AI voice agent metrics: measure call AI like a sales floor, not a chatbot

Chatbot dashboards report sessions, messages and "resolution". None of that tells you whether a phone agent is earning its place. Judge it the way you'd judge a new person on the phones: did they reach people, have useful conversations, book the right work, and hand over what they couldn't handle? This article defines those numbers, shows how to calculate them and gives you a dashboard to copy.

It's part of our guide to AI voice agents in Australia. The same discipline applies whether your agent answers inbound calls or rings enquiries back.

The call funnel, defined precisely

Write the definitions down before you start, and don't change them mid-month. Most arguments about whether voice AI "works" are really arguments about definitions.

For inbound, connect rate becomes "answered rate" and should be close to 100%; the interesting numbers are greeting drop-off, booked rate and escalations. For outbound call-backs, connect rate is the one that moves most with timing, which is why speed to lead matters.

SMS metrics belong on the same page

Calls and texts work as one system, so measure them together. A low connect rate with a high SMS reply rate tells you people prefer to respond in writing, and that changes your cadence.

• SMS reply rate: replies ÷ texts sent, by template.
• Reminder confirmation rate: confirmed ÷ reminders sent.
• No-show rate: before and after introducing reminders.
• Opt-out rate: opt-outs per hundred commercial texts. Rising opt-outs are an early warning on frequency or relevance.

We cover cadences and the Spam Act rules behind them in SMS follow-up after a call.

Escalation rate: the number that tells you about scope

Track escalations by reason (caller asked for a person, urgent job, complaint, agent failed to understand, out-of-scope request). Each reason means something different:

• Many "asked for a person" escalations in the first seconds: the opening is off-putting or callers expected a human. Shorten the greeting; check the voice.
• Many "didn't understand" escalations: audio quality, accents or vocabulary. Read those transcripts first.
• Many out-of-scope escalations: the agent is being given calls it wasn't designed for. Either widen its scope deliberately or route those calls elsewhere.
• Almost no escalations at all: suspicious. Check transcripts for the agent bluffing through calls it should have handed over.

There is no universal "good" escalation rate; it depends on your call mix. What matters is the trend and the reasons.

QA sampling: read calls every week

Numbers tell you where to look; transcripts tell you what happened. Every week, have a person score a random sample of calls plus every call with a complaint, an escalation for "didn't understand" or an opt-out.

Start with 20 calls a week, or all of them if you have fewer. Any compliance score of 0 is a stop-and-fix issue, not a trend to watch; the rules behind that row are in ACMA rules for AI outbound calls.

A dashboard template you can copy

One screen, weekly, compared with the previous four weeks. The numbers below are illustrative, made up to show the layout, not benchmarks.

Add one line for cost: credits or minutes used this week. Then sanity-check value with your own figures: bookings made by the agent multiplied by your average job value and win rate. Our guide to measuring hours saved by AI agents explains how to count the time side without fooling yourself, including the review time you spend on the agent.

Take a baseline before the agent goes live

The most common measurement mistake is having nothing to compare against. Before switching anything on, record four weeks of the current state from your phone system and CRM: calls missed, voicemails left, time from web enquiry to first contact, bookings per week and no-show rate. If you can't get a number automatically, sample it by hand for one week; a rough baseline beats none.

Then change one thing at a time. If you launch after-hours answering, reminder texts and outbound call-backs in the same week and bookings rise, you won't know which one did it, or whether one of them is quietly hurting you. Stagger launches by a fortnight and annotate the dashboard on the day each change goes live.

Finally, be honest about seasonality. Trades are busier before Christmas, clinics quieter in January. Compare against the same weeks last year where you can, and treat a single good week as a hint, not a result.

Metrics we'd ignore, or treat with suspicion

• Average call length on its own. Longer can mean engaged or confused. Pair it with outcome.
• Sentiment scores. Automated sentiment on short phone calls is noisy; your QA reviewer's judgement is better.
• "Containment" or "resolution" rate borrowed from chatbots. A call the AI "contained" by never offering a person is a failure, not a win.
• Calls per hour. Throughput is the AI's easiest number to improve and the least connected to revenue.

Where the numbers come from

None of this works unless every call writes a structured outcome somewhere: who, when, connected or not, outcome, next action, escalation reason. With Fulcrum Lyre, call outcomes are logged to the Fulcrum CRM and recapped in your channel, so the dashboard is a query rather than a spreadsheet you rebuild every Friday. Whatever tools you use, insist on that structure before you go live; retrofitting it later means a month of unusable data.

Review weekly for the first two months, then fortnightly. Change one thing at a time (the opening, the reminder timing, the qualification questions) so you can tell what moved the numbers.