What's the single most important voice AI metric to track?

There isn't one. Answer rate tells you if calls get picked up, containment tells you how many finish without a human, and order accuracy tells you if the food is right. A great answer rate means nothing if accuracy is poor. Read them together, and weight them toward whatever is costing you the most today.

What counts as a 'good' containment rate?

For straightforward order-taking and FAQ calls, somewhere in the 60–85% range is a reasonable expectation once your menu is trained, meaning most calls finish without a human. But the right number depends entirely on your call mix. A place with heavy catering and complaint volume will and should hand off more. Measure your own baseline before judging it.

How long before the metrics stabilize?

Give it a few weeks of real call volume. Early numbers are noisy because the menu training is still settling and you're catching edge cases. Watch the trend line, not any single day. If a metric is still drifting the wrong way after a month, that's a signal to dig in.

Do these metrics prove ROI on their own?

Not by themselves. After-hours orders captured and conversion get you close, but you still have to tie them back to revenue and compare against your plan cost. The dashboard gives you the inputs. The ROI math is yours to run against your own ticket sizes and margins.

Voice AI Metrics: The KPIs That Tell You If It's Actually Working

Most voice AI dashboards will happily show you a wall of numbers. The harder question is which ones actually tell you whether the thing is earning its keep, and which are just decoration. A high "calls answered" count feels great right up until you learn half those calls ended with a frustrated customer hanging up.

Here are the metrics worth watching, what each one really measures, and roughly what "good" looks like, framed as a range rather than a promise. Your baseline is the only number that matters, so measure your own before you judge any of these.

Answer rate: are calls even getting picked up

This is the floor. Answer rate is the share of inbound calls that get answered at all, versus rung out or dumped to voicemail. It's the whole reason a lot of restaurants look at voice AI in the first place, because the cost of a missed call compounds in ways a single lost ticket never shows.

With a 24/7 agent, this number should sit near the top of its range, because the system doesn't get slammed during the Friday rush the way your staff does. If your answer rate isn't close to complete, something is misconfigured, not just busy. This is the easiest KPI to read and the least interesting once it's healthy, so don't stop here.

Containment: how many calls finish without a human

Containment (sometimes called resolution or automation rate) is the percentage of calls the AI handles start to finish without transferring to a person. It's the metric that most directly reflects how much work the system is taking off your staff.

A reasonable expectation for routine order-taking and FAQ calls is somewhere in the 60–85% range once your menu is trained. But read that number in context. A restaurant with heavy catering, complaints, and special-event calls will hand off more, and should. That's not failure. As we covered in how AI handles the weird calls, a clean handoff with context is a feature, not a defect. Chasing containment toward 100% by forcing the AI to muscle through calls it shouldn't is how you end up with angry customers and a great-looking dashboard.

Order accuracy: is the food actually right

This is the one that touches your kitchen and your refunds. Order accuracy is the share of AI-taken orders that go out correct, without a modification error, a missed substitution, or a wrong item. It's harder to measure cleanly than a call metric because it lives partly in your POS and partly in what walks out the door, but it's worth the effort.

Watch it early and watch it after any menu change, because that's when errors creep in. A useful proxy is your rate of order corrections or remakes tied to phone orders. If accuracy dips after you add a new set of modifiers or a seasonal menu, that's a training gap, not a lost cause, and it's fixable by tightening the menu setup.

Average handle time: fast, but not too fast

Average handle time is how long a typical call takes end to end. Shorter is generally better for throughput, and an AI that doesn't put callers on hold usually beats a slammed staffer on this. But handle time is a metric you read alongside accuracy, never alone.

A suspiciously short handle time can mean the agent is rushing people off the line or dropping calls, not that it's efficient. A slightly longer call that ends in a correct, complete order beats a fast one that generates a remake. Treat handle time as a health check, not a target to minimize at all costs.

Conversion and average ticket: is it selling

Conversion is the share of order-intent calls that end in a placed order. Average ticket is what those orders are worth. Together they tell you whether the agent is actually turning calls into revenue, or just politely answering questions.

A voice agent that consistently offers a documented add-on or reads the full modifier list can nudge average ticket, but be honest with yourself about attribution. Don't credit the AI for tickets that would have been the same on a human call. The cleaner signal is conversion on calls that used to go unanswered entirely, because those were revenue you were flat-out losing before.

After-hours orders captured: the revenue you couldn't get before

This is the KPI that most directly justifies a 24/7 system, and it's the one hardest to argue with. After-hours orders captured is the count and value of orders taken when your phone would otherwise have been dark, on nights, holidays, and the pre-open and post-close windows.

Unlike conversion during staffed hours, there's little attribution debate here. Nobody was answering that phone at 11pm. Every completed order in that window is incremental. Track both the volume and the dollar value, and this is usually the first number that makes the pricing math pencil out for an operator on the fence.

Reading the dashboard honestly

A few habits keep these numbers useful instead of flattering:

Watch trends, not single days. One bad Saturday tells you nothing. A metric drifting the wrong way over three weeks tells you plenty.
Never read a metric in isolation. Containment without accuracy, or handle time without conversion, will mislead you. The story lives in how they move together.
Establish your own baseline first. Every range above is a general approximation. Your restaurant's mix of order types, menu complexity, and call volume sets the real bar.
Segment by daypart. Peak-hour calls and 2am calls behave differently and hide each other in a blended average.

If a number looks too good, check what it's costing you somewhere else on the board. If one looks bad, check whether it's the system or your menu training before you conclude anything. You can always talk through your own numbers with someone who's seen a lot of these dashboards.

The bottom line

No single metric proves voice AI is working, and any vendor who points at one is selling you the flattering half of the picture. Answer rate proves calls get picked up, containment proves work is coming off your staff, accuracy proves the food is right, and after-hours capture proves it's making money you couldn't make before. Read them together, trust your own baseline over any published range, and let the trend line, not a good night, tell you the truth.