Pipeline
Coaching software mostly produces compliments.
Ask a model to review a call and it will find something positive to say about almost anything, because that is what it is shaped to do. Which makes the output useless precisely on the calls that needed reviewing.
So the instruction has a floor written into it.
The floor
A call that lost contains at least three findable mistakes. Find them.
Written into the instruction, in those terms: a call that was lost, or where the rep talked more than two-thirds of the time, or where sentiment finished low, contains at minimum three distinguishable mistakes — find them, quote them, and list them separately.
Followed by what to look for: opening with a monologue instead of discovery, ignoring what the prospect just said, selling past the stated need, guessing at a price or a fact, dismissing a competitor rather than understanding why they are being considered, and not confirming a concrete next step.
And in the same breath, the counterweight: every entry still needs a verbatim quote — do not invent mistakes.
- Missed moves are their own list —
Openings the prospect gave that the rep did not take — a stated pain not dug into, a buying signal not pursued, a stakeholder mentioned and never asked about. Usually more useful than the mistakes.
- What went right is only collected on a won call —
And lessons only on a lost one. So "what worked" is grounded in something that actually worked.
- The measurements involve no model at all —
Share of talking, longest unbroken stretch, back-and-forth — computed arithmetically from the timings, identical whether the words came from audio or a transcript file.
- A third party is counted honestly —
Someone who is neither the rep nor the prospect counts toward the total but neither share — so a call with a partner on it shows shares that do not add to one, rather than quietly inflating somebody.
Comparison
Against the median — and the median excludes you.
Seven measures over a rolling four weeks, each ranged against the team's middle rather than against a target somebody invented.
The median is calculated without the person being looked at, so nobody is measured against a figure they helped set. And when somebody is the only person with calls in the window, the comparison is withheld and says so rather than comparing them to themselves.
It is deliberately not a rules engine. The intent, written into the code, is that a rep reads "you are third on discovery depth", not "you are a bad rep".
- Share of the talking68%58%
- Longest unbroken stretch1m 51s34s
- Back-and-forth3854
- Objections resolved74%66%
- Deals won31%28%
- Weeks start on Monday —
Because that is when the stand-up is.
- No calls returns a reason, not a blank page —
With a suggestion — upload one, or widen the window.
- Live coaching exists and is additive —
Prompts during the call, from a cheaper pass over what has been said so far. The full analysis still runs at the end and remains the record. It is off by default, because it costs materially more.
›Can a rep see their own coaching?
Yes — it is built to be opened by the person it is about.
›Does a manager get a leaderboard?
No. There is no ranking and no per-person severity colouring. The comparison is to a median, which is a position, not a place in a queue.
›Can I share one moment from a call?
Yes — a link scoped to a range you choose, optionally password-protected, with an expiry and a revoke. Nothing auto-generates a highlight.
›What languages?
The live path is configured for multilingual transcription. The coaching analysis itself is written in English, so a call in another language will transcribe better than it coaches.