The metrics that flatter you
The first dashboard most teams build shows call volume, containment rate, and average handling time, and all three can improve while the service gets worse. Containment — the share of calls handled without a human — is the worst offender, because the easiest way to raise it is to make escalation harder. An agent that stubbornly refuses to transfer will post an excellent containment number while callers hang up in frustration, and those hang-ups do not appear anywhere in that metric. Average handling time has the same problem in reverse: it falls when calls end early, and calls end early both when the agent resolves things efficiently and when the caller gives up. Neither number is useless, but neither can be read on its own, and a dashboard showing only these will tell you everything is fine right up until someone complains.
Abandonment is your most honest signal
The clearest indicator that something is wrong is callers hanging up mid-conversation with the agent. It requires no interpretation and no survey: the caller decided this was not worth continuing. Track it as a rate, then break it down by where in the flow it happens. Abandonment clustered at one step points at a specific broken question — often one that asks for something people do not have to hand, or that the agent keeps mishearing. Abandonment spread evenly across the call usually points at something systemic, such as latency or an unnatural speaking rhythm that makes the whole interaction tiring. This single breakdown will direct your improvement work better than any other report, and it is usually the cheapest one to build.
Look at distributions, not averages
Averages hide the failures you most need to see. A reasonable-looking mean call duration can be a blend of very short calls where the agent failed immediately and very long calls where it looped without resolving, and the average of two failures can look exactly like success. Plot the distribution instead. Very short calls are worth sampling for immediate misunderstanding or callers who bailed the moment they realised it was automated. Very long calls are worth sampling for loops, repeated re-asking, and confirmation cycles that never terminate. The same applies to latency: an acceptable average can conceal a tail of responses slow enough to make callers repeat themselves, and it is the tail that drives the complaints.
Listen to real calls on a schedule
Every team that runs voice agents well does one unglamorous thing: someone listens to real recordings every week, on a fixed cadence, whether or not anything looks wrong. Ten calls a week, chosen partly at random and partly from the abandonment and long-duration buckets, will surface problems no metric captures — an agent that is technically correct but sounds curt, a phrase that confuses people, an accidental implication that the caller has been penalised. Rotate who listens, and include someone who is not on the build team, because the people who wrote the prompts have stopped hearing them. This is also where you find the scenarios worth adding to your regression suite.
How this works in Persistence
Persistence includes monitoring alongside building and testing, so the loop from noticing a problem to changing the agent to verifying the change stays inside one place rather than spanning three tools and a spreadsheet. Calls can be reviewed with their transcripts and the agent's reasoning together, which matters because the interesting question is usually not what the agent said but why it said it. The practical payoff is speed: a bad call you spot on Monday can become a saved test scenario and a shipped prompt change without waiting on an export. If you are still evaluating platforms, ask what the post-launch story looks like — plenty of tools help you build an agent and leave you alone with it afterwards.
Key takeaways
- Containment rate on its own is a vanity metric — an agent that never transfers can be failing every caller.
- Abandonment during the agent conversation is the most honest early warning signal you have.
- Watch the distribution of call durations, not the average; the interesting failures live in the tails.
- Listen to a fixed number of real calls every week. No dashboard substitutes for hearing the agent get something wrong.
