Make a test call
Ring the number from a phone that is not in your workspace and hold an ordinary conversation — a real question, an interruption, and something you know it cannot answer. Then read the call record: the transcript, the summary, and the response timings. A call that felt fine and reads badly is more common than the reverse, which is why both halves matter.
A test worth running#
- Ask a straightforward question you know the answer to. This checks grounding.
- Interrupt the answer halfway through. On the realtime engine this checks barge-in; a first attempt inside the first couple of seconds of a greeting is deliberately ignored, so interrupt properly rather than instantly.
- Ask something you know is not in your Knowledge. This checks whether uncertainty is expressed or covered up.
- Ask for a call back, or a quote. This checks whether a commitment becomes a follow-up rather than a sentence nobody records.
What to listen for#
| Symptom | What it usually means | Where to look |
|---|---|---|
| A long pause before the first word | Prompt size, or the model behind the line | The response timings on the call record |
| Answers that are correct but endless | Reply length is not being held down; a spoken reply should be far shorter than a written one | Voice tuning, then the transcript |
| It talks over you, or cannot be interrupted | Barge-in behaviour — the realtime engine supports it, the turn-based path does not | Which engine the line uses |
| Confident nonsense | No grounding material, so nothing specific to answer from | Knowledge, and the facts extracted from it |
| Silence after answering | A line-health problem rather than a conversation problem | Needs You, which line health escalates into by itself |
Reading the timings honestly#
Response time is measured from the end of your speech to Connect's first spoken word, and the numbers that matter are known. Around 2.5 seconds is the working target; past about 4 seconds a caller has already decided nobody is there. On the fastest configuration the first token has been measured at roughly 0.6 seconds, while a native-audio model sits nearer 1.2 to 2.2 seconds, and the best full call recorded a median reply of 3.3 seconds.
Those figures are worth carrying because they set expectations correctly: a well-tuned line is quick but not instant, and chasing the last few hundred milliseconds through settings has limits — a large share of the delay is the model rather than the configuration. Reply length is the lever that is genuinely in your hands; a median spoken reply of about nineteen words reads as brisk, and the same answer at forty words reads as a lecture.
The response watchdog sits above all of this: if nothing comes back within a few seconds the call is nudged rather than left in silence, because silence on an answered line is the worst failure a phone system has.
After you hang up#
Open the call on
#/calls.Result The transcript and a summary are attached to the call record, and the call is on the caller's timeline alongside their mail.
Read the summary against the transcript.
Result You find out whether the summary is a fair account or a flattering one — which is what anybody reading it next week will rely on.
Check whether anything you asked for became a follow-up.
Result A promise made on a call and recorded nowhere is the failure mode worth catching before customers meet it.
If the caller was new, check that a person record was created and that a second call from the same number would find it.
Result One person, one record, across channels — which is what makes the next conversation start from something rather than nothing.
Questions#
Can I listen to the recording afterwards?
No. Call recording is not enabled on the live carrier — the capability exists in the telephony layer but is not switched on, so the transcript and the summary are the record of what was said. Plan reviews around reading rather than listening.
The call sounded fine but the transcript is a mess. Which do I trust?
Both, for different things. The transcript is what the rest of Connect reasons over afterwards — the summary, the follow-ups, the relationship record — so a poor transcript degrades everything downstream even when the caller had a pleasant time.
How many test calls should I make before going live?
Enough to have heard each of the four situations above at least once, and at least one from a number the workspace has never seen. Repeating the same friendly question ten times proves less than four different awkward ones.