AI Voice Agent KPIs: 8 Numbers That Show Whether It’s Working

Your dashboard says the agent handled most calls. But a customer says the line disconnected, a transfer never reached your team, or an appointment appeared without the right details. Which version should you trust?

Trust the one you can verify against what actually happened on the call. A “handled” count only shows that the agent picked up and responded. It can’t tell you whether the line stayed connected, whether the transfer reached a person, or whether the booking was right. One headline number will always hide those gaps, so you need to measure outcomes, not attempts. These 8 AI voice agent KPIs do that: answer rate, handled-without-staff rate, mid-call hang-up rate, transfers that reached a person, bookings correct and kept, repeat calls within 7 days, cost per resolved call, and listened-call pass rate. Together, they tell you what happened, not merely what the agent attempted.

PRO TIP

Do not start by asking whether a KPI is “good.” First ask what was counted, what was excluded, and which independent system can confirm the result.

8 AI Agent Performance Metrics

Repeat calls within 7 days is the AI-agent version of first-call resolution (FCR), the measure human teams have used for years. Once you have a few hundred surveyed calls a month, add customer satisfaction (CSAT) on handled calls and compare it with transferred calls.

If the agent answers inbound calls

In the first month, check whether the agent is set up correctly before you judge how well it performs. Most early problems are setup problems, not performance problems, and these three KPIs expose them fastest. Answer rate shows whether calls are reaching the agent at all. Handled-without-staff rate shows whether it has the knowledge and call flows to finish what it starts. Correct, kept bookings show whether it is writing the right details into your calendar. They are also the numbers you bought the agent to move, so if one is off, you know exactly where to fix it.

If the agent makes outbound calls

Swap four rows. Answer rate becomes connect rate, which is calls a person picked up ÷ calls dialed. Handled without staff becomes outcome per 100 connected calls, whether that is bookings, qualified leads, or confirmed renewals. Bookings correct and kept stays as it is.

Then add two compliance numbers. Consent coverage is dialed numbers with recorded prior express consent ÷ numbers dialed. Opt-outs honored on time is opt-outs actioned within 10 business days ÷ opt-outs received.

Count Only Real Conversations, or Every Rate Looks Better Than It Is

Two KPIs in the table are divided by real conversations: automated resolution rate and abandonment rate. Repeat contact rate and cost per resolved call build on them. So how you define a real conversation affects half the scorecard.

A real conversation is an answered call where the caller spoke:

Real Conversations = Answered Calls − Spam − Robocalls − Silent Calls

These calls tell you nothing about the agent, and leaving them in makes your rates inaccurate. In an analysis of 450,702 calls to 503 AI receptionists, Upfirst found that raw hang-up rates reflect a phone number’s spam traffic more than the voice agent. That is vendor data, but it matches what most operators see in their own logs.

Every real conversation then needs one outcome tag.

• Booked
• Question answered
• Message taken because the caller asked to leave one
• Transferred
• Hung up with no outcome

A call without a tag doesn’t count as handled.

Check Transfers and Bookings at the Far End, Not in the Agent’s Log

A transfer succeeds only when your phone system shows a staff member answered. “Transfer started” or “transfer connected” in the agent’s log may only mean the platform dialed the number, even if the call then rang out or went to voicemail.

Once a week, match the agent’s transfer list against your phone system’s answered calls. Then sort the transfers by reason:

Planned: Calls the agent is set up to hand over, such as emergencies, billing requests, or callers who ask for a person.

Forced: Calls the agent couldn’t handle.

Planned transfers are working as intended, as long as someone picks up. Forced transfers show where the agent falls short. If one type of request keeps ending in forced transfers, either improve how the agent handles it or send those calls straight to your team.

Use this transfer script as the test standard:

Transfer script:

For bookings, check both accuracy and the eventual outcome. Match each reported booking to the calendar: Correct person, service, date, time, time zone, and contact details. Then calculate the show rate and compare it with appointments booked by staff. A booking that exists but has the wrong service or never triggers a reminder is not a clean success.

Use explicit confirmation rather than a vague “you’re booked”:

Booking script:

A short script makes pass/fail review easier because the required fields are audible in the call and visible in the transcript.

What A Good Number Looks Like For Your Business

Ranges as published by each vendor, October 2026. Response delay is the gap between the caller finishing a sentence and the agent starting to reply. The “slowest 5%” figure is the 95th percentile (P95).

What Each Resolved Call Costs You

Cost per minute hides calls that cost money and fix nothing, such as spam, hang-ups, and calls that end in a transfer. Cost per resolved call shows what you pay for each problem the agent solves. Compare it with what the same types of calls cost when staff handled them. 

Cost per resolved call = total agent cost ÷ (calls handled without staff − repeat contacts)

Total agent cost includes every minute (including spam, hang-ups, and transferred calls), plan and phone number fees, and the staff time spent reviewing calls.

For Example:

Reusing the funnel, 500 answered calls averaging 2 billable minutes is 1,000min. At Botphonic’s published Pro PAYG rate of $0.12 per talk minute, as of October 2026, that is $120.

The cost per resolved call is about 2.3 times the per-call figure. This example leaves out staff review time, so add yours for the full figure.

Cost per resolved call also treats every resolved call the same, but a booking that becomes a paid job is worth more than an answered question. See Botphonic’s success stories for how the same calls look once you count jobs won.

Scripts To Fix The Most Common Voice Agent Failures

Start with these scripts when the review points to a specific failure:

Short Inbound Greeting:

Recovery Script:

Human-Request Script:

For outbound calls, have counsel approve the opening and opt-out language. A plain operational template is:

Outbound Script:

Do not treat that wording as proof of TCPA compliance; consent, identification, disclosure, and opt-out requirements depend on the campaign and jurisdiction.

Your First 30 Days: A Week-By-Week Measurement Plan

If you’re just setting up, Botphonic documentation covers call handling, spam screening, and syncing bookings to your other tools.

What Belongs On Your AI Voice Agent KPI Dashboard (Scorecard)

Conclusion

An AI voice agent dashboard is only as honest as its denominator and its far-end checks. Count real conversations, confirm transfers in the phone system, verify bookings in the calendar, connect containment with repeat calls, and divide total cost by resolved outcomes, not activity.

Then segment the results before diagnosing a problem and change only the layer that failed: Script, workflow, handoff, or scope. Review a small, balanced call sample every week and compare rolling 4-week totals so that normal low-volume variation does not trigger unnecessary rewrites.

These practices give you 8 numbers you can trust and one practical action each week. Start by filling the scorecard with last month’s call log.