A transfer metric is misleading when “the system dialled someone” counts as success. During a pilot, define completion as the intended person accepting a usable conversation, then report the stages that prevented that outcome.
Handoff completion needs an explicit denominator
AI-to-human handoff completion measures whether an eligible caller request reaches an appropriate human destination. Define eligibility before the pilot: business hours, allowed destinations, caller consent to wait, and any verification prerequisites.
Keep requested, offered, answered, accepted, bridged, and resolved as separate events. An answered destination might be voicemail. A completed audio bridge might connect the wrong department. A successful handoff does not establish that the customer’s underlying issue was resolved.
For a proposed pilot, assign one interaction reference and separate attempt identifiers. Count callers and attempts independently. Otherwise repeated offers to the same unavailable employee can inflate both activity and apparent success.
Calculate a fictional pilot result
Suppose 40 eligible callers request a person. Thirty receive a confirmed human connection, six receive an agreed callback instead, and four leave without a next step. Live handoff completion is 30 divided by 40, or 75 percent. The callbacks belong in a separate follow-up measure.
If those callers generated 52 ringing attempts, reporting 30 divided by 52 answers a different question: attempt efficiency. Both can be useful, but their labels and denominators must remain visible. These figures illustrate calculation only; they are not observed project results.
| Pilot question | Required evidence |
|---|---|
| Did the employee accept? | Explicit acceptance or reviewed equivalent |
| Could both parties hear? | Audio check or authorised review |
| Was the destination appropriate? | Routing reason and staff disposition |
| Did a callback happen later? | Linked follow-up outcome |
Sample failures as well as successes for human review. A dashboard can report completion while recordings reveal caller confusion or context missing from the staff screen.
Segment results before changing the prompt
Compare staffed periods, destinations, transfer modes, and provider routes. A poor answer rate at one understaffed desk is unlikely to improve through conversational wording alone. Similarly, one-way audio requires transport investigation.
Small pilots have uncertain rates. Publish counts beside percentages and avoid ranking destinations with only a few calls. Operational changes during the pilot should be annotated so their effects are not attributed to the original configuration.
The voice integration case study establishes related engineering work without claiming pilot outcomes. Use cross-provider transfer testing to investigate route-specific failures before expanding traffic.
Handoff measurement FAQ
Does voicemail count as a human handoff?
Keep it separate unless the approved workflow explicitly defines voicemail as its intended destination.
Should caller abandonment count as failure?
Report it in the eligible-caller funnel and record when it occurred. Do not silently remove difficult cases.
Which metric belongs beside completion rate?
Show waiting time, fallback acceptance, context availability, and unresolved callers so a higher connection rate does not hide a worse experience.