What is the difference between resolution rate, containment rate and deflection rate?
Containment rate is the share of AI conversations closed without a person. Resolution rate is the share where the customer's problem was actually solved. Deflection rate measures customers who used self-service instead of opening a ticket at all.
Where each metric is measured
The three are often used as if they meant the same thing, and vendors add to the confusion by using "resolution" as a billing word. They measure different moments. Deflection happens before a conversation starts, containment is decided when the conversation ends, and resolution is only known later, when you see whether the customer came back. If you are planning an AI support agent, our page on AI customer service agents for existing customers covers the build; this page is about measuring it honestly.
| Metric | Question it answers | Formula | When you know it |
|---|---|---|---|
| Deflection rate | Did the customer help themselves instead of contacting us? | Zendesk's version: help center sessions / users who opened tickets, shown as a ratio such as 4:1 | Over a period, from help center and ticket data |
| Containment rate | Did the AI finish the conversation without a person? | AI conversations closed with no human reply / all AI-handled conversations | When the conversation closes |
| Resolution rate | Was the problem actually solved? | Contained conversations with no repeat contact on the same issue within your window / all AI-handled conversations | After your waiting window, for example 7 days |
Resolution rate is always equal to or lower than containment rate, because every resolved conversation was contained but not every contained one was resolved. A customer who gives up and closes the chat window is contained. They are not resolved.
How do Zendesk, Intercom and Gorgias define a resolution?
Zendesk counts a "verified resolution" only after a language-model check; conversations that fail it are "contained". Intercom counts a resolution when the customer confirms or stays silent for 24 hours. Gorgias counts a request no person touches for 72 hours.
Each vendor's definition is also its billing unit, so it is worth knowing which of the three metrics above it is closest to. This mapping is our reading of each vendor's help articles, checked October 30, 2026.
| Vendor term | Vendor definition | Closest to |
|---|---|---|
| Zendesk "contained resolution" | The AI agent handled the interaction to completion without the customer requesting further assistance; also where the LLM check fails | Containment |
| Zendesk "verified resolution" | Passed an LLM check of the conversation text for whether the request was satisfactorily resolved | Between containment and resolution |
| Zendesk "assisted escalation" | The AI contributed before a person completed the resolution | Neither; a handoff |
| Intercom Fin "confirmed resolution" | The customer confirms the answer was satisfactory | Resolution, at that moment |
| Intercom Fin "assumed resolution" | The customer disengages for 24 hours after Fin's last answer without asking for more help | Containment |
| Gorgias "automated interaction" | A request fully resolved without a human agent, with no human involvement needed within 72 hours | Containment with a 72-hour check |
| Zendesk "self-service score" | Help center sessions divided by users in tickets | Deflection |
Two vendor details matter here. Zendesk introduced its three tiers on May 18, 2026, and only verified resolutions draw on the billing allowance. Intercom deducts a resolution if the customer returns to the same conversation for more help, even in a later billing period. Neither catches a customer who gives up on chat and phones or emails instead. That is the gap your own resolution rate has to close. The billing side of these definitions is compared in our Zendesk AI vs Intercom Fin vs Gorgias pricing guide.
How do you calculate each rate from your own data?
Export one month of AI-handled conversations, count those closed without a person (containment), then remove any whose customer contacted you again on the same issue within 7 days on any channel (resolution). Keep the denominator the same for both.
One month, three numbers (illustrative)
Here is a worked example with illustrative numbers, to show how far apart the three can sit for the same month.
- 1,000 conversations start with the AI agent.
- 280 are handed to a person: the customer asked for one, the topic was on the high-risk list, or the AI failed twice.
- 720 close without a person. Containment rate: 720 / 1,000 = 72%.
- Of those 720, 90 customers contact you again about the same issue within 7 days, by chat, email or phone.
- 630 stay solved. Resolution rate: 630 / 1,000 = 63%.
- Separately, your help center had 4,000 sessions and 1,000 users opened tickets. Zendesk-style self-service score: 4,000 / 1,000 = 4:1.
Some teams divide resolved conversations by contained ones instead (630 / 720 = 87.5%). That number is useful for spotting early closes, but it is a different metric. Whatever you report, write the formula next to it.
How to spot a repeat contact
Match on the customer record, not the conversation. In most helpdesks and in a CRM such as GoHighLevel, every chat, email and call is attached to one contact. A simple rule works: any new conversation from the same contact within 7 days, tagged with the same topic, counts as a repeat. We usually build this as a small scheduled report in the helpdesk or an n8n workflow that tags the original conversation as "not resolved".
What is a good AI resolution rate for customer service?
There is no neutral public benchmark, because every vendor measures resolution its own way. Intercom says Fin averages 76% across 12,000+ customers by its definition. A better target is your own baseline, raised topic by topic.
Intercom's figure comes from its own website, checked October 30, 2026, with no date or method attached. By Intercom's definition, a resolution includes an assumed resolution after 24 hours of silence, so it sits closer to containment than to the resolution rate defined on this page. It is a useful reference for what Fin reports, and not a benchmark for your own 7-day number.
What a good rate looks like depends on your ticket mix. Order status and policy questions can be contained at high rates because they need one lookup and one answer. Refunds, complaints and technical faults should not be contained at high rates, because a person should handle them. A single blended number hides this. Report it by topic:
| Ticket topic | What to expect | What a high number may mean |
|---|---|---|
| Order status, hours, shipping policy | Highest containment and resolution | Working as designed |
| Booking changes, resend invoice | High, if lookups and actions are connected | Working as designed |
| Returns and exchanges | Middle; exceptions go to a person | Check exceptions are not being refused wrongly |
| Technical troubleshooting | Lower; many need a person | Customers may be giving up; check repeat contacts |
| Refunds, complaints, legal | Should be low by design | A problem: these should reach a person |
The topic-by-topic sort is the same one we use in an audit, set out on our customer service agent page. Our research on AI automation success rates covers what independent studies say about results more broadly.
Why can a high containment rate hide unhappy customers?
Because containment only records that no person got involved. A customer who could not reach a person, gave up, or came back on another channel still counts as contained. Without a repeat-contact check, a frustrating agent can report a high rate.
Containment rises when it is hard to reach a person. That is why it should never be the only number a support team is judged on. Watch for these signs:
- Containment goes up while CSAT on AI-closed conversations goes down.
- Phone or email volume rises after chat containment improves.
- Many assumed resolutions (silence) and few confirmed ones.
- Handoff rate near zero on topics that need judgment, such as complaints.
Each of these is fixed in the design: a clear path to a person, handoff triggers after two failed answers, and a "did that solve it?" question before closing. Our page on whether an AI agent will replace staff covers why people still handle the cases at the bottom of the table above.
Which AI support metrics should you report each month?
Report containment and resolution side by side, with the gap between them, plus CSAT on AI-closed conversations, handoff rate by topic and the error rate from a weekly sample of conversations. Add deflection if you run a help center.
| Metric | Why it is on the report |
|---|---|
| Containment rate | How much work the AI takes off the team |
| Resolution rate (7-day) | How much of that work was actually done |
| Gap between the two | Customers who gave up or came back; the number to shrink |
| CSAT, AI-closed vs human-closed | Whether automation is hurting the experience |
| Handoff rate by topic | Where content or rules need work |
| Wrong-answer rate from a weekly sample | Protects the brand; vendor reports do not show it |
| Vendor billable resolutions | What you pay for; reconcile it against your own resolution count |
Setting up this report, and keeping the agent's knowledge current as it shows gaps, is part of what support after a build covers.
More questions about AI support metrics
Is deflection rate the same as containment rate?
No. Deflection is about contacts that never happen: a customer reads a help article or uses self-service and does not open a ticket. Containment is about conversations that did start with the AI and ended without a person. A chatbot conversation that closes without escalation is contained, while a help center visit with no ticket afterwards is deflected. Track them separately.
What is Fin AI resolution rate?
Intercom defines Fin's resolution rate as the percentage of conversations Fin resolves without human involvement. A resolution is either confirmed by the customer or assumed when the customer leaves without asking for more help, with 24 hours of silence after Fin's last answer counting as assumed. Intercom states an average of 76% across 12,000+ customers on its website.
What is a Zendesk automated resolution?
It is the unit Zendesk uses to measure and bill AI agent usage. Since May 18, 2026, conversations fall into three tiers: assisted escalation, contained resolution and verified resolution. Only verified resolutions, which pass a language-model check after the conversation window ends, are drawn from your allowance. The window is 72 hours on email, 2 hours by default on messaging and the hang-up on voice.
Does silence mean the customer's problem was solved?
Not always. Some customers go quiet because the answer worked; others because they gave up. Intercom counts 24 hours of silence as an assumed resolution, and Zendesk runs a language-model check on the conversation text. Neither sees a call or email the customer makes afterwards, so add your own repeat-contact check across channels before calling a conversation resolved.
How do you raise an AI agent's resolution rate?
Work from the gap between containment and resolution. Read the repeat contacts, find the topic, and fix the cause: a missing or outdated help article, a lookup the agent cannot do, or an action it is not allowed to take. Add the question to a test set so it is checked before future changes. Raising containment alone, by making handoff harder, lowers real resolution.
Do these rates change what you pay?
Yes, on per-resolution pricing. Intercom charges $0.99 per Fin outcome, Zendesk US$2.00 per pay-as-you-go automated resolution, and Gorgias $0.90 to $1.50 per automated interaction, each using its own definition. A higher vendor-reported resolution rate means a higher bill, so reconcile the billed count against your own resolution rate every month.
Sources and further reading
- Zendesk help: about automated resolution tiers, checked October 30, 2026
- Zendesk help: about the automated resolutions platform (prior to resolution tiers), checked October 30, 2026
- Zendesk help: calculating your self-service score, checked October 30, 2026
- Intercom help: Fin AI Agent outcomes and resolution rate, checked October 30, 2026
- Fin help: Fin pricing, outcomes, updated September 17, 2026, checked October 30, 2026
- Fin website: average resolution rate claim, checked October 30, 2026
- Gorgias help: how you're billed for using Gorgias, checked October 30, 2026
- Gorgias pricing: AI Agent interactions, checked October 30, 2026
