Use correctly completed, policy-compliant tasks as the economic denominator, with autonomous and human-rescued outcomes reported separately. This is the desk’s recommended measurement design.
| Measure | Proposed definition |
|---|---|
| Autonomous completion | Unique eligible tasks completed correctly without human intervention, divided by unique eligible tasks assigned to the agent |
| Overall completion | Unique eligible tasks completed correctly, including human rescue, divided by unique eligible tasks assigned |
| Record accuracy | Audited task records meeting every required field and policy check, divided by audited task records |
| Handoff success | Escalations reaching the approved destination with required context, or a correctly owned deferred pathway, divided by attempted escalations; report live and deferred routes separately |
| Repeat-contact rate | Unique tasks generating additional contact within the pre-agreed observation window, divided by unique tasks handled |
| Policy failures | Counts and severity of prohibited statements, actions, disclosures or communication-control failures |
| Fully loaded cost per successful task | Total cost within the defined workflow boundary, divided by unique correct, policy-compliant completions |
Include the cost of unsuccessful attempts in the numerator. Include telephony, AI usage, integration operation, monitoring, human handling, rework and repeat contact where attributable. Show implementation cash outlay separately from any explicitly justified steady-state allocation.
Keep cash savings, capacity released, avoided hiring and incremental contribution from additional completed work separate. For missed-call recovery, deduct demand already recovered through existing callbacks and test whether fulfilment capacity exists.
Set pass/fail thresholds and the observation period before launch. Choose the evaluation size around the workflow’s call mix, decision needs and consequences of error; this quiz does not supply a universal minimum call count.