Platform · Conversation intelligence
Every call is scored. Here is the whole rubric.
Six dimensions, weighted, adding to one hundred, published with the two thresholds that turn a score into a verdict. Nobody in this category publishes their weights, and a score you cannot argue with is not evidence.
Included at $499 per seat per month
The composite, in full
points, summing to 100
- Talk balance25
- Discovery depth20
- Objection handling20
- Monologue control15
- Filler discipline10
- Opener strength10
Two thresholds are the entire verdict. There is no third number held back, and no model deciding the band on its own.
Read this before you quote the number
Guidance, not a measured target. Competitor-derived heuristic with no peer-reviewed basis, not yet calibrated against SAGARIS outcomes.
When the call ends
Conversation intelligence in seven steps. Three of them stop for a person.
The moment a call ends an ordered pipeline turns it into a record: a summary, a disposition, a CRM stage, a drafted follow-up, a reschedule decision, an invite and the qualification signals. Three of those seven produce something a human approves rather than something that has already gone out.
The post-call pipeline, in order
per completed call
- 1Call summary
Topics, objections, interests and next steps, structured from the transcript.
- 2Disposition
The call is classified, and the classification drives everything below it.
- 3CRM stage
The contact's canonical stage is projected from what the call actually was.
- 4Follow-up emailWaits for you
Drafted from the recap in step one. There is no send path in the pipeline.
- 5Sequence reschedule
The recovery action is decided, and an opt-out ends the sequence instead.
- 6Calendar inviteWaits for you
Composed when a meeting was booked, and only when one was.
- 7BANT signalsWaits for you
Extracted onto the lead as a suggestion the rep confirms, never as fact.
Every step is fail-soft. One unconfigured model records itself on that step and the other six still run, so a bad afternoon costs you a summary rather than the whole record.
Eight named objections it recognises
Objection handling is twenty of the hundred points, so the classifier has to name what it heard rather than score a feeling. A ninth bucket, unknown, catches what it will not guess at, which is the honest half of any taxonomy this short.
- price
- timing
- authority
- build it ourselves
- competitor
- status quo
- feature gap
- trust
Why this is different
We own the phone system, so consent is a gate and not a checkbox.
Conversation intelligence usually sits beside the telephony and asks it for a recording. Ours is the telephony: an Asterisk PBX we run on our own trunk, placing and receiving the calls. That moves the consent decision to the moment before recording starts, which is the only moment it can actually prevent anything.
Where recording needs consent first
one cell per US area code
145 / 372
Area codes where recording cannot start until consent is captured on the call.
227 / 372
Area codes where the statute allows one party to consent.
14 all-party states, plus Vermont held as unresolved rather than assumed permissive. The map is generated from the NANPA registry with the file date and the checksum of that file pinned in the source.
And when it does not know
A number the registry cannot place resolves to no jurisdiction, and no jurisdiction is treated as consent required. The gate refuses the case it does not understand instead of waving it through, which is the difference between a gate and a setting.
The gate sits above the recorder
The jurisdiction check runs before recording is armed on the channel, not after the audio exists. That is only possible because the box taking the call is ours. A tool bolted onto someone else's phone system can ask for a recording; it cannot decline to make one.
The unknown case fails closed
An area code the registry does not carry resolves to no jurisdiction, and no jurisdiction is treated as consent required. Vermont sits in the same bucket by choice, because it has no general eavesdropping statute to rely on rather than because a court settled it.
One system, one record
The call, the score and the contact record are the same system, so nothing is exported and no second vendor ends up holding your customers' voices. The score is attached to the record it was made from, which is also the only place it means anything.
The questions buyers actually ask.
- MOST ASKED
No. The six weights are a constant inside the scoring module and the scoring function takes only a signals object, so there is no per-workspace override to configure. The reason we have not built one is the same reason we publish the numbers: the rubric is not calibrated against outcomes yet, and letting every team retune an uncalibrated heuristic would turn one unverified number into fifty of them.
We will not quote you an accuracy figure. We will show you the arithmetic.
Bring a call to a demo and we will score it in front of you: which dimension lost the points, what that dimension was worth, and which band the composite landed in. The disclaimer will be on the same screen, because a number without its provenance is a number you cannot use in a review.