Task completion rate
Task completion rate is the share of eligible calls or tasks in which an agent achieves the intended outcome end-to-end according to a written success definition.
This metric maps closely to business value because it asks whether the work was actually done. A successful greeting, accurate transcript, or plausible response may support the process, but none proves that the caller obtained the intended result. For an appointment workflow, completion might require creating the booking for the correct person, service, date, and time and then confirming those details. Merely attempting the booking would not qualify.
The denominator can change the result substantially. A rate calculated across all incoming calls includes wrong numbers, spam, callers who disconnect immediately, and requests outside the system’s scope. A rate limited to calls with a qualifying intent answers a different question. Restricting it further to connected calls, or to calls where required systems were available, changes the question again. Reports should state the denominator, exclusions, and treatment of calls containing multiple tasks.
Each workflow needs its own written definition of “complete.” For information requests, completion may require delivering an approved and relevant answer. For lead intake, it may require collecting every mandatory field and saving the record successfully. For scheduling, changing, or canceling an appointment, completion should include confirmation that the intended action occurred in the system of record. Definitions should also explain how partial completion, caller abandonment, duplicate actions, and later corrections are scored.
Task completion is related to containment, but the measures are not interchangeable. Containment asks whether an interaction avoided live human intervention. Task completion asks whether the intended outcome occurred. A transferred call can still produce a completed task if the automation gathers the necessary context and a person finishes the workflow. A contained call can fail if it ends without a transfer but leaves the caller’s request unresolved.
Measurement can combine several evidence sources. Reviewers can grade transcripts against workflow-specific rubrics, but transcripts should be checked alongside action logs and downstream records when the task changes external state. Simulated calls can test known scenarios repeatedly, including interruptions, corrections, and unavailable resources. A maintained validation set supplies representative examples with ground-truth intents, required actions, and expected final states. Human review remains useful for ambiguous cases and for auditing automated graders.
The rate is most informative when segmented by workflow, language, call condition, and failure reason. Teams should track whether failures came from speech recognition, reasoning, missing knowledge, policy boundaries, an unavailable integration, or incorrect action execution. That breakdown turns a headline outcome metric into a practical guide for improvement without weakening appropriate escalation rules.