Most call-center dashboards are good at counting activity. Dials, connected calls, talk time, calls per hour. All useful, and all silent on the one question a sales director really asks: did that conversation make us money without costing us goodwill? Four numbers answer it far better than a wall of activity charts: conversion, basket value, hang-ups under 10 seconds and do-not-call requests. Each one is easy to measure badly, so it is worth being precise about what goes into it.
Conversion: decide what you are dividing by
"Conversion rate" can mean three different things in the same meeting. One manager divides orders by dials. Another divides by connected calls. A third divides by conversations that lasted more than a minute. All three get a different answer, and all three can be defended.
Pick one denominator and write it down. For most outbound teams the cleanest choice is orders per human-answered call. Dials punish you for bad data and voicemail. Long conversations only is a filter that hides your weakest openings, because those calls end early.
Imagine 1,000 calls answered by a real person and 80 orders. That is 8 percent. It is a number you can improve, compare by campaign, by hour of the day and by script version. Keep a second, softer figure beside it: the share of calls that ended with a clear next step, such as a callback booked or an SMS with a payment link sent. In considered purchases, those calls are often tomorrow's orders.
Basket value: the number that catches cheap wins
Conversion alone can be gamed, even by accident. A team that offers a discount in the first thirty seconds will close more calls and earn less per call. That is why conversion should never travel without average basket value.
For e-commerce phone sales, track at least two things: the average order value on calls, and the share of orders that include an extra item beyond the one the customer first asked about. That second figure is your upsell and cross-sell rate, and it shows whether the agent, human or AI, is actually listening for needs rather than reading an offer.
Here is a simple example. A customer calls back about a delayed parcel of running shoes. A good agent solves the delay, then asks whether the customer needs socks or insoles for the new pair. If that works one time in ten, it will show up in basket value long before it shows up in anyone's opinion of the script.
Compare like with like. Orders from a reorder campaign and orders from cold leads have different natural baskets. Segment first, then judge.
Hang-ups under 10 seconds: a verdict on your opening
A call that ends in under ten seconds is the shortest feedback a customer can give. Nobody heard your offer. They heard your first sentence, or the pause before it, and decided.
Track this as a share of human-answered calls, and break it down by queue, caller ID, time of day and script version. Patterns appear quickly. A spike in the evening may mean you are calling at the wrong time. A spike on one caller ID may mean people have learned to reject it. A spike after a script change tells you the new greeting is worse than the old one.
One trap deserves a warning: make sure you are not measuring machines. If voicemail greetings and dead lines are counted as short human hang-ups, the metric turns to noise. Answering machine detection that decides from live audio, with a time limit you can tune per queue, keeps the figure honest. Alphacall, for example, classifies each answer as human, voicemail, unavailable or invalid number, so the short-call rate describes people only. Too slow a decision creates its own problem, though: a customer who says "hello" three times into silence will hang up and be counted against your opening.
When the rate is high, listen to twenty recordings of calls under ten seconds before changing anything. Often the cause is plain: the agent takes too long to say who is calling and why, or the first line sounds like a survey.
Do-not-call requests: treat them as a signal, not an annoyance
Some teams hide this number because it feels like failure. That is a mistake. A do-not-call request is the customer telling you precisely what they want, and honoring it quickly protects your reputation and your future lists.
Measure two things. First, the rate of requests per human-answered call, by campaign and by data source. A list that produces many requests is a list with a problem: old consent, a bad purchase source, or a mismatch between who you call and what you offer. Second, the time between the request and the number actually being suppressed. The right answer is "immediately, automatically". If it depends on someone exporting a spreadsheet on Friday, a customer can get called again on Thursday.
Rules on calling hours, consent and suppression lists differ between countries, so check what applies to your markets with your own advisers. What you can do operationally is make the process boring and reliable: one suppression list shared by all queues, the request recorded the moment it is spoken, and calling windows enforced by the system, not by memory.
Reading the four numbers together
Each metric alone can mislead. Together they tell a story.
- Conversion up, basket value down: you are probably discounting too early or skipping needs discovery.
- Conversion flat, short hang-ups up: the opening is losing people, and the agents who stay are doing more work to compensate.
- Conversion up, do-not-call requests up: pressure is working today and costing you the list tomorrow.
- Basket value up, conversion down: the upsell may be arriving before the customer has agreed to the main purchase.
A short weekly routine keeps this practical:
- Fix the definitions: denominator for conversion, what counts as a short call, what counts as a request.
- Exclude voicemail, unavailable and invalid numbers from all four rates.
- Split every figure by queue, source, hour and script version.
- Listen to a small sample of recordings from each extreme: best baskets, shortest calls, every do-not-call request.
- Change one thing at a time, then wait for enough calls before judging it.
The same discipline applies whether the caller is a person or an AI voice agent. If anything, it matters more with an agent, because you can change a script on Monday and see the effect in the numbers by Wednesday. Real-time supervision helps here: listening in on live calls lets a supervisor catch a weak greeting the same hour, not the following month. The aim of a human plus AI setup is not to remove people from the loop but to give them cleaner data and handovers with full context, so their time goes to the conversations that need judgment.
See the numbers on a real call
Metrics are easier to trust once you have heard the calls behind them. If you would like to see how an AI voice agent opens a conversation, handles an objection, offers a second item and hands over to a colleague with context, you can watch it happen. Book a demo and see Alphacall on a live call, then bring your own four numbers and compare.