A voice programme is working when outcome metrics move: collections recovered, renewals saved, leads qualified. Track a short stack of operational metrics behind them: contact rate, containment or automation rate, task completion, average handle time and CSAT. Watch outcomes, not call volume. Vanity metrics like calls dialled tell you the system is busy, not that it is working.
Why do most voicebot dashboards measure the wrong thing?
Most voice dashboards are full of numbers that feel like progress and prove nothing. Calls dialled. Minutes of talk time. Bot uptime. All of them go up when the system is busy, whether or not it is working.
The question a voice programme has to answer is narrow: did the outcome improve, at a lower cost, without hurting the customer? Every number worth keeping ladders up to that one.
A busy contact centre and a working one are not the same thing.
Vanity metric
- Calls dialled
- Minutes of talk time
- Containment on its own
- Bot uptime
Outcome metric
- Recovery against baseline
- Renewals and policies saved
- Task completion rate
- Cost per outcome
The metrics that tell you a voice programme is working
Five operational metrics carry most of the signal. Define each one carefully, because the names get used loosely.
- Contact or connect rate: the share of attempted calls that reach a live person. A low connect rate points at your dialling strategy or your data, not the bot.
- Containment or automation rate: the share of conversations the AI finishes on its own without a human. This is the number that drives cost. A mature full-stack deployment handles 60 to 80% of interactions autonomously.
- Task completion rate: of the calls the bot handled, how many reached the actual goal, such as a payment promised, a renewal booked, or a query resolved? High containment means nothing if the task did not complete.
- Average handle time: how long a conversation takes. Useful as an efficiency check, dangerous as a target. Cut it by rushing customers and you wreck completion and satisfaction.
- Customer satisfaction (CSAT): measured through post-call surveys or sentiment scoring. The test is simple: does the AI match or beat your human agents on the same process?
What about the metrics that actually pay the bills?
The operational metrics tell you the machine runs. Outcome metrics tell you it earns its keep. These are the ones to put in front of the board.
- Collections: recovery rate and value recovered against baseline, not calls made. A bot that dials all day and recovers nothing is a cost, not a solution.
- Retention and renewals: policies or contracts saved, and churn avoided, measured against what you lost before.
- Lead qualification: qualified leads passed to sales and how they convert downstream. Where AI runs full qualification flows, we see roughly 3x the human benchmark conversion.
- Cost per outcome: the metric that ties it all together. In India, the hybrid model runs at around 30% lower cost per outcome.
What does good look like?
Good is not an absolute number you copy from a vendor deck. It is relative to two things: your human baseline and your business goal.
For automation, a mature deployment contains 60 to 80% of interactions without a human. For task completion and CSAT, good means the AI matches or beats the agents it replaced on the same process. For outcomes, good means the recovery, save or qualification rate moves against the baseline you measured before you started, and the cost per outcome falls.
Measure the baseline first. A programme with no before number cannot prove an after.
Where do teams go wrong with voicebot metrics?
The common trap is celebrating containment while task completion quietly sits low. The bot is handling calls; it just is not finishing the job. Connect rate looks healthy; the conversations go nowhere.
If a number on your dashboard does not ladder up to the outcome you are paid to move, it is decoration. Start with that outcome, then work backwards to the handful of metrics that predict it.
Vanity metric vs outcome metric
| Vanity metric | Outcome metric |
|---|---|
| Calls dialled | Recovery rate against baseline |
| Minutes of talk time | Renewals and policies saved |
| Containment rate on its own | Task completion rate |
| Average handle time | Cost per outcome |
| Bot uptime | Qualified leads that convert |
Frequently asked questions
What is containment rate in a voicebot?
Containment rate, also called automation rate, is the share of conversations the AI handles end to end without passing the customer to a human. It is the main driver of cost savings, and a mature full-stack deployment contains 60 to 80% of interactions. On its own, though, it can mislead: high containment with low task completion means the bot is ending calls without finishing the job.
What is the difference between containment and task completion?
Containment measures whether the AI handled the call without a human. Task completion measures whether the call achieved its goal, like a payment promised or a renewal booked. A bot can contain a call and still fail the task by giving up, misunderstanding, or ending early. Track both together; containment without completion is a false positive.
What is a good automation rate for an enterprise voicebot?
A mature full-stack voice deployment automates roughly 60 to 80% of interactions, with humans handling the 20 to 40% that need real judgement. The exact figure depends on the use case: simple reminders automate higher, complex disputes lower. Chasing 100% automation usually backfires, because the hard cases are exactly the ones that need a person.
Which metric matters most for a collections voicebot?
Recovery: the value actually collected against your pre-programme baseline, and the cost per rupee recovered. Calls dialled and minutes spoken are vanity numbers; a collections bot earns its place only if recovery goes up or cost per outcome goes down. Track promise-to-pay and kept-promise rates as leading indicators of that recovery.
How do you measure voicebot CSAT?
Use short post-call surveys, or sentiment scoring on the conversation itself, on the same scale you use for human agents. The comparison is the point: does the AI match or beat your agents on the same process? Watch CSAT alongside average handle time, because pushing calls to end faster can quietly drag satisfaction down.