Resolution Rate, Not Containment: How to Measure Agentic Voice AI

Resolution Rate, Not Containment: How to Measure Agentic Voice AI

Short answer: Containment rate measures how many calls never reached a human. Resolution rate measures how many callers actually got what they called for. As voice agents move from answering questions to completing tasks, resolution rate — paired with escalation quality and cost per resolved contact — is the only metric that tracks real value.

Key takeaways

• A call can be fully "contained" and completely failed. The caller hangs up, calls back tomorrow, or churns silently.

• Agentic voice AI — agents that call APIs to book, reschedule, update or pay — makes containment even more misleading, because the agent can now fail after appearing to succeed.

• The metric set that survives scrutiny: resolution rate, repeat-contact rate within 72 hours, escalation quality, task-level success on write actions, and cost per resolved contact.

• Most voice agent failures in production are integration and edge-case failures, not speech failures.

What is containment rate, and why is it misleading?

Containment rate (also called deflection rate) is the share of inbound contacts handled entirely by an automated system without transfer to a human agent. It became the default voice AI metric because it is trivial to instrument — you already know whether a transfer happened.

The problem is that it counts an absence, not an outcome. These all count as contained:

• The caller got a correct, complete answer and left satisfied.

• The caller got a wrong answer, believed it, and acted on it.

• The caller gave up and hung up.

• The caller was told to "visit our website" for something the agent should have done.

• The caller hung up and called back an hour later, where they counted as a fresh contact.

Only the first is a success. Reporting all five as one number is how a dashboard shows improvement while customer satisfaction falls.

What changed with agentic voice AI?

First-generation voice agents mostly talked. They answered FAQs, read out balances, and routed calls. The worst outcome was an unhelpful answer.

Agentic voice AI completes tasks. It calls tools mid-conversation: checking real availability, moving an appointment in a scheduling system, updating an address in a CRM, issuing a refund, taking a payment through a compliant handoff. That changes the failure surface completely.

An agent that talks can only be unhelpful. An agent that acts can confirm a booking that never wrote to the calendar, update the wrong record, double-charge, or promise a slot that a race condition just gave away. All of those calls are contained. Some of them are worse than a busy signal.

Resolution rate vs containment rate

How do you actually measure resolution?

Resolution needs a verifiable definition per intent, decided before launch. The pattern:

1. Define the completing event for every intent. "Reschedule" resolves when the calendar row changed. "Balance enquiry" resolves when the balance was stated and no repeat contact follows. If you cannot name the event, you cannot claim resolution.

2. Instrument the backend, not the transcript. A voice agent saying "you're all set" is a claim. A row in your booking table is evidence. Log tool-call outcomes with the call ID so every conversation carries its own receipt.

3. Track repeat-contact rate within 72 hours, split by intent. This is the cheapest silent-failure detector in the business, and it catches the failures no transcript review would flag.

4. Score escalation quality separately. A fast, well-contextualised transfer is a good outcome. Measure how often the human received the caller's context versus how often the caller had to start over.

5. Report cost per resolved contact, not cost per call. A cheap call that resolves nothing is pure overhead plus a second contact.

6. Watch abandonment inside the agent — where in the flow callers hang up. Abandonment concentrated at one step is a design bug, not caller impatience.

Our data: [Insert your own before/after table: containment rate, resolution rate, and repeat-contact rate for one deployment. A single honest comparison, including where the numbers disagree, is the most citable asset on this page.]

Where do agentic voice agents actually fail?

Not usually where teams expect. In production, the recurring failure modes are:

• Integration edges. The record exists twice. The phone number has no match. The API times out at three seconds and the agent confidently continues.

• Write actions without confirmation. The agent takes an irreversible action on an ambiguous instruction. Every mutation needs an explicit readback before it executes.

• No idempotency. The caller repeats themselves, the tool fires twice, the appointment is booked twice.

• Latency dead air. A four-second backend query with no holding line reads as a dropped call, and the caller hangs up mid-success.

• Hallucinated authority. The agent confirms a price, a delivery date or a slot it cannot actually see. This is a prompt design failure, and it is the single most damaging one.

• Broken handoff. The transfer works, but no context travels with it, so the caller repeats the whole story to a human — and rates the experience on that.

What does a resolution-first scorecard look like?

Report these five together, per intent, weekly:

Containment can stay on the dashboard. It just cannot be the headline, and it must never be the number a vendor is paid against.

How do you keep it from regressing?

Voice agents drift when prompts, models, or backends change — usually all three, separately, without coordination.

Build a scenario suite from real calls: your fifty most common paths plus your worst twenty edge cases, including the hostile caller, the wrong number, the caller who changes their mind twice, and the one who speaks two languages. Replay it before every release. Add every production failure to the suite as a permanent regression test, so the same bug cannot ship twice.

Then review recordings weekly, sampled by outcome rather than at random — every unresolved call and every abandoned call, not a convenience sample of the ones that went well.

Frequently asked questions

What is a good resolution rate for a voice AI agent? It depends entirely on intent mix, so cross-company benchmarks are close to meaningless. Simple, well-integrated intents like appointment rescheduling reach far higher resolution than open-ended troubleshooting. Set your baseline from your own human agents on the same intents.

Is containment rate useless? No — it is a useful cost signal and a fine secondary metric. It becomes harmful only when it is the headline number or the term of a vendor contract, because it can be improved by making transfers harder rather than by solving problems.

How is resolution rate different from first call resolution? First call resolution is the long-standing human contact centre metric measuring issues closed in one contact. Resolution rate for voice AI applies the same logic but verifies completion against a backend event rather than agent self-reporting.

Should a voice agent be allowed to take irreversible actions? Only with an explicit readback and confirmation immediately before execution, idempotency protection on the tool call, and a scoped permission model. Anything genuinely irreversible — large refunds, account closures — is usually better routed to a human.

How do you measure escalation quality in practice? Sample transferred calls and check whether the human agent had the caller's identity, intent and conversation summary at pickup, and whether the caller had to repeat information. Report it as a percentage of transfers arriving fully contextualised.