Guide

How AI support vendors count a resolution

If you pay an AI support agent per resolution, the vendor's definition of resolved is part of the price. Put the published rule beside your own ticket history before you check the bill.

Take one published rule. Fin lists $0.99 per resolution. The rule, in short: the customer either confirms the answer was satisfactory (confirmed resolution), or exits the conversation without requesting further assistance (assumed resolution). The clock: 24 hours. The correction: deducted if the customer returns to the same conversation, even across billing periods. Those few lines decide what the price buys.

When you pay per resolution, the definition of resolved is part of the price. Two teams can pay the same rate and buy different things, because the rules count different conversations. So read the rule first. Then put it beside your own tickets.

There are always two rules in play. One is the vendor's published rule. The other is the rule your team would write if you wrote it down. Most of the time they agree. The useful part is the short list of conversations that count under one and not under the other.

Four things your resolution definition has to say

The general outcome definition has four fields. For a support team, they become four practical questions:

  1. What ends the conversation? A customer can say yes, or a period of silence can end it.
  2. What is the window? The rule needs a clock for waiting, checking, or reopening.
  3. What undoes the count? A reopen, a later human handoff, or the customer returning can change the answer when the published rule says so.
  4. Who decides? It can be the vendor's rule, a second model, or a rule your team defines.

Published resolution definitions, side by side

These rows come from our outcome definitions registry. Each rule is a short summary of what the vendor publishes, with the date we read it. The sources are listed under Notes and on each vendor's page. A field says “Not published” when we found no published wording for it. Follow the source before you rely on a row: vendors change these pages.

Vendor and productPricePublished rule, in shortWindowCorrection ruleAuthority
Fin (formerly Intercom)
Fin

resolution. Read .
$0.99 (Published list)the customer either confirms the answer was satisfactory (confirmed resolution), or exits the conversation without requesting further assistance (assumed resolution)24 hoursdeducted if the customer returns to the same conversation, even across billing periodsVendor rule
HubSpot
HubSpot Customer Agent

resolved conversation. Read .
$0.50 (Inferred)
from $1.00 per conversation, effective 2026-04-14
resolved when the agent handles it with no human handoffNot publishedNot publishedVendor rule
Lorikeet
resolution. Read .
$0.99 (Published list)
Start plan, chat, email, and SMS; $0.90 on the Scale plan; voice $1.50 and $1.20; plans billed annually
the customer decides what counts as resolvedNot publishedNot publishedBuyer-defined rule
Salesforce
Salesforce Help Agent

resolution. Read .
$2.00 (Inferred)
400 Flex Credits
against a published definitionNot publishedNot publishedVendor rule
Zendesk
Zendesk AI agents

contained resolution. Read .
$0.00 (Observed)
free
The AI agent handled the interaction to completion without the customer requesting further assistanceemail: 72 hours after the last email; messaging: 2 hours after the last message by default, adjustable up to 72 hours; voice: at hangupNot publishedVendor verification
Zendesk
Zendesk AI agents

verified resolution. Read .
$1.50 (Observed)
$1.50 on a committed pack, $2.00 pay as you go, reported second-hand; tiers verified
The AI agent responded to a customer’s question, and the customer did not request further assistance. [...] Conversations that pass this verification are considered a Verified resolutionemail: 72 hours after the last email; messaging: 2 hours after the last message by default, adjustable up to 72 hours; voice: at hangupNot publishedVendor verification

Who decides, and how to dispute a count

Two more fields matter once you check a bill: who decides that a conversation counted, and what the vendor publishes about disputing a count. Each answer carries the date we reviewed the vendor's pages.

Vendor and outcomeWho decidesHow to dispute a countReviewed
Fin (formerly Intercom)
resolution
The customer's reply, or the customer's silence, read under the vendor's rule. We did not find how a reply is classified as affirmative in the reviewed sources. SourceNot found in the sources we reviewed on 2026-09-19.
Zendesk
verified resolution
A large language model run by the vendor reads the conversation text. We did not find its criteria, threshold, or error rate in the reviewed sources. SourcePublished: contact the vendor's customer support to dispute the tier assigned to a ticket. We did not find a deadline, a credit rule, or a stated result in the reviewed sources. Source

A vendor appears here once we've reviewed its published policy pages field by field. The registry shows every field for each one.

What you can check in your helpdesk export

Your export usually has enough history to test a billed list against a rule. You don't need message text for any of this. IDs, statuses, flags, and timestamps are enough.

  • Reopened after the count: look for status history, a reopen flag, and timestamps after the resolution time.
  • Handed to a person later: look for assignee changes, handoff events, and the time each happened.
  • The same customer and issue counted twice: compare conversation identifiers, customer identifiers, subjects or categories, and nearby timestamps. Remove names and email addresses before sharing the file.
  • Closed by silence: compare the last customer message, the last agent reply, the close time, and any later reply. Satisfaction ratings can add context, but they do not replace the published rule.

Pull the billing month plus a week or two after it. Some published rules take a resolution back when the customer returns later, so the days after the month closes are part of the answer.

Expect findings in both directions. You may find conversations that were counted and don't meet the rule. You may also find conversations that meet the rule and were never billed. And a check that finds nothing is a good result: it means the count and your records agree.

How Spoolis fits

A Spool is a versioned acceptance policy. “Counts only if not reopened in 7 days and no person took over” is one such policy. The vendor's published rule can be another. Both are explicit, both can be read by either side, and the same evidence can be checked against each. The acceptance lifecycle keeps the window and later correction explicit, while the glossary defines the record each system consumes.

Check the bill

Once you have the vendor's rule and your own counted conversations, check your AI agent bill against your tickets.

Scope

This comparison records what each vendor publishes. It does not score vendors or interpret a contract. Prices and definitions can change, so each row includes the date the source was retrieved.

Notes

How AI support vendors count a resolution · Spoolis