Dashboard → Health — an at-a-glance answer to "is the agent working?".

What it shows:

• Traffic-light status — green / yellow / red, computed from the last 24h of calls
• AI replied % — out of "engaged" calls (excludes <10s caller hangups)
• Silent calls — ≥10s where the bot never spoke (real failures)
• Caller hangups — <10s, the caller dialled and hung up
• Turned away — calls refused before the AI could answer, because a plan limit was reached. Shown only when there are some, and any at all turn the status red — these callers never reached you.
• Line noise — how often line hiss was heard as words (higher on old narrowband phone lines)
• Telephony status — connected phone lines and calls in progress (admins only)

«Where interactions came from» on the home page counts only what CUSTOMERS sent you — messages you send first (campaigns, reminders) are named on a separate line underneath, never mixed in. The same definitions as the volumes table below it, so the two can never disagree. Note that a message is one row: a person who writes five times is five incoming messages but one person, which is why the header also shows how many PEOPLE contacted you.

The status card on the Overview page shows the same last-24-hours signal, so the two never disagree. If you had calls in that window it reports on calls — identical to this page. If you had none — on a text-only plan, or on a line that gets a call or two a day — it switches to your conversations instead and reads "AI answered, all channels", counting every WhatsApp / Telegram / Instagram / widget conversation and showing how many an operator took over; on a line with calls it adds that there were none today. On a line that usually takes ten or more calls a day, a day without a single one is not normal: the card turns amber, reads "No calls in 24 hours" and says how many you usually get — check your phone line. That card is always the last 24 hours — it answers "is my AI working right now", so the period buttons above the chart do not change it.

«How the agent speaks, by week» — every phone call is checked shortly after it ends for what makes a good answer: whether it answered the question, in the caller's language, with your facts said exactly, without repeating itself or promising what it did not do. The table shows, week by week, how often each went wrong. Read it as a trend, not as a single number.

«The four things that matter, by week» — the first table of the voice section. Untrue statements: how many times that week the agent claimed something that did not happen, or handed a caller over saying it could not make them out right after they had said a whole, clear sentence; the goal is zero. Settled where given a chance: of the calls where the customer did not ask for a person in their first words, the share settled on the first contact. Asked twice — got a person: the agent offers to help once when someone asks for a person, and the second ask must always put them through; below 95% is worth a look. How many customers ask for a person is deliberately not here: people always will, and that is their choice, not the agent's failure. The manner of the agent is the table «How the agent works» below.

«How the agent works, by week» — each call is also rated the way a manager rates a receptionist, from 1 to 5, on seven traits: answers first (says what it can before asking anything back); one thought at a time (answers and stops); facts said naturally (the hours for the day asked, not the whole timetable); keeps the thread (never asks again what the caller already said); about the customer (what to do, never what it cannot do); solves it itself (hands over only what needs a person); tone and language. A dash means the trait did not come up in enough calls that week. Start with the lowest number. Low on facts or on solving it itself usually means an answer is missing: add it to Settings → Say these word for word or to the Knowledge Base.

«The line, by week» — what the network did to the caller's voice on its way to the agent, measured by the phone system at the end of every call: the typical packet loss, jitter and round-trip, and the share of calls on a bad line (loss above 2%). A bad line means the caller's own connection or distance — on WhatsApp, how far they are from the nearest WhatsApp server — not the agent. When callers say the voice sounded «robotic» or «electric», look here first: if their calls show loss, it is the line; if not, write to us.

«Did the customer get their answer, by week» — a day after each call it is counted whether the customer's question was settled on the first contact: no person was needed (no transfer, no call-back promised) and the same customer did not call or write again within 24 hours. «Handed to a person» and «Came back within a day» are the two ways it was not. Coming back is not always a failure — a second question, a photo sent in the chat — so read the weekly trend, not one call. A rising «handed to a person» usually means the agent lacks an answer: add it to Settings → Say these word for word or to the Knowledge Base.

«How the conversation flows» — the second table measures turn-taking, not words: the agent starting to speak while the caller was still mid-thought; the caller hanging up while the agent was still talking; the caller asking «hello?» in the middle of the call (they could not hear the agent or thought the line was dead); more than a third of the agent's speech never heard because it kept being cut; the agent failing to make out the caller twice; the caller having to repeat themselves; the agent stopping when the caller only said «yes» or «uh-huh»; a silence of more than two and a half seconds after the caller finished; the agent talking on for more than two seconds after the caller started to speak over it; the agent starting most of its replies with the same word; the agent asking for a phone number or a name the caller had already given; the agent talking about itself — what it can or cannot do — instead of what the customer should do. Each column is the share of calls where it happened; the last column is how long the caller waits for an answer — the typical wait, and after the slash the wait that one turn in ten is longer than. Lower is better everywhere.