Call analytics: how the AI calls went, and what they cost

Whether it is working, working how, which prompt is better, and when you need people on the queue. Cost is not a footnote here — you pay the carrier directly, so it is the number that decides whether a cheaper model is worth trying.

The Call Analytics page covers calls your AI agents handled. It is ordered by the question being asked rather than by what is easiest to plot.

Call analytics — outcomes, agents, hours and cost

Is this working?

The KPI strip at the top: how many calls, how long they ran, what they cost. This is the glance that tells you whether anything needs a closer look.

Working how?

The outcome split. An agent that handles every call alone and an agent that transfers every call are both plausible, and they mean opposite things — one is doing the job, the other is an expensive switchboard. The split is what separates them.

Handled — the agent finished the call itself.
Handed off — the agent transferred to a human, which for a qualification agent is success and for a support agent may not be.

Which prompt is better?

The per-agent table. When you change an agent’s instructions, this is where the change shows up — compare the same agent before and after, or two agents doing the same job with different prompts.

When do I need people on the queue?

The hourly pattern. If your handoffs cluster at particular hours, that is when a human needs to be reachable, and staffing to it is cheaper than an agent transferring into an empty queue.

Cost

You pay Telnyx directly, so cost is not a footnote on this page — it is the number that decides whether a cheaper model is worth trying. The figures are the carrier’s own, reported back when the call ends, not an estimate we calculate.

A call that cost nothing usually means it never connected. Worth checking rather than celebrating.