolution and 25 percent fewer repeats as proof of success. The second used falling quality and unhappy customers as the admission. The repeat-inquiry number was good news in 2024, and in hindsight it simply did not cover the conversations that were deteriorating: emotionally charged ones, multi-step ones, ones that needed records from several systems, ones that needed an exception approved.

Aggregate metrics eat stratification. If nine of ten conversations are order lookups and address changes, a 67 percent automation rate looks excellent. How badly the remaining tenth collapsed is invisible in 76 percent, because both the numerator and the denominator dilute it.

Where the meter sits is who eats the cost

Strip it to the plainest layer. A ticket marked resolved comes back three days later for the same problem. The buyer paid for two human sessions, one AI session, and a slice of goodwill. On the buyer ledger that ticket has negative value. On the vendor ledger it is a clean resolution record.

Both ledgers can be true at once because the meter sits in different places. The vendor meters at the instant the conversation ends. The second contact falls outside the window.

No malice is involved. Any billing system has to put its unit on something observable and settleable, and the end of a conversation is the easiest thing in support to settle. The choice looks like an implementation detail. It is actually the thing that decides which direction cost moves.

Retail has the identical structure. Revenue is recognized when goods ship, and returns land afterwards, which is why retailers watch return rates instead of shipment totals. Support AI is at the stage of reporting shipments only, and reporting them convincingly, because shipments really are up.

Why would a vendor not stretch the window on its own? To know whether a ticket returned within thirty days, the vendor has to connect to the buyer ticket system and keep watching for thirty days, which puts its own failure rate into a contract. There is no commercial reason to volunteer for that. This is my inference rather than a documented finding, and it is testable: watch whether any vendor writes a no-return window into standard pricing before 2027.

That yields a criterion more useful than any benchmark score. Any AI service billed on a state at the end of a session will leave the cost of return with the buyer. Support is the first scenario that scaled, and code review, financial reconciliation and resume screening will reach the same billing question, because all of them can look finished and come back days later.

Running the arithmetic

Use the parameters from the vendor documentation: 50,000 conversations a month, eight dollars per human interaction, 0.99 dollars per AI resolution, 76 percent resolution rate.

AI cost is 38,000 times 0.99, or 37,620 dollars. The remaining 12,000 conversations go to humans at eight dollars, 96,000 dollars. The two together are 133,620 dollars, or 2.67 dollars per conversation. All-human handling is 400,000 dollars, eight dollars each. The saving is 66.6 percent.

Now add the third line. Assume 8 percent of the conversations marked resolved return within thirty days for the same issue. To be explicit, 8 percent is not survey data. It is a parameter I set to show how far one point moves the total, and an industry average would be useless for your budget. At eight dollars per human interaction, 38,000 times 8 percent is 3,040 returns, 24,320 dollars. Total cost becomes 157,940 dollars, 3.16 dollars per conversation, and the saving falls from 66.6 to 60.5 percent.

One convention is easy to miss. Eight dollars per human interaction is fully loaded cost including management and scheduling, not an agent wage, because a handed-off conversation also consumes queue time, QA and supervision. Budget models built on hourly wages systematically overstate what automation saves, and they always err in the same direction: human cost understated, automation overstated.

The rate of change is what matters. Every additional point of return rate costs 3,040 dollars a month, about 36,500 dollars a year. Every additional point of resolution rate absorbs 500 more conversations and saves roughly 3,500 dollars a month. The two numbers land in the same order of magnitude, and that is the part worth remembering: one step forward on resolution can be fully erased by one step forward on return rate, and the return rate never appears in a procurement deck.

It also explains the opening scene. Resolution rose fifteen points, which should have saved real money. Labor cost rose ten percent instead, because the return rate moved at the same time. Two variables pushed in the same direction, and only one of them reached the report.

Three changes the buyer controls

First, replace the acceptance unit with the repeat contact rate inside a thirty-day window. The contract language is concrete: same customer, same issue, reopened within thirty days counts as unresolved. This needs no product change from the vendor. It changes a definition, and definitions are something the buying side can put on the table.

Second, make the vendor share the second contact. It is technically feasible, since the vendor already sees ticket state; what is missing is willingness to widen the meter. A vendor that agrees has a real handle on its own quality. A vendor that refuses tells you what you are buying, which is better than signing a contract that looks cheap while leaving the largest cost line on your side.

Third, stratify the sample by issue type instead of sampling totals. Pull 200 resolved conversations, revisit them after thirty days, and have a person judge each. Stratify on three axes: emotional intensity, number of steps, number of systems touched. Report the return rate per stratum. The Klarna lesson is that aggregate metrics cannot see strata, and the worst stratum is usually where second contacts come from.

All three share one property: none of them requires switching vendors. Swap in another product that also meters at the end of a conversation and you buy the same structure with the problem deferred by a year.

What 3.6 billion dollars bought

The agreement was signed on June 15, 2026 and closed on September 10, with Salesforce acquiring Fin. The announcement disclosed the completion and a customer base of more than 30,000 companies. The deal value of roughly 3.6 billion dollars comes from CMSWire tracking of the transaction. The wording in the announcement is measurable outcomes.

Three things were bought: a resolution-rate measurement system already running at scale, an installed base of thirty thousand companies, and the right to define what counts as resolved.

The third is the most valuable and the least noticed by buyers. When the definition sits with the platform, renewal conversations are about price and never about definition. Writing acceptance criteria into a contract recovers part of that definition. Leaving them out means accepting the other ledger. On a single purchase the difference is small. After consolidation it becomes structural, because thirty thousand companies now share one measurement convention, and that convention still stops at the end of a conversation.

So my read on the deal is this: it did not solve the measurement problem. It made the measurement problem into infrastructure. Once an industry buys a ledger as the standard, correcting it costs far more.

Our judgment

Where the meter sits decides who carries the cost. That is a criterion for predicting the next case of the same kind, not a summary of this one.

Two predictions follow, both falsifiable. Within twelve months a second generation of outcome-priced agent products will appear that moves the measuring window from the end of a conversation to a number of days without a return, and the first vendors to do it will set price, because buyers can finally separate 76 percent from actually fixed. Second, products that keep metering at the end of a conversation will be asked to sign return-sharing clauses in their next procurement round. Check back in a year.

Three questions for anyone signing this contract now. What is your thirty-day return rate, and who measures it? How long is the measuring window, and is it in the contract? Have you stratified the sample, and which issue types resolve worst? Answer all three and you are buying an auditable service. Answer none and you are buying a notification that a conversation ended.

The same ledger applies well beyond support. Any automation that defines finished at its own boundary carries the same gap. You spot it not by the completion rate it reports but by where its measuring window stops.

Sources and method

Fin resolution rate of 76 percent averaged across more than 12,000 customers, about two million resolutions a week, and 0.99 dollars per resolution: vendor documentation, vendor claim, not independently benchmarked. fin.ai/learn/ai-resolution-rate and fin.ai/learn/per-resolution-vs-per-conversation-ai-pricing

Definition of resolution rate as conversations closed without human involvement: Intercom help documentation, primary. intercom.com/help/en/articles/8205718-fin-ai-agent-outcomes

Completion of the Fin acquisition, customer base above 30,000 companies, and the measurable outcomes wording: Salesforce press release, primary, September 10, 2026. salesforce.com/news/press-releases/2026/09/10/salesforce-completes-acquisition-of-fin

Deal value of roughly 3.6 billion dollars and the June 15, 2026 signing: CMSWire tracking of the transaction, secondary. cmswire.com/customer-experience/salesforce-acquires-fin

Klarna statement of May 2025: Bloomberg, primary, paywalled. bloomberg.com/news/articles/2025-05-08/klarna-turns-from-ai-to-real-person-customer-service

Klarna February 2024 figures, and the alternative agent-equivalent arithmetic: press summary, secondary, figures not independently verified. twig.so/blog/klarna-ai-customer-support-efficiency

Parameters for the worked example: eight dollars per human interaction and 0.99 dollars per resolution come from vendor documentation; 50,000 conversations and the 76 percent rate follow the same example; the 8 percent thirty-day return rate is a parameter I set and is labeled as such in the body, not survey data. Every division and percentage in this piece can be recomputed from those inputs.