Call analytics, how to tell whether your voice agent is handling calls well

It is the Analytics section of the panel. It brings together every call in a period and tells you how many your agent handled, how they went and where it got things wrong, without you having to listen to them one by one.

At the top you choose the period, today, 7 days, 30 days or 12 months, and if you have more than one agent, which one you want to see or all of them together. Everything below is worked out from the real calls in that period. The tests you run from the panel do not count.

For an insurance broker, for example, it answers three questions. How many claims the agent is taking, whether it is leaving them complete, and whether at any point it told a customer something that is not true.

Outcome. How many calls there were, split into inbound and outbound, what share was completed, how many the agent passed to a person on your team and how long they last on average. The first three come compared with the previous period, so you know whether things are going up or down. A rise in calls passed to your team is not flagged as bad, because sometimes that is exactly what the agent should do.

How they end. Whether the caller hung up, the agent hung up when it was done, the call was passed to someone, voicemail picked up or it was cut off for silence. On the calls your agent makes, you also see whether the line was busy, nobody answered or the number did not exist.

Volume. Calls by day, by hour and by status. It tells you when people call most, which is exactly when the agent most needs to be well tuned.

Business outcomes. If you asked for a yes or no detail in your agent’s Post-call analysis, such as “claim registered” or “wants a call back”, here you see in how many calls it came out yes and in how many no.

And if you have chat agents, further down you see their conversations, the agent’s messages, how many were closed for inactivity and their sentiment.

The most useful part is Quality. When a call ends, an automatic reviewer reads every call, not a sample, and looks for the turns where the agent said something that does not add up. A turn is each time the agent speaks.

The reviewer is another AI model, and it reads the call with the same things your agent had in front of it. Its instructions, its knowledge base and what your system sent back when the agent checked something. That is how it tells apart opening hours the agent took from your knowledge base and opening hours it made up.

It flags three things. Made it up, when the agent gives a detail that is nowhere to be found. Contradicted itself, when it says something that goes against its instructions or against what it has just said. Misunderstood, when it gets wrong what it was told. On agents that hang up on their own, it also flags Hung up too early. Each flag is marked Serious or Minor.

Flagged turns lists them, with the reason written out in full and a link to the call, Open the call. Open it and read the sentence. The reviewer can be wrong too, which is why you always see the reason and not just a label.

Next to it are other figures that are not opinions, because the system measures them. Confidence understanding is how sure the system was that it understood what the caller said. The agent went quiet counts the times it fell silent because of a noise or because what was said did not reach it. Say it is a virtual assistant shows in how many calls the agent said so at the start. And Data discarded by the engine, on flow agents, counts the details the agent was about to accept and the system rejected, such as a policy number with digits missing.

Functions. A function is the request your agent sends to your system in the middle of a call, to look up a policy or register a claim. Here you see how many times each one ran, how many times it failed, how long it took and the most frequent error. The ones that fail come first.

Conversation phases. On a flow agent, the kind that follows a set of steps, you see which phases callers go through and which one they are in when they hang up. If many calls end in the phase where the policy number is asked for, something there needs a look.

Latency. How long the agent takes to answer, split into steps, from the moment the caller stops talking. Understand what was said, know they finished, search the knowledge base, think the answer and start speaking, plus how long it takes to greet when it picks up. For each step you see the median, the middle value, and the p90, which only the slowest answers go over, one in ten.

If a step or a phase cannot be measured with your agent, for example phases on a single prompt agent, the panel does not show a zero. It tells you why there is no figure, so a gap never looks like good news.

First, open Flagged turns and read the calls. If at the broker the agent gave office hours that are not in its instructions, the problem is not the reviewer, it is that the agent is missing that detail. Add it to its instructions or to the knowledge base.

Second, look at Functions. If the request that registers the claim fails, the error tells you whether your system did not respond, took too long or the address is not valid. That gets fixed in your system or in the function, not in what the agent says.

Third, look at How they end and, if your agent is a flow agent, Conversation phases. Lots of calls cut off for silence, or always stuck at the same step, show you where the conversation gets stuck.

Then publish the change and look again in the next period. The Outcome figures come compared with the previous one, and Flagged turns tells you whether the same thing keeps happening.

You can also ask Maite, the panel’s assistant. She can read your latest calls, with their summary, sentiment and transcript, and suggest changes to the agent.

Analytics gives you the overall picture. To look at a single call, with its summary, intent, urgency and sentiment, meaning whether the caller ended up happy or annoyed, there is Call history, where you can also filter by sentiment.

Calls are kept for as many days as your account keeps them, 30 by default. If you choose a longer period, the panel tells you there is no data before that date.

Anyone on your team with access to the panel can see the analytics, including people who join as Viewer and can only look.

If you would rather have the Lera team review and adjust your agent, we will send you a quote. And the Advanced plan includes a quarterly review of the agent with us.

Does it review every call? It is not a sample. The reviewer reads the calls once they end, without keeping anyone waiting. It only has a monthly safety cap, around 3,500 calls, so that a fault cannot run up the cost, and Analytics shows how many it reviewed. Audio tests in the panel are not reviewed and don’t count; the outbound test is, because it is a real call.

What if the reviewer flags something that was fine? Nothing. It is an AI model giving its opinion and it is sometimes wrong. That is why you have the reason and the call right there, so you can check for yourself.

Why do I see a zero in one place and an explanation in another? Where a figure cannot be measured with your agent, the panel explains why. Data discarded by the engine only exists on flow agents, so on a single prompt agent there is no point looking at it.

Can I see the analysis of each call? Yes, in Call history, with the summary and the details you asked for in Post-call analysis.

How much does it cost to try? €20 free credit, no card needed. Create your agent, call it yourself a few times and see how the calls look in Analytics.