Containment, resolution and transfer: the voice AI metrics teams confuse

Voice AI dashboards often start with a tempting number: containment. It appears to answer a simple question: how many calls ended without a human? The problem is that a caller can leave without a human for many reasons. The issue may be resolved, the caller may have accepted a useful next step, the system may have failed or the person may have given up. Treating all of these outcomes as success rewards the wrong behaviour.

Dring's analytics, quality workflows and Agent Factory are better understood through three separate concepts: containment, resolution and transfer. They can be connected in one scorecard, but they should not be collapsed into a single rate. The distinction is especially important when a company is deciding whether to expand a subscription, launch a new workflow or change an agent policy.

Containment describes what happened to the channel

Containment means the call ended without a human agent joining that interaction. It is a channel outcome, not a customer outcome. A routine order status call may be contained and resolved. A payment dispute that ends after a failed verification attempt may also be contained, but it is not resolved. Define the event precisely: did the agent end the call, did the caller hang up, did a tool fail or did the system move the case to another channel?

Keep caller abandonment separate. A call that ends before the agent has answered the question should not be counted as positive containment. Track silence, early hang-up, transfer request and system termination as separate events. Clear event definitions make the dashboard more honest and give the quality team a place to investigate.

Resolution describes the customer's job

Resolution should be tied to the intent and an observable next step. For an order question, it may mean the customer received the current status and did not need another action. For a return request, it may mean eligibility was confirmed and the label was sent. For an appointment, it may mean a booking was created and confirmed. For a complex support issue, resolution may require a human handoff that successfully reaches the responsible team.

Write a resolution definition per workflow. Do not use one universal rule such as “caller said thank you” or “call lasted under two minutes.” A polite ending can hide an unanswered question. If the outcome is a callback or specialist review, the case should remain open until that next step is completed or intentionally closed.

Transfer is not automatically failure

A transfer can be the correct result. A caller asking for a supervisor, a patient raising a clinical question, a suspected payment dispute or a safety report may need a trained person. Penalising every transfer encourages agents to delay or resist escalation. Measure transfer appropriateness and handoff completion instead.

A useful transfer passes identity state, intent, facts captured, action attempted, reason for escalation, language and customer expectation. If the human asks the caller to repeat everything, the transfer may be technically completed but operationally poor. Dring's handoff design treats context preservation as part of the outcome.

Build a metric tree, not a vanity leaderboard

At the top level, report volume, containment, resolution, transfer and abandonment. Under resolution, show workflow-specific completion. Under transfer, show transfer reason, reach rate, time to human ownership and rework. Under quality, show factual accuracy, policy adherence, next-step clarity and customer feedback. This lets a manager see both efficiency and care.

Use denominators that explain scope. A rate across all calls may hide a small, high-risk workflow. Break results down by intent, language, channel, customer segment, agent version and time period. Dring's 62-language technical capability inventory spans voice, WhatsApp, SMS and email, so language should be a normal reporting dimension rather than a footnote. The ten-language public launch-priority set still requires locale/workflow validation on the actual path before production.

Include the time after the call

Many outcomes happen after hang-up. A message is delivered, a return label is used, a meeting is attended, a specialist responds or the customer calls back. Add an outcome window that matches the workflow. For a simple status question, same-call completion may be enough. For a callback, evaluate whether the callback happened and whether the customer still needed help.

Use CRM write-back to keep the post-call state visible. The data model guide explains why intent, action, owner, timestamp and final outcome should be separate fields. Sector Insight can then reveal which contained calls later reopen, which may indicate overcounted success.

Guard against metric gaming

Every metric creates pressure. If teams are rewarded only for containment, they may shorten calls, avoid human handoff or define resolution too generously. Add counter-metrics: repeat contact, complaint, handoff request, failed action, reviewer safety and customer effort. Review calls close to the boundary, not only the best ones.

Use the Agent Factory to test changes that improve one metric without harming another. A new opening may lift containment but reduce clarity. A stricter handoff rule may lower containment but improve resolution for complex cases. Release decisions should consider the complete scorecard.

Example scorecard

• Containment: call ended without human participation, with abandonment reported separately.
• Resolution: workflow-specific completion confirmed in the allowed outcome window.
• Transfer quality: correct route, context passed and human ownership reached.
• Customer effort: repeat explanation, repeat contact and time to next useful action.
• Safety and policy: reviewer-rated accuracy, guardrail adherence and escalation quality.
• Learning: recurring themes, corrected outcomes and test cases added to the next release.

Containment is useful when it is placed in its proper role. It tells you something about channel load. Resolution tells you whether the customer's job was completed. Transfer tells you where human judgement belongs and whether the organisation made that transition well. Together they create a more credible view of voice AI than any one headline percentage.

Make the denominator part of the story

A metric is easier to trust when the denominator is visible. “Resolution rate” might mean resolved calls among all offered calls, answered calls, eligible workflow calls or cases that reached a final outcome window. Those populations answer different questions. Put the denominator, date range, workflow scope and exclusions next to the rate so an operations leader can interpret the result without opening a separate data dictionary.

Keep a stable definition for trend reporting, and publish a second view when the workflow changes. For example, a new identity step may temporarily increase transfers while making completed outcomes more reliable. Explain the change rather than forcing the new period into an old formula. Clear definitions protect the team from metric gaming and make conversations about agent improvement much more productive.

Further reading

• Twilio on conversation intelligence and actionable call data
• McKinsey on efficiency and customer experience in care operations
• NIST on measuring and managing AI system risks