Monday's dashboard says the AI agent "handled" thousands of calls, yet complaints are up. Inbound AI agent metrics show what really happened, from containment rate to handoff reasons and dead air. Here are ten, each with a formula and a healthy range.

1. Containment rate: calls the AI finishes alone

Containment rate is the share of AI-answered calls that end with no human transfer. The formula is calls resolved with no human transfer ÷ total AI-answered calls.

A caller who gives up and hangs up often counts as "contained", so containment alone can mislead. Gartner analysed 432 AI service use cases. Only a quarter showed positive ROI, and 42% could not be measured.

A healthy range is the highest containment you can hold while resolution (metric 2) and repeat contact (metric 9) stay steady. Derive it per call type. Measure a baseline for a few weeks, read a sample of contained transcripts, and raise the target only while those two metrics hold. Status and FAQ calls will sit higher than disputes.

2. Resolution rate: contained and actually solved

Resolution rate is the share of issues that were truly solved. The formula is issues confirmed solved ÷ total issues. Confirm a fix in two ways. First, check that the same caller did not contact you again about the same issue within 7 days. Second, add a short post-call check, such as asking "Did that fix it?" or scoring a "resolved" criterion on the transcript.

A contained call that comes back two days later is contained but not resolved. That gap hides behind a high containment number.

A healthy range is a resolution rate close to your human team's on the same call type. Set the target by sampling contained transcripts and checking whether the problem was really solved.

3. Handoff rate and handoff reasons

Handoff rate is the share of AI-answered calls passed to a person. The formula is transfers ÷ AI-answered calls.

A healthy handoff rate is above zero, because callers need a route to a human. Gartner found 87% of 3,566 customers say a company using GenAI for service must offer a way to reach a human.

Tag every handoff with one reason code:

  • Caller asked for a human
  • Sensitive or disputed issue
  • Missing knowledge
  • Tool failure
  • Low confidence

Fix the top reason first. Missing knowledge and tool failures are yours to remove. Requested handoffs are not.

Set escalation rules in advance. A billing dispute should go straight to a person. Buyers raise this design question on their own, asking what happens to a disputed amount.

A healthy range is a rate that falls by call type as knowledge gaps close. The share of handoffs caused by missing knowledge and tool failure should shrink fastest.

4. Handoff quality: does the human get context?

Handoff quality is the share of handoffs that reach a person with the caller's details and a summary of the call. The formula is handoffs delivered with summary and caller details ÷ all handoffs. Aim for close to 100%.

A caller who has to repeat their name, account and problem will feel the AI wasted their time. Track the repeat-question rate beside it. Listen for the human asking something the caller already said.

Check one more thing: the AI should use a polite closing line before the transfer.

5. Dead air and silence rate

Dead air is the share of calls with a silent gap longer than your threshold. The formula is calls with a silent gap over your threshold ÷ total calls. Also track total silent seconds per call.

A pause of a few seconds feels like a dropped call to many callers, so a slow bot feels broken.

Silence has five usual causes:

  • Speech-to-text delay
  • Slow model replies
  • Slow CRM or tool lookups
  • Telephony latency
  • A carrier that connects but delivers silence

Set your threshold by listening to a sample of calls and noting the gap where callers start saying "hello?". A healthy range is a dead-air rate that is low, flat or falling, with no gap long enough to trigger that. For a lookup, have the agent say "One moment while I check that."

6. Score every AI call

Score your AI agent the way you score a human agent. The formula is weighted criteria points earned ÷ points possible, averaged across all AI calls. Weight the criteria that matter most, such as accuracy, disclosures and resolution, more heavily than politeness.

Sampling hides problems. Manual QA usually covers only 2% to 5% of calls. An AI agent that fails in a rare way can pass a sample and still fail many callers.

A healthy score sits at or above your human team's average and rises week over week. A flat line means nobody acts on the findings.

Insight7 scores every call against a weighted scorecard your team owns, and rolls scores up by person, project and organisation. Managers can review and override the AI's ratings to calibrate them.

7. Critical errors: one miss fails the call

Critical error rate is the share of calls with at least one serious mistake. The formula is calls with at least one critical error ÷ total calls. For compliance and wrong-information errors, the target is zero.

A checklist score can hide what matters. QA leaders say a call can look perfect on procedure and still leave the customer angry. So one critical error should fail the whole call.

Define your critical errors up front:

  • A false promise, such as a refund the policy does not allow
  • A missed disclosure
  • A wrong balance or due date
  • Sharing account data before verifying the caller

Regulated buyers, banks first, fear an AI that "goes rogue" or makes a public mistake. Scanning every call for disclosures and data breaches, with alerts to QA and risk staff, builds the audit trail.

8. Time to answer and call duration

Track three numbers together: answer time, abandonment and handle time. Answer time is the wait before the AI picks up. Abandonment is abandoned calls ÷ offered calls. Average handle time is total talk time ÷ calls.

Healthy looks like this:

  • Answer time is near instant.
  • Abandonment is at or below your human queue's rate. Treat a rise as a pickup or routing fault.
  • Handle time is steady or falling while resolution holds.

Read handle time beside resolution (metric 2).

9. Repeat contact and caller sentiment

Repeat contact rate is contained calls where the same caller returns about the same issue within 7 days ÷ contained calls. Healthy means below your own baseline from the first month and falling. Pair it with sentiment shift, which compares the caller's mood at the start and the end of the call. Healthy means most calls end neutral or better.

Consumer surveys show a rising preference for humans, yet 69% of those who prefer humans would switch to AI that fully resolves their issue, so resolution decides whether they come back.

Check sentiment tags against how calls actually felt. Tags keyed to single words can miss irony and tone. Read a sample of "negative" and "positive" calls, then tune the rules.

10. Cost per resolved call versus human

Cost per resolution is your total AI cost ÷ resolved calls. Total cost is platform fees + telephony + QA cost. Divide by resolved calls, not answered calls, so callers who gave up do not flatter the figure.

Add your three monthly costs, then divide by the calls confirmed resolved under metric 2. Compare the result with your human cost per resolved call. Healthy means lower than the human figure with quality metrics held.

The pressure to show this is real. Gartner found AI spending by service leaders rose 38% while overall service budgets rose 2%. Champions need a case built on their own cost to serve.

Related reading: How to Turn Failed Customer Calls Into Test Cases for Your AI Voice Agent, Containment Rate vs Resolution Rate vs CSAT for AI Agents Explained and Human-in-the-Loop Improvement Workflow for AI Agents With Approval Steps for Script Changes.

Questions people also ask

What is a good containment rate for an AI agent?

It depends on the call type. Simple status and FAQ calls can run high, while disputes and sensitive issues should reach people. Set a target for each call type. Trust it only while resolution and repeat contact hold steady.

What is containment rate in a contact centre?

It is the share of AI-answered calls that end without a human taking over. Tag how each call ended: AI closed it, caller hung up early, or transferred. Only the first counts as contained, so abandoned calls stay out.

How do you calculate containment rate?

Divide calls the AI finished with no transfer by all AI-answered calls. Use the same period and denominator every time. Report it by call type, since one blended figure hides weak spots.

What is the difference between containment rate and resolution rate?

Containment tells you whether a person was involved. Resolution tells you whether the caller's problem was solved. Resolution can never exceed containment on the same calls, so track it as a share of contained calls.

How do you measure dead air in AI voice calls?

Use a transcript with timestamps for each speaker. Measure the gap between the caller finishing and the agent starting. Flag gaps over your threshold. Then check system logs to see whether transcription, the model, a lookup or the carrier caused each one.

How do you track why an AI agent hands calls to humans?

Record a reason code from a fixed list at the moment of transfer. Read a sample of transcripts to check the codes. Rank reasons weekly and fix the top one you control.

How often should you review inbound AI agent calls?

Score every call automatically. Review critical errors the same day. Have a person read a sample of the lowest-scoring calls each week. Recalibrate the scorecard whenever human and AI scores disagree.

The takeaway

Ten metrics only help if they point at the caller's outcome. Insight7 scores every call against a scorecard your team owns, so you see those outcomes. Rule of thumb for your team: count what the caller got, not when the call ended.

Book a demo of Insight7