Passer au contenu
Opérations 7 novembre 2026 10 lecture minimale

How to Measure AI Voice Agent Containment

Containment rate formula for AI voice agents, the right denominator, how hang-ups and repeat calls fake success, and the metrics to report beside it weekly.

D
7 novembre 2026

Réponse rapide

Containment rate is the share of calls an AI voice agent handles from start to finish without passing them to a person. Divide contained calls by the calls that reached a real conversation, leaving voicemail out. Then subtract false containment: hang-ups, wrong answers and callers who ring back within a fixed window. Report it beside transfer rate, abandons, repeat contacts and satisfaction, and listen to a weekly sample of contained calls.

Containment is the number every AI voice agent dashboard wants to show, and the easiest number to flatter. A rate can climb because the agent is better. It can also climb because callers give up. The formula looks the same either way, so the method around it matters more than the formula.

This guide sets out a definition you can defend, the denominator choices, the three ways containment is faked by accident, and a weekly routine that keeps the number honest. It makes no benchmark claim, because the right target depends on which calls you give the agent. The numbers in the worked example are illustrative only.

What containment means, and the formula

A call is contained when the AI voice agent handled it from the first greeting to the end without transferring it to a human. The rate is:

containment rate = contained calls / calls in the denominator x 100

That is the whole formula. Everything difficult sits in two places: what counts as contained, and what goes in the denominator. A plain definition written down once, and kept, is worth more than a clever one.

Write the definition in a sentence your whole team can repeat: "A call is contained when the agent reached a recorded outcome and no transfer to a person was requested or made." Note that this says nothing about whether the caller was happy. That is deliberate. Containment is a routing measure. Whether the call went well is a different question, covered below.

Choosing the denominator

The denominator decides what the rate means. Three common choices give three different numbers from the same week of calls.

DenominatorWhat it includesBest used for
All calls offered or dialledIncludes no-answer, busy, voicemail and abandons before the greetingCapacity planning, not agent quality
Répondre aux appelsEvery call someone or something picked up, including a caller who hangs up during the greetingInbound lines, where callers choose to ring
Connected conversationsOnly calls where a person spoke with the agent, with voicemail and silent pick-ups removedOutbound campaigns, and the cleanest quality read for inbound

Leave voicemail out of outbound

On an outbound campaign, a call that lands on voicemail has nobody to contain. If it sits in the denominator, a list with a poor connect rate looks like a poorly performing agent. Keep voicemail, no-answer and busy as their own outcomes, and report them next to the rate. The connected-conversation denominator then answers one question: when a person did talk to the agent, how often did the agent finish the job?

Decide the edge cases once

  • A caller who hangs up in the first seconds. Decide a cut-off, and either exclude those calls or count them as their own bucket. Do not leave them to fall into "contained" by default.
  • A caller who asks for a person and the agent declines. Count this as a failure of the call, even though no transfer happened. See the next section.
  • A call transferred by design. Some call reasons are meant to go to a person, such as a complaint. Do not count a planned transfer against the agent. Segment by call reason instead.

False containment: three ways the number lies

False containment is a call that counts as contained and was not a success. It has three usual sources.

1. Hang-ups

A caller who loses patience and hangs up produces a call with no transfer. Unless you tag it, it looks like a clean finish. The fix is to require a recorded outcome for a call to count as contained. If the agent never reached an outcome such as a booking made, an answer given or a message taken, the call is an abandon, not a containment.

2. Repeat calls

A caller whose problem was not solved rings again. The second call may be contained too, and now one problem has produced two "successes". Choose a fixed window, and when the same number calls again about the same reason inside it, reclassify the first call as not resolved. The window is your own policy. A booking question might use 24 hours, a billing question several days. State it beside the metric and keep it fixed so weeks stay comparable.

3. Wrong answers

An agent that answers quickly and wrongly is contained. A quoted price that is not in the catalogue, a booking for the wrong day, an invented policy: all of them close the call without a transfer. This is the failure the DialerBee agent vocal IA is built to limit, with answers drawn from your uploaded documents and only catalogue prices spoken. Even so, you find wrong answers by listening, not by counting. The weekly sample below is how.

Containment versus verified resolution

Containment tells you the call was not routed to a person. Verified resolution tells you the problem was solved. Keep them as two separate numbers.

A practical way to verify resolution without surveying every caller:

  1. Start from contained calls with a recorded outcome.
  2. Remove any caller who rang again about the same reason inside your window.
  3. Remove calls flagged as wrong, in the weekly listening sample, and apply the sample's error rate to the rest as an estimate.

The result is a conservative resolution estimate. It is the number to use when you compare cost: see AI voice agent pricing and cost per resolution, where the denominator is resolved issues rather than minutes. It also links to the older idea of résolution du premier appel, which asks the same question of human agents. For ways to raise it, see how to improve first call resolution.

Segment by call reason

One blended rate hides everything useful. A line that handles opening hours, bookings and complaints will show a middling number made of one very high rate, one medium rate and one rate that should be near zero by design.

Tag every call with a reason, and report containment per reason. Three groups usually emerge:

  • Reasons the agent should own. Hours, booking, reschedule, status questions. A low rate here is a defect to fix.
  • Reasons that are mixed. Questions whose answer depends on the account. A moderate rate is expected, and the transfer path matters more than the rate.
  • Reasons that should go to a person. Complaints, distress, legal threats. A high rate here is a warning, not a win. Make sure the human handoff works for these calls.

Report it beside the metrics that expose it

Containment on its own invites the wrong behaviour. Put these next to it on one weekly page.

MétriqueWhat it catches
Transfer rateThe mirror of containment. If it falls while abandons rise, callers are not getting through to a person.
Taux d'abandonCallers who hang up before an outcome. The main source of false containment.
Repeat contactsThe same number calling again about the same reason inside your window.
Caller satisfaction (CSAT)Whether contained calls felt resolved to the caller. Use a short post-call question only where you have the caller's consent to ask.
Handling timeA very short average can mean quick answers, or quick hang-ups.

For the standard formulas behind transfer, abandon and handling measures, see call center KPIs and formulas. For how the two approaches differ at the design level, see conversational IVR versus an AI voice agent.

Build the report from each AI call's recording, transcript and outcome

Every DialerBee AI call produces a result: the recording, the transcript and the outcome, pushed into the CRM so the AI call sits in the same customer history as the human ones. That is enough to build the report, and it needs no guesswork about what the agent did.

  1. Résultat gives you the contained flag and the call reason. A call with a recorded outcome and no transfer is contained.
  2. Transcription lets you search for phrases such as "speak to a person" or "that's not what I asked", which are signs of trouble in a call that never transferred.
  3. Enregistrement is the evidence for the weekly listening sample.

The site does not describe a built-in containment dashboard, so treat the metric as one you define and build. Use the reports and insights exports and the analytique dashboards for the surrounding numbers, such as abandons and call outcomes, and join them to the AI results in a spreadsheet or your BI tool.

Listen to a sample of contained calls every week

The number cannot tell you whether the answers were right. Only a person can. Set a fixed routine:

  • Each week, pull a random sample of contained calls. Make it large enough that a reviewer can finish it, and keep it the same size week to week.
  • Add every contained call that was very short, every call with a repeat contact, and every call where the transcript contains a request for a person.
  • Listen to the recording, not only the transcript. A transcript can read smoothly over a call that sounded wrong, with long silence, talking over the caller or a mispronounced name.
  • Score each call as resolved, partly resolved or not resolved, and record why.

The share scored "not resolved" in the sample is your estimate of false containment. If it moves, the headline rate has stopped meaning what it used to. For a fuller review method, see the QA scorecard in how to QA AI voice agent calls.

Worked example (illustrative numbers only)

These figures are invented to show the arithmetic. They are not a benchmark and not a DialerBee result.

StepAppels
Outbound calls dialled1,000
Voicemail, no answer and busy (removed)600
Connected conversations (the denominator)400
Transferred to a person80
Hung up with no recorded outcome60
Reached an outcome with no transfer260
Of those, repeat contact inside the window30

Read the same week three ways:

  • Naive rate: calls with no transfer, 320 of 400, is 80%. This counts the 60 hang-ups.
  • Containment with a recorded outcome: 260 of 400 is 65%.
  • Containment net of repeats: 230 of 400 is 57.5%.

If the listening sample then finds that one in ten of the 230 was wrong, the resolution estimate is about 207 of 400, or roughly 52%. The headline fell from 80% to about 52% without the agent changing at all. Only the method changed. This is why the definition has to be written down before the first report goes out.

When a higher containment rate is the wrong goal

Pushing the rate up has costs, and some of them are invisible on the dashboard.

  • It can hide the handoff. If the agent is rewarded for not transferring, it may be tuned to hold on to calls it should release. A caller who asks for a person should get one.
  • It rewards easy call mixes. A rate that rises because you stopped sending hard calls to the agent is a routing change, not an improvement.
  • It ignores calls that should leave. Complaints, vulnerable callers and high-value decisions often belong with a person.
  • It can trade satisfaction for cost. A call contained with a bad experience still costs you the customer.

A better goal is a high rate of verified resolution on the reasons the agent owns, with a clean and quick handoff on the rest. Treat containment as a diagnostic, not a target.

Foire aux questions

What is a good containment rate for an AI voice agent?

There is no honest universal number, because containment depends on the call reasons you give the agent. A booking line and a billing dispute line can contain very different shares of calls. Set your own baseline from the first four weeks of listened-to calls, then track the trend per call reason. A rising rate only counts if the sampled calls still sound resolved.

Is containment the same as resolution?

No. Containment says the call ended without a transfer to a person. Resolution says the caller's problem was actually solved. A caller who hangs up in frustration is contained but not resolved. Measure resolution by checking the outcome, by looking for a repeat contact within your window, and by listening to a weekly sample.

Should voicemail count in an outbound containment rate?

Leave it out of the denominator. A call that reached voicemail never had a person to contain, so counting it dilutes the rate and hides how the agent performs on real conversations. Report voicemail and no-answer as separate outcomes beside the rate.

How long should the repeat-call window be?

Pick a window that matches how long a genuine fix takes in your business, such as 24 hours for a booking question or several days for a billing question, then keep it fixed so weeks stay comparable. The number is a policy choice, not a standard. Write it next to the metric.

Can I trust containment alone to judge the agent?

No. A rate can rise because the agent got better, or because it became harder to reach a person. Always report it beside transfer rate, abandons, repeat contacts and a satisfaction measure, and listen to a sample of contained calls every week.

Where does the data come from in DialerBee?

Each AI call leaves a result with its recording, its transcript and its outcome. Those three items are the raw material for the report. The site does not describe a built-in containment metric, so define the formula yourself and build it from those results and from your report exports.

Lectures complémentaires

Prêt à voir DialerBee en action ?

Demandez une démo en direct de 15 minutes, ou démarrez un essai gratuit de 14 jours — sans diaporama, sans engagement. Les appels démarrent dès que votre propre trunk SIP ou PBX est connecté et approuvé.

Essai gratuit de 14 jours · crédit IA gratuit · sans carte · 11 langues · BYOC · contrôles qui contribuent à la conformité

Voir le site complet en anglais →