Does the dialogue behave the way a human conversation behaves, particularly under disruption?

Conversational quality is the dimension most correlated with whether an agent feels like an assistant or a script being read aloud, and it is disproportionately revealed not in the smooth part of a call, but in the disrupted part. Any agent can sound competent when the caller follows the happy path in order. What separates a genuinely well-built agent is what happens when the caller interrupts, changes their mind mid-sentence, or asks the agent to recall something from four turns earlier.

Two agents can score identically on accuracy and resolution and still produce completely different caller impressions, because this dimension governs the texture of the exchange rather than its outcome. We weight interruption handling and context retention most heavily here, because in our testing they are the two behaviors most tightly correlated with a caller trusting the agent enough to use it again.

Does the agent respond fast enough to sustain the illusion of a real conversation?

Latency is the dimension where voice diverges most sharply from every other AI modality. A chat interface can pause for two seconds and the user barely registers it, there's a cursor blinking, a sense that something is happening. A voice agent that pauses for two seconds sounds like the call dropped. The tolerance for delay in a spoken exchange is governed by decades of research on human turn-taking, where the natural gap between one speaker finishing and another beginning sits in the low hundreds of milliseconds. Anything meaningfully past that band reads as hesitation, confusion, or failure, regardless of what's actually happening in the pipeline.

This is also the dimension most distorted by averages. A mean response time can look healthy while a long tail of the distribution, the calls that hit a slow backend lookup, a cold model start, or a network hiccup, quietly produces the exact freezes that erode trust fastest. We score latency at every stage of the pipeline independently: speech recognition, language model inference, and speech synthesis, because a spike in any single stage is a different engineering fix, and averaging across them erases the information needed to find it.

What did the interaction leave the caller feeling, beyond whether the task got done?

Experience metrics exist because task completion and satisfaction are not the same variable, and treating them as interchangeable is one of the more consequential measurement errors a voice AI team can make. A caller can get exactly what they asked for and still leave the interaction irritated: repeated back their account number twice, spoken over mid-sentence, made to feel like they were talking to something that wasn't quite listening. None of that shows up in a completion rate. All of it shows up in whether that caller calls back next month, or quietly moves to a competitor.

This is also the dimension with the weakest direct instrumentation and the strongest predictive value. Survey response rates are low, so we lean heavily on behavioral proxies: how sentiment moves across the arc of a call, whether the same caller rings back within a short window, and specific linguistic markers of frustration that appear reliably before a formal complaint ever does.

Did the agent correctly understand what was said, before anything downstream had a chance to go right?

Accuracy is the foundation layer of the entire stack, and errors introduced here are the most expensive to catch late, because they don't announce themselves; they masquerade as failures somewhere else. A misheard word becomes a misclassified intent, which becomes the wrong flow, which becomes a resolution failure three steps downstream from where the actual mistake occurred. Teams that only measure task completion frequently spend days debugging a "resolution problem" that was, in fact, a transcription problem the whole time.

The complicating factor is that accuracy measured under clean, quiet, lab-recorded audio tells you almost nothing about accuracy on a real phone call from a moving car or a warehouse floor. We test this layer exclusively under realistic acoustic conditions: background noise, accented speech, imperfect telephony, because the gap between clean-room performance and real-world performance is frequently the single largest number on this page.

Did the agent accomplish what the caller actually needed, independent of how the call sounded?

Resolution is the dimension that answers to the business, not to the transcript. It is possible, and in our testing common, for a call to read as a complete success in every qualitative sense: polite tone, clear diction, a confident closing line, and still fail entirely, because the action the agent claimed to take was never committed on the backend. This is the single largest source of false-positive success we encounter in pre-launch testing, and it is invisible to any evaluation method that only reads the transcript.

Treated rigorously, resolution metrics require ground truth outside the conversation itself: a booking system, a CRM record, a payment ledger. We verify outcomes against the system of record, not against what the agent said it did, because the two diverge more often than most teams assume before they start measuring it directly.