Call Containment: How to Measure and Improve It

Call containment rate is the share of calls an AI voice agent or IVR handles from start to finish without transferring the caller to a human. It tells you how much of your call volume the automated layer absorbs, which makes it the first number most teams report after launching a voice agent.

It is also the easiest number to inflate. A caller who gives up and hangs up counts as contained. That is why containment is only useful when you measure it next to resolution.

How do you calculate call containment rate?

Divide the calls the AI agent completed without a human by all calls the AI agent answered, then multiply by 100. A call is not contained if it was transferred to a person or if a supervisor took it over mid-call.

Call containment rate formula: Containment rate = (AI-answered calls − Transfers − Take-overs) ÷ AI-answered calls × 100

10,000 AI-answered calls, 2,600 transferred and 200 taken over by a supervisor: 7,200 ÷ 10,000 = 72% containment. 900 of the 7,200 contained calls ended unresolved: 6,300 ÷ 10,000 = 63% resolved containment.

Report resolved containment too: only contained calls where the caller's task was completed.

The second line of that example is the one to report upward. Resolved containment counts only the contained calls where the caller actually got what they called for. The gap between the two numbers is the group of callers who left without help, and they will show up again as repeat calls.

Why is containment rate on its own misleading?

Containment measures where the call ended, not whether it worked. Traditional IVR reporting made this worse: anyone who did not reach an agent counted as contained, including people who hung up on a menu loop. Three checks keep the number honest:

• Pair it with resolution. Track the share of contained calls marked successful, by call reason.
• Watch repeat calls. If containment rises and callbacks rise with it, the agent is ending calls, not resolving them. The first call resolution benchmarks article explains how to join AI and human records to catch this.
• Look at sentiment on contained calls. A rising share of negative sentiment on calls that never reached a person is an early warning.

A hang-up is not a resolution. Count a call as contained only when the caller left with what they called for. Everything else is a repeat call waiting to happen.

How is containment different from deflection and resolution?

These terms get used interchangeably in vendor material, but they count different things. Agree on which one you report before you compare numbers across teams or vendors.

• Containment rate: What it counts: Calls the AI finished without a person. Blind spot: Counts frustrated hang-ups as success.
• Resolved containment: What it counts: Contained calls where the task was done. Blind spot: Needs a clear success definition per call reason.
• Deflection rate: What it counts: Contacts moved away from live agents to self-service. Blind spot: Says nothing about whether the issue got solved.
• First call resolution: What it counts: Issues solved with no repeat contact. Blind spot: Needs AI and human call records joined.

What containment rate should you expect from an AI voice agent?

There is no reliable cross-industry benchmark for AI voice agent containment. Published figures use different definitions (some exclude transfers the caller asked for, some count hang-ups as contained), and most come from vendors. The rate also depends almost entirely on which calls you route to the agent and what systems it can act in.

What does exist are individual deployments with published results. Everise contained 65% of voice calls to its internal service desk with a Retell AI agent covering issues such as account access and software problems. In its Matic Insurance case study, 80% of customers complete AI-handled calls without asking for a human, while its scheduled-call agent transfers to licensed agents by design. Both are single companies with specific call mixes. Use them as proof that high containment is possible, then set your own baseline per call reason during a pilot.

How do you measure call containment in Retell?

Retell records why each call ended and whether it met your success criteria, so you can build containment and resolved containment from the same data. The analytics dashboard documentation lists every chartable metric.

1. Chart the transfer rate. In Analytics, add a chart with Call Transfer Rate, the share of calls transferred. Containment is the remainder once you also exclude take-overs.
2. Break calls down by how they ended. Chart Call Counts by Disconnection Reason. call_transfer and call_take_over are escalations. user_hangup, agent_hangup and inactivity are endings without a person, and they need the resolution check.
3. Add resolution. Chart Call Successful Rate next to it. Customize the prompt behind the built-in call_successful field so "successful" means the caller's task was completed.
4. Record why calls escalate. Under Post call extraction, add a selector such as transfer_reason (caller asked for a person, policy requires a human, agent could not complete the task). Any extraction field can be used as a chart breakdown.
5. Check transfers that fail. Retell's AI quality assurance reports transfer success rate and transfer wait time, and an alert on transfer call failures tells you when callers are being dropped instead of handed off.

Break every chart down by agent version. When you change a prompt, you will see within days whether containment moved and whether resolution moved with it.

How do you improve containment without trapping callers?

Raise containment by making the agent able to finish more tasks, not by making it harder to reach a person. The changes that move it most:

• Route only the call reasons the agent can complete end to end.
• Connect the agent to the systems that close the request (booking, CRM, order lookup).
• Load approved answers into a knowledge base and keep it current.
• Review transfer reasons weekly and fix the most common one first.
• Turn failed calls into test cases and rerun them before each release.
• Keep a one-step path to a person for callers who ask.

Keep the exit open. A caller who asks for a person should get one quickly, with context passed along through a warm transfer so the human agent does not start from zero. That transferred call then shows up in your human team's average handle time, which is where a good handoff pays for itself. To test a change safely, Retell's A/B testing can split a phone number's traffic between two agent versions (for example 80/20) so you compare containment and resolution on live calls before rolling out.

Teams running inbound call center operations usually start with a few call reasons the agent can complete end to end, prove resolved containment there, and expand from that baseline.

The containment rate that matters is the one you measure on your own calls. Retell runs pilots on real call traffic, so you can see containment and resolved containment by call reason before you sign anything. Talk to the Retell team about a pilot.