What Counts as a Resolution in AI Support (and What Should Not)
Everyone is moving to pay-per-resolution. The interesting question is not the price. It is what counts as a resolution in AI support.
If a clarifying question counts as a resolution, you will be billed for confusion. If a frustrated customer who got a grounded-but-useless answer counts, you will be billed for damage. If an internal preview chat counts, you will be billed for your own QA.
I would rather lose the row than keep a definition that does that.
What counts as a resolution in AI support
A support conversation, chat, email, or phone, is a resolved conversation when three things are true at once.
The answer was grounded in your knowledge, not generated from the model's prior.
A human did not have to take over.
The customer was not waving a red flag: a failed attempt, an error, a stall, a person who is already angry.
If any one of those fails, it is not a resolution. You should not pay for it. We should not report it as one.
That is the definition we productized on August 4, 2026. A conversation is billable under per-resolution pricing only when the answer was grounded, the customer was not waving a red flag, and a human did not have to step in. Handoffs and dashboard previews are out. The product page has the short version. The longer version is the exclusion list, because that is where vendors hide.
What should not count as a resolution
A handoff
The AI tried, then a person finished it. That is assist, which is valuable. It is not a resolved conversation. If you pay per resolution, a handoff billed as a close is a tax on the hard tickets. This is the difference between AI for customer support that sits on the helpdesk you already have, and a bot that treats every transfer as a win.
A stall
"Which product are you on." "Let me look that up." An error, then an apology. Something was sent, so the dashboard counted an answer. Nothing was resolved. We stopped counting clarifying questions and errors as answers on August 3, 2026 for that reason. Internal data the bot collected no longer leaks into the reply. When there is no answer, it escalates. See the IrisAgent changelog for those two shipping notes.
A preview
Someone on your team tested the bot in the dashboard. That is QA. Billing it is charging you to inspect the thing you already bought.
A red-flag chat
The answer may have been grounded. The customer was still stuck, or already furious. A clean source citation does not make that a win. We disqualify those so the bill matches what actually got resolved. How we measure accuracy is the eval side of the same line.
A deflection that is only a deflection
The customer did not open a ticket. That can be good. It is not the same metric as a resolved conversation. Mix them and you cannot tell whether the AI closed the issue or hid it. What is ticket deflection is the other term; keep the two formulas apart.
Resolution rate vs deflection
Deflection asks: did a ticket get created. Resolution asks: did the customer's issue actually end.
Those can move in opposite directions. A bot that asks two clarifying questions and then gives up will look strong on deflection and weak on resolution. A bot that answers from a live status page during an outage, then stops, may create no ticket and still be the right behavior. A bot that writes a fluent wrong refund policy will resolve the chat and create a bigger ticket next week.
This is why how we measure accuracy uses two eval sets, not one. The resolution set: questions the AI should answer, scored for correct, grounded, and cited. The hallucination set: questions it must not answer. On that set, a confident wrong answer is a failure and a clean decline is a pass.
A vendor who only shows you the first set is showing you the half of the job that inflates the rate.
The four-way split
Outcome | What happened | Counts as a resolution? |
|---|---|---|
Grounded resolve | Answer cited your knowledge, no human, no red flag | Yes |
Handoff | A person had to take over | No |
Stall | Clarifying question, error, or no-answer that bought time | No |
Preview | Internal dashboard test, not a real customer | No |
If a vendor's resolution rate cannot be rebuilt from this table, you were shown activity.
Coverage is the other missing denominator
Even a clean resolution rate can lie if it is computed on a slice.
The AI saw 40 percent of cases. The dashboard still reports a resolution rate as if the sample is the operation. A 70 percent resolution rate on 20 percent coverage is a pilot. A 40 percent resolution rate on 90 percent coverage is a production system.
Ask for both numbers: of conversations the AI actually processed, how many were grounded resolves, how many were handoffs, how many were stalls. Then ask what share of the queue that processed set was.
The eval question I would ask this fall
Per-resolution pricing is about to get copied into every deck. Ask for the exclusion list before you ask for the unit price.
What happens on a handoff. What happens on a clarifying question. What happens on a preview chat. What happens when the customer is already angry. What share of the queue the rate was computed on.
If they cannot say, the number is a demo.
I would rather publish a smaller resolution count than a prettier one. What counts as a resolution is the product. The exclusions are how you keep it honest.
Frequently Asked Questions
What counts as a resolution in AI support?
A support conversation, chat, email, or phone, is a resolved conversation when three things are true at once. The answer was grounded in your knowledge, a human did not have to take over, and the customer was not waving a red flag such as a failed attempt, an error, a stall, or a person who is already angry. If any one of those fails, it is not a resolution and should not be billed as one under per-resolution pricing.
What should not count as a resolution?
Handoffs, stalls, preview chats, and red-flag conversations should not count. A handoff is assist, not a resolved conversation. A stall is a clarifying question, an error, or a no-answer that bought time. A preview is an internal dashboard test. A red-flag chat may have a grounded answer, but the customer was still stuck or already furious. Clarifying questions and errors stopped counting as answers in IrisAgent metrics on August 3, 2026.
How is resolution rate different from deflection?
Deflection asks whether a ticket got created. Resolution asks whether the customer's issue actually ended. Those can move in opposite directions. A bot that asks two clarifying questions and then gives up looks strong on deflection and weak on resolution. A fluent wrong refund policy can resolve the chat and create a bigger ticket next week. Keep the two formulas apart.
What is per-resolution pricing?
Per-resolution pricing bills only when the AI delivers a resolved conversation. The unit price is the easy part of the contract. The exclusion list is the product: no charge on handoffs, clarifying questions, errors, dashboard previews, or red-flag chats. If a vendor cannot rebuild their resolution rate from that four-way split, you were shown activity.
Why does coverage matter for a resolution rate?
A clean resolution rate can still lie if it is computed on a slice. A 70 percent resolution rate on 20 percent coverage is a pilot. A 40 percent resolution rate on 90 percent coverage is a production system. Ask both numbers: of conversations the AI actually processed, how many were grounded resolves, handoffs, and stalls, and what share of the queue that processed set was.
