Measuring ROI on a voice automation project

ROI on a voice automation project is measurable with real precision if you define the right baseline metrics before launch — most of the ambiguity we see comes from not doing that upfront.

The metrics that actually matter, and the ones that don't

"Calls handled by the automated system" is the metric vendors like to lead with, but it's incomplete on its own — a system that handles a high volume of calls poorly, driving customers to call back repeatedly or escalate angrily, isn't actually delivering value even with a high raw handling number. We track four metrics together: containment rate (percentage of calls resolved without human transfer), average handle time compared to human-agent baseline, caller satisfaction specifically on automated interactions (measured via a brief post-call survey), and — critically — repeat call rate, since a caller who calls back within 24 hours because the automated system didn't actually resolve their issue represents a hidden cost that raw containment numbers miss entirely.

Establishing the baseline before you build anything

You can't measure improvement without knowing your starting point. Before any voice agent build, we establish baseline numbers from the existing system: current average call volume by type, current average handle time by call type, current after-hours call volume (often simply lost or voicemail today, representing pure opportunity), and current staffing cost per call. This baseline is what every post-launch number gets compared against — without it, "the voice agent is working well" is a feeling, not a measured claim.

The cost side of the equation, calculated honestly

Voice AI infrastructure has real, ongoing per-minute or per-call costs (the underlying speech-to-text, language model, and text-to-speech usage), which need to be weighed against the labor cost of human-handled equivalent calls, not just presented as "free" automation. We build this cost model explicitly — for a call type costing roughly $4-6 in fully-loaded human agent time versus a fraction of that in automated handling cost, the ROI case is straightforward; for a call type that's genuinely complex and requires extensive escalation even when routed to the voice agent first, the automation cost can end up additive rather than substitutive, and we flag that honestly rather than force-fitting every call type into an automation narrative.

A concrete example with real numbers

A home services company deployed a voice agent for appointment scheduling and basic service inquiries, with a pre-launch baseline of 1,100 monthly calls, average handle time of 4.2 minutes for a human agent, and a fully-loaded agent cost of roughly $28/hour. Post-launch, the voice agent achieved a 68% containment rate on scheduling-specific calls (the highest-volume category), with average handle time for those automated interactions at 2.1 minutes — actually faster than the human baseline, since the system didn't need small talk or hold time for looking up availability. Post-call satisfaction surveys on automated interactions averaged 4.1 out of 5, comparable to their human-agent baseline of 4.3.

Calculated against fully-loaded labor cost avoided and the voice AI platform's per-minute cost, the client's own finance team calculated payback on the initial build investment within approximately 5 months, with ongoing monthly savings after that point.

How Ndakum approaches it

Every AI Voice Agent engagement starts with establishing this baseline — you get a real, measured ROI case, not a vendor promise, before and after launch.