The most useful AI voice agent KPIs are task completion, containment, transfer rate, handoff quality, first-call resolution, abandonment and failure, response latency, interruption recovery, capture accuracy, and cost per resolved call.
Together they show whether calls are resolved, not just answered. This guide gives a formula and a measurement method for each, so you can judge a production voice agent, compare vendors, or build a call analytics dashboard.
Why one metric is never enough
Containment alone can look healthy while callers give up. Macha puts it this way: containment counts sessions that did not escalate, which is not the same as sessions that succeeded.
Read the ten KPIs in four groups: outcomes (1, 2, 5, 10), escalation (3, 4), call health (6, 7, 8), and accuracy (9).
| # | KPI | Core question |
|---|---|---|
| 1 | Task completion rate | Did the caller accomplish the goal? |
| 2 | Containment (automation) rate | Did automation carry the call without a human? |
| 3 | Transfer rate | How often was a human needed? |
| 4 | Successful handoff rate | When transferred, did the handoff work? |
| 5 | First-call resolution | Was the issue resolved on first contact? |
| 6 | Abandonment and failure rate | Did callers give up, or did the call break? |
| 7 | Response latency | Did the agent answer fast enough? |
| 8 | Interruption recovery rate | Could it handle barge-ins and corrections? |
| 9 | Intent and capture accuracy | Did it understand and record the right details? |
| 10 | Cost per resolved call | Did it create measurable value? |
The 10 AI voice agent KPIs
1. Task completion rate
Formula: Completed tasks / calls with a task intent × 100
This is the clearest success signal. Count a task only when the system of record confirms it, such as a booking in the calendar or a created ticket.
Segment by intent so a weak workflow, such as rescheduling, is not hidden by strong ones. LuMay's Google Calendar and Salesforce integration pages describe checking availability, writing bookings, and updating records from call outcomes, which gives you system events to count.
2. Containment (automation) rate
Formula: Interactions resolved without a human / interactions entering automation × 100
It shows how much demand automation carries alone. Treat it as an efficiency metric, not a quality metric, and do not count a call as contained if the caller was stuck, misunderstood, or left without finishing the task.
Add a repeat-contact check. Parloa treats a repeat contact within seven days about the same issue as a sign that a contained interaction did not resolve anything.
3. Transfer rate
Formula: Calls transferred to a human / AI-handled calls × 100
Split transfers into planned (payments, policy, VIP, complaints) and unplanned (confusion, failed verification, frustration). Unplanned transfers are your improvement backlog.
Contact-center tools treat this as its own outcome. Google's Dialogflow CX analytics reports live agent handoff as a conversation metric and ranks intents by human escalations. Transfer and containment will not always sum to 100%, because abandoned and failed calls sit outside both.
A rising transfer rate is an early warning. Read the transcripts behind it before you change prompts or routing, and never optimize it in isolation.
4. Successful handoff rate
Formula: Handoffs where a person answered with full context / total handoffs × 100
A transfer is not a successful handoff. It succeeds only if the caller reaches the right person, does not have to repeat the issue, and the summary arrives with the call. Check what the agent passed along:
- caller identity or verification status;
- reason for calling and reason for escalation;
- details already collected and actions already attempted;
- the CRM or ticket reference.
LuMay's Voice Agent features page describes handoffs with a live conversation summary, so the human arrives informed. Audit recorded samples to confirm it works.
5. First-call resolution (FCR)
Formula: Calls resolved on first contact / total calls × 100
Call Centre Helper gives FCR as total contacts resolved on first contact divided by total contacts. Define “resolved” in writing before launch. HiveDesk offers one typical definition: the customer does not contact the company again about the same issue within seven days, and each team should set its own window.
FCR catches what containment hides: calls that end politely but solve nothing. Join call outcomes with downstream CRM, ticket, and repeat-contact data.
6. Call abandonment and failure rate
Formulas: Abandonment = (calls offered - calls handled) / calls offered × 100. Failure = calls ending in a technical error / total calls × 100.
Keep the two apart so you do not blame conversation design for an infrastructure problem. Technical causes include telephony faults, dropped media, API errors, and failed transfers.
Exclude very short abandons. Call Centre Helper says it is common practice to exclude calls abandoned in the first five seconds, since many are misdials. Amazon Connect defines abandonment as contacts disconnected by the customer while in queue, divided by contacts queued.
For failures, watch call quality. Twilio flags carrier and SIP calls with 5% cumulative packet loss, average jitter above 5 ms, or internal RTP traversal above 150 ms. LuMay's Twilio integration page covers phone and messaging workflows on that infrastructure.
7. Response latency
Formula: Time from the caller's last word to the first agent audio, reported as median, p95, and p99
Averages hide the slow calls callers remember. Track percentiles by intent and time of day, and report sample counts, because a p99 from 50 calls is one or two slow calls. LuMay's testing guide covers how to measure it.
For context, ITU-T G.114 says that keeping one-way delay below 150 ms gives most applications essentially transparent interactivity, and recommends not exceeding 400 ms for general network planning. That covers transmission only, so measure model response time separately.
Break the gap into recognition, model, speech synthesis, and network stages, as LuMay's latency page does. That page publishes no fixed benchmarks, so set targets from your own caller testing.
8. Interruption recovery rate
Formula: Interruptions handled correctly / total caller interruptions × 100
Handled correctly means the agent stops promptly, hears the new input, keeps its place in the workflow, and answers instead of restarting its script. Also count false interruptions caused by noise, echo, or short acknowledgements.
Platforms expose the controls. Twilio ConversationRelay lets you set whether speech can interrupt playback, tune interrupt sensitivity, and ignore short fillers such as “okay.” Review recordings flagged for overlapping speech, since these failures hide inside aggregate completion numbers.
9. Intent and information-capture accuracy
Formulas: Intent accuracy = correctly classified calls / audited calls × 100. Field accuracy = correct captured fields / audited fields × 100.
An agent can sound fluent while recording the wrong detail. Audit a random weekly sample against human-reviewed transcripts or confirmed downstream records. Score names, dates, phone numbers, and account identifiers separately, because small errors break downstream records.
For transcription quality, word error rate is (substitutions + deletions + insertions) / words in the reference transcript. Pay extra attention to high-risk fields such as identity, payment, and compliance answers.
10. Cost per resolved call
Formula: (Platform + telephony + integration + review effort + human time on transfers) / resolved calls
Divide by resolved calls, not handled calls, so failed automation is not counted as savings. Compare the result with your human cost per resolved call over the same period, and avoid claiming savings unless the baseline, volume, and outcome definitions match. Add workflow outcomes, such as bookings or qualified leads, to show revenue impact.
A worked example (hypothetical numbers)
Assume 1,000 calls enter automation in a month. These figures are illustrative, not LuMay results.
| KPI | Inputs | Result |
|---|---|---|
| Task completion | 680 confirmed of 800 calls with a task intent | 85% |
| Transfer rate | 150 of 1,000 | 15% |
| Successful handoff | 120 of 150 transfers answered with context | 80% |
| Abandonment and failure | 50 of 1,000 | 5% |
| Containment | 1,000 - 150 transferred - 50 abandoned or failed = 800 | 80% |
| First-call resolution | 720 with no repeat contact in the chosen window | 72% |
| Cost per resolved call | Assumed $1,500 total cost / 720 resolved | About $2.08 |
The gap between 80% containment and 72% FCR is the point: 80 calls ended without a human but came back or went unresolved. Containment alone would have hidden them.
Where to start in the first month
Do not try to tune all ten KPIs at once. A practical order:
- Before launch: baseline your human-only performance and define “resolved” in writing.
- Weeks 1 to 2: watch the foundations first, which are intent and capture accuracy, response latency, and technical failures. If these are weak, every other number is unreliable.
- Weeks 3 to 4: add task completion, containment, transfer and handoff quality, and the repeat-contact check behind FCR.
- Ongoing: report cost per resolved call and workflow outcomes once the definitions are stable, and review them weekly.
This is a suggested sequence, not a standard. Adjust it to your call mix and risk level.
How to build your measurement stack
Most KPIs need joined data from four sources:
- Telephony: call state, duration, and quality, such as Twilio Voice Insights.
- Conversation: transcripts, intents, and outcomes.
- Systems of record: bookings, tickets, and CRM updates.
- Human QA: a weekly audited sample.
LuMay's features page lists a real-time dashboard covering volume, resolution rates, escalation triggers, common intents, call duration, and sentiment trends, plus automatic transcripts and structured summaries. Handoff quality and capture accuracy depend on QA sampling on any platform.
To feed your own BI tools, LuMay's integrations docs describe a REST API connector and a post-call webhook target, though they list no endpoints or payloads, so request the specification. AI integration services covers connecting agents to existing systems. LuMay's security and governance page describes audit trails and tenant-configurable PII masking.
Launch checklist
- Define every formula and time window before go-live.
- Baseline human performance first.
- Segment by intent, language, and time of day.
- Review p95 latency weekly.
- Alert on repeat contacts.
Explore the LuMay Voice Agent and its inbound call handling, or book a demo to map these KPIs to your calls.
