Calls

Calls is where you check what actually happened. Every call the assistant answered has a record with the outcome, the reason, the transcript and the exact playbook version that produced it.

Everyone in your account can read calls. Group managers see only their own communities'.

The list#

Calls in progress appear at the top and update live, so you can watch a call as it happens. The list of past calls, and an open call's page, reload as calls change: at most once every 5 seconds, and only while the page is on screen. A call's page stops reloading while you are showing its sensitive answers.

Filters#

What each outcome means#

A call records the furthest point it reached. A caller who turned down a tour and then hung up, or whose call reached the 20-minute limit after a callback was taken, shows that outcome, not Caller hung up. A call the assistant ended over abusive language is recorded as Caller hung up unless something stronger had already happened.

One call#

Click the date in any row.

The page has these panels.

Summary#

A three-line summary written after the call, plus the outcome, the community, the playbook version, the caller's masked number and the length. Test calls are marked Yes: not billed and not in analytics.

What the caller told us#

Every answer the assistant collected, by its short name.

Something the caller offered before being asked, and never confirmed, is here too, marked mentioned, not confirmed. It is not an answer: no rule read it, your reports do not count it, and it was not written to a CRM field. See Questions and flows.

Answers you marked Sensitive — health and care details — are hidden behind a Show button. Revealing them is written to the audit log, so your account has a record of who looked.

Rule that fired#

Which booking rule decided this call, and what it did.

This is the fastest way to answer "why didn't it book them?". The panel also lists the details the rule matched on, so you can see what the assistant knew at the time.

Tour#

When a booking was made: the time, the kind of tour, who gives it, the confirmation number from your CRM, and whether it is still confirmed or has been cancelled.

Callback#

When a callback was taken: the caller, their number, the best time to reach them and when it is due.

Guardrail hits#

Times a safety check stepped in on this call — an off-topic request, a blocked sentence, a caller trying to change the rules. Most calls show None. The agent stayed within its guardrails.

Recording#

When recording is switched on for your account and the caller did not object, press Play recording. The link stops working after a few minutes; open the call again for a new one.

If there is no recording you see either There is no recording for this call or, when the caller objected, No recording: the caller asked not to be recorded. See Call recording.

Transcript#

Turn by turn, with what the assistant did along the way written as short sentences: "Checked who is calling", "Searched the knowledge", "Checked open times", "Booked the tour", "Transferred to a person".

Call debugger#

Everything the assistant's code decided on the call, in order, beside the question map. Press Open the debugger at the bottom of the page.

• The map numbers the path the call took and ticks the questions it answered. Click a line of questioning to see only what happened on it.
• What happened is the timeline, one sentence per event:
  • where the call moved;
  • every rule checked, and why it matched or not;
  • each answer saved, and anything the caller volunteered, shown as mentioned, not confirmed until they confirmed it;
  • each talking point said before the offer, one left out because an answer or lookup it uses was not in, and a Say this point the assistant reworded;
  • each tool the assistant used, including tools from your other systems, with how long it took;
  • guardrail hits;
  • what code told the assistant to do on each turn.
• Open a row to see the details: what a tool was sent and what came back, every rule in a set with the reason it did or did not match, and the facts the rules read.
• Problems only narrows the timeline to failed tools, refused moves and guardrail hits. It is the quickest way to answer "what went wrong on this call?".
• Timing shows how long the caller waited on each turn, and which step was slow.
• Click the time on any row to jump to that moment in the transcript.

Sensitive answers are hidden in the debugger too, and so is anything that would give them away. Press Show sensitive answers to see them. As with Show, that is written to the audit log, and so is opening the debugger.

When a caller asked not to be recorded, the debugger keeps what the assistant decided but not what the caller said or answered. Calls from before the debugger was added have none.

What the model was sent

Press Show prompts to see exactly what the language model read on each step of the call: its instructions (the playbook brief, the location's details, the next question and the answers so far), the notes for that turn, the conversation, tool calls and their results, and the tools it could use. If the assistant left out something you expected, this shows whether the fact was there to use.

• Every test call and test chat keeps its prompts.
• Real calls keep them only when our team turns prompt capture on for your account. Ask support if you need it while you investigate something.
• Prompts are not masked, so each look is written to the audit log like Show sensitive answers.
• When a caller asks not to be recorded, the prompts kept so far are deleted and no more are kept.

Callbacks#

The Callbacks tab next to All calls is your team's work queue.

Each card shows the caller, their number, the reason, the best time to reach them and when it is due. Overdue ones are marked in red. Press Mark done once your team has called back.

Closing a callback's commitment or task on Today or in My tasks closes the callback too. Done marks it done; Dismiss marks it Dismissed, closed without a call back. Dismissed callbacks are listed with the done ones, never count as kept, and can still be marked done later.

A callback marked done from a connected AI app or through the API can carry an optional note of what happened. The card then shows it after Note:.

If a callback is still open after the community's target time, your account's owners and admins are emailed, and the community's notify numbers are texted when texting is set up. See Human backup.

Viewers can read callbacks but cannot mark them done.

Using this well#

• Read every Booking failed call the same day. It usually means a CRM problem that will affect others.
• Filter by Version after a publish to see whether the change did what you hoped.
• Skim a few Caller hung up transcripts each week. They show you where the conversation loses people.
• Use Top unanswered questions in the call metrics (or Knowledge gaps, with the add-on) to decide what to add to Knowledge.

How long calls are kept#

Transcripts, the call debugger and recordings are deleted automatically after the retention period set on the playbook's Advanced tab: 90 days for each by default. See Data and privacy.