A leader sees a high containment rate and reports it as proof the AI agent works. Containment only counts conversations that never reached a human. It also counts callers who gave up. So you can pay for an agent that looks good on the dashboard while customers leave unhelped. In the containment rate vs resolution rate vs CSAT debate, report resolution first, with CSAT and containment beside it.

Three metrics, three different questions

Containment, resolution and CSAT each measure a different thing about an AI agent. Containment tracks whether a human was needed. Resolution tracks whether the problem was solved. CSAT tracks how the customer felt afterward. They can move in opposite directions, so none of them can stand in for the others.

  • Containment rate is the share of conversations the AI finishes with no human. Formula: conversations with no handoff divided by all conversations. It answers: did the AI keep it?
  • Resolution rate is the share of conversations where the customer's issue was actually solved. Formula: resolved conversations divided by all conversations. It answers: did the AI fix it?
  • CSAT is the customer's own rating, usually from a short post-contact survey. Formula: satisfied responses divided by total responses, times 100. It answers: did the customer like it?

This matters because measurement is where AI projects fail. Gartner looked at 432 AI customer-service use cases. Only a quarter showed positive ROI, and 42% could not be measured at all. If your dashboard has one number, you are probably in that 42%.

How containment rate is calculated and counted

Containment is a simple division. Take 1,000 conversations. If 700 end with no transfer to a human, containment is 700 divided by 1,000, or 70%.

The hard part is deciding what counts as contained. The usual rule is no transfer and no ticket opened for a person to follow up. Vendors draw that line in different places:

  • Some count a call as contained if the AI simply ended it.
  • Some exclude callers who hung up in the first few seconds.
  • Some count a callback request as contained, because no human spoke yet.

Ask each vendor for its exact rule before you compare numbers.

Market claims also blur the terms. A pitch may say an agent "resolves" a share of inbound calls when it only describes where calls ended up. Nothing in that tells you the customer's problem was fixed. Treat such a figure as a containment number until someone shows you the transcripts.

Why a high containment rate can mislead

A high containment rate can hide failure, because it counts every conversation that ended without a human, whatever the reason. A caller who hangs up in frustration counts as contained. So does one stuck in a loop of "I didn't catch that" until they quit.

The metric also rewards the cheapest outcome. Say a team is told to push containment from 70% to 85%. The easiest way is to make the human option harder to find. The agent stops offering a transfer, and the number rises while customers get less help.

Customers expect that exit to exist. Gartner found that 87% of 3,566 customers say a company using GenAI for service must offer a way to reach a human. So a containment rate that climbs because the escape hatch is blocked is a warning sign, not a win.

Buyers in regulated sectors raise handover design on their own. They want to know if it is polite, what happens on a disputed amount or sensitive issue, and whether anyone chases a handover nobody picked up. Our advice is to track the handover as closely as containment itself.

Resolution rate: what it needs to be honest

Resolution rate is only honest when something other than the bot decides the issue was solved. You can measure it three ways, alone or combined:

  • Customer confirms. The agent asks whether that fixed it. This is fast, but people who gave up often stay silent, and some say yes just to leave.
  • No repeat contact within 7 days. A customer who comes back about the same issue was not helped. This catches quiet failures, but it arrives late. Match the window to how long your issue type takes to resurface.
  • QA review of the transcript. A reviewer checks the call against the customer's intent. This is the most accurate method, and it is costly unless every call is scored.

Resolution can still mislead. The AI may mark its own work resolved. The intent may be mislabelled, so a right answer to the wrong question counts. A call can also be perfect on procedure and still leave the customer angry.

The prize is large. In 2026 surveys, 69% of people who prefer humans said they would switch to AI if it fully resolved the issue. The fight is resolution, not channel.

So judge resolution by reading calls. Score every conversation against a weighted scorecard instead of a sample. Let one critical error fail the whole call.

CSAT: useful, but a thin and skewed sample

CSAT is a useful signal, but it comes from a small and lopsided group of customers. Use a 1 to 5 survey. CSAT % is the number of 4 and 5 ratings divided by all responses, times 100.

Few customers answer, so the score describes them and not everyone. The very happy and the very angry reply, while the quiet middle does not. If the survey only appears at the end of a successful flow, callers who gave up never see it. One score also cannot say which step failed, and a different question or channel gives a different number.

So read CSAT per intent, next to the transcripts. A billing-dispute score and a password-reset score should never be averaged into one figure. Recurring CSAT survey calls can also feed retention work, as one input beside call evidence.

For the customers who never answer, AI can infer sentiment from the call itself. The Sentiment Analysis Tool returns sentiment, emotional tone and friction points from an uploaded conversation. Treat that as a cross-check, not a replacement.

Read the three together: a quick decision table

Read the three numbers as a pattern, not as separate scores. Each row below pairs a pattern with its likely cause and your next move.

Containment Resolution CSAT Likely cause Next action
High Low Low or missing Callers give up or loop, and the survey never reaches them Read abandoned calls. Make the human option easy to find.
High High Low The answer is right, but the tone or effort is wrong Check transcripts for long steps, repeated questions and cold wording.
Low High High Healthy. Humans handle the complex work. Keep it. Watch cost per resolved conversation.
Falling Rising Steady or rising The agent hands off hard cases sooner Usually fine. Confirm with a transcript sample.
High High High A real win, or the bot is marking its own work resolved Audit a sample of "resolved" calls by hand.

When the pattern is unclear, let resolution break the tie. It is the one number that says whether the customer's problem was fixed.

Score the calls to check all three

Audit the agent's "resolved" claims by scoring the calls yourself. Insight7 evaluates 100% of customer conversations, in place of a 2% manual sample. Each call is scored against a weighted scorecard your team owns, and resolution can be one of the scored criteria.

Starter rubrics on the Criteria screen give you a scorecard to edit instead of writing one from scratch. Alerts can watch for low scores and send a weekly digest of the weakest calls by email or Slack. Your QA manager can then read the worst "contained" calls first and check whether the bot's resolved flag was right.

Which metric to report to leadership

Lead with resolution rate, verified by transcript review. Put CSAT beside it. Show containment as a cost and capacity metric, next to the escape-hatch rate, which is the share of conversations where the customer asked for a human.

Leaders want proof now. Gartner reports that customer-service AI spending rose 38% while overall service budgets rose 2%. That money needs an outcome behind it, and with 42% of use cases unmeasurable, a verified resolution number is your best defence.

Add three supporting numbers:

  • Cost per resolved conversation. Total agent cost divided by verified resolutions.
  • Repeat-contact rate. The share of customers who return within 7 days.
  • Handoff quality. How often a transferred customer got help without repeating themselves.

Cost per resolved conversation also helps a champion carry the case upstairs, because buyers weigh it against their own cost to serve a customer.

Use one page with three headline tiles and a trend line for each. Review weekly in operations and monthly with leadership. Never blend the three into one score, because a blended number hides the failure you need to see.

Related reading: How to Turn Failed Customer Calls Into Test Cases for Your AI Voice Agent, AI Voice Agent QA Scorecard Template With Criteria for Accuracy, Compliance and Customer Experience and How to Improve AI Voice Agent Performance From Real Customer Conversations.

Questions people also ask

What is containment rate in AI customer service?

Containment rate is the share of AI conversations that end without a human joining. The formula is conversations with no handoff divided by all conversations. It measures how much work the AI kept, so it speaks to capacity and cost. It says little about whether customers got help.

What is a good containment rate for an AI agent?

There is no single good number, because vendors count containment differently. Set your own baseline per issue type, such as password resets versus billing disputes. Raise it only while resolution and CSAT hold steady. A rate that rises while those two fall is a bad sign.

What is the difference between containment rate and resolution rate?

Containment asks whether a human was needed. Resolution asks whether the problem was solved. A caller who hangs up angry is contained but not resolved. A transfer that ends with a fixed issue is resolved but not contained. Resolution matters more.

How do you calculate resolution rate for an AI agent?

Divide verified resolved conversations by all conversations in the period. Decide "resolved" from transcript review or customer confirmation, never from the bot's own flag. Remove any customer who returned on the same issue within 7 days. Report the result per intent, since averages hide weak spots.

Is CSAT a good metric for AI agents?

CSAT is a useful check, but a weak main metric. Few customers answer, and callers who gave up rarely see the survey. Use it per intent beside resolution and transcripts. It shows where customers feel friction. Read those calls to learn why.

Can containment rate be too high?

Yes. A very high rate can mean customers cannot reach a person. Disputes, fraud and complaints often need a human, so a rate near 100% is suspicious. Read the abandoned calls. If the escape-hatch rate falls as containment climbs, the agent may be blocking exits.

What is first contact resolution and how does it relate to resolution rate?

First contact resolution is the share of issues solved in the customer's first contact. It is resolution rate with a time test added. If the same customer returns on the same issue within your window, such as 7 days, the first contact did not count as resolved.

How do you measure customer satisfaction for AI voice calls?

Offer a one-question survey right after the call, by SMS or a spoken prompt, on every ending, including transfers and hang-ups. Keep wording and scale the same each time. Add AI scoring of the call itself to cover customers who never reply.

Which AI agent metrics should you report to executives?

Report verified resolution rate, CSAT per intent and containment with the escape-hatch rate. Add cost per resolved conversation and repeat-contact rate. Keep it to one page. Lead with whether customers were helped and what each help cost.

Say it in one line

Resolution shows the customer's problem was fixed, containment shows how much work the AI kept, and CSAT shows how customers felt. Keep all three on the page, and check "resolved" by reading calls, not by trusting the bot's flag. Rule of thumb: resolution first, CSAT beside it, containment for cost.

Book a demo of Insight7