Call Coaching

Overview

Call Coaching answers the question a rep has the moment a call ends: was that good, and what do I fix next time. Each call is scored on four dimensions, Clarity, Effectiveness, Control and Next Step, which roll into a 0 to 100 composite and a 1 to 4 grade. Expand a call and every variable shows the published benchmark it was scored against, so no number on the card is unexplained.

What it is for, on the page. Under the headline, this page says what it produces, what makes it run, and where the output lands. Every Work page reads that line from one registry, so the three answers cannot drift from the page.

It is the per-call companion to the weekly behavioral read, which compares your won calls to your lost ones. That read needs several closed calls before it can say anything. This one needs one transcript.

The four dimensions

Effectiveness (weight 35) asks whether you ran real discovery and landed the value message: question depth, consequence development, whether the pain was quantified, the value message itself, how many questions you asked and how evenly you spread them. Control (25) asks whether you guided and listened: talk-to-listen ratio, objection handling, patience before you speak, how often the conversation switched speakers, and whether the buyer ever got a run at telling their story. Clarity (20) covers your longest monologue, conciseness, filler rate, talk speed and whether you signposted. Next Step (20) asks whether you secured a concrete dated advance, and it is weighted up despite having few signals because a set next step is the strongest single per-call predictor of the outcome.

Consequence development

When a buyer names a problem, the next question matters more than the next slide. This signal asks whether you found out what the problem costs, blocks or spreads to, in the buyer's own terms, and whether the buyer said more. A question like "When I hear the backlog doubled, renewals are usually next to slip. Is that showing up for you?" scores near the top. A stock "how does that affect you?" scores about half. Describing the product in the next turn or two, without asking, scores near zero, and the card quotes the line where you went to the pitch.

The signal is judged by the AI coaching pass, so it shows only on calls you ran the pass on, and it always quotes the call. When the buyer named no problem at all, it reads not measured, buyer named no problem and its weight moves to the other Effectiveness signals. When the quote the model gave is not in the transcript, it reads quote not found in transcript. If you keep a consequence map in Settings, under Brand & Company, the card also names the row your question matched, for example Matched your map: renewals slip. That match compares words and spends no extra model call.

What changed in the weights. Question depth used to give full marks for reaching implication questions, which is the same behavior, so it now gives full marks for problem and need-payoff questions instead. To make room, the Effectiveness weights moved: question depth 20, consequence development 15, pain quantified 20, value message 20, question count 15, question spread 10. They still sum to 100.

Rubric versions

Every score remembers which version of the rubric it was graded on. Scores from before consequence development was added are version 1; new ones are version 2. Nothing is re-scored behind your back: a call keeps its version until you run the AI coaching pass on it again.

Because version 2 weights Effectiveness differently, a composite from one version and a composite from the other are not the same measurement. When your window holds calls from both, the composite trend compares calls within the newer version only and says Rubric changed on the date it changed. Per-signal trends carry no such note, because a signal means the same thing in both versions. Composites for newly scored calls can shift by a few points because of the new weights, not because you changed anything.

What a signal we could not measure does

It is excluded, and its weight moves to the other signals in its dimension. It is never scored zero.

This matters. A transcript without timings makes patience and talk speed unmeasurable: calls ingested from your call recording now carry per-line times when the vendor supplies them, but older calls and pasted transcripts do not. Scoring those zero would tell you that you interrupt constantly, when the truth is that we could not measure it. A dimension where nothing was measurable reads "not scored" rather than showing a number, and every card states how much of the rubric its score actually covers, so a thin score never passes as a complete one.

The AI coaching pass, and the citation check

Eleven of the signals are computed straight from the transcript with no model involved, and those run automatically on the weekly pass. Nine more need a model reading the call: question depth, consequence development, whether pain was quantified, the value message, objection handling, conciseness, structure, mutual agreement on the next step, and who owns what. Those run when you ask for them, on one call at a time, and they draft only. Nothing is sent and nothing is written to your CRM.

How to ask. Every call on the panel carries a Run AI coaching pass button in its summary row. Press it and the nine judged signals are scored for that call alone; the card then reads AI coaching pass has run and the button becomes Run AI pass again. A re-run asks for confirmation first, because it spends a second model call on a call that already has a score. Only one pass runs at a time, so clicking across several cards cannot queue up spend.

Every AI score has to cite a quote from the call, and the quote is checked against the transcript. Matching ignores capitalization, spacing and punctuation, so a re-styled quote still passes; it does not ignore words, so a paraphrase does not. A score whose quote is not in the call is dropped, its signal goes unmeasured, and the card tells you it happened. It is not scored zero, because a made-up value message is not evidence of a bad value message.

When a pass cannot run, the card says which of the reasons it was: the call has no transcript to read, the model is not configured on this deployment, the model call itself failed, or the model was cut off before it finished writing its answer. That last one is a budget, not a fault in your call, and it says so: a very long transcript can outrun the output the pass is allowed, and the fix is a shorter transcript rather than a re-run. None of these is shown as a call that simply has not been coached yet, because they need different things from you.

Your claims

Below your calls sits a section for the three or four value claims your pitch rests on. Name them once, and every time you run the AI coaching pass on a call, the same pass records whether you said each claim and what the buyer did next. Claims carry no weight in the call score. A call scores the same with or without them.

Getting started. With no claims saved, the section suggests up to four, picked from your Brand & Company messaging in Settings and your latest value-framing brief. It prefers sentences with a number or a named outcome. Nothing is saved until you edit the list and press Save. You can keep up to five; three or four works best. Changing a claim's wording keeps its history, and the row tells you the date the wording changed so you know older marks judged the older text. Removing a claim archives it with its marks.

The five marks. Repeated means the buyer restated the claim in their own words. Asked means they asked about it. Pushed back means they doubted or rejected it. Moved on means you said it and the buyer went somewhere else. Not stated means you never made the claim on that call.

Every mark quotes the call. A stated mark carries your line and the buyer's line, and both are checked against the transcript the same way the AI scores are. A mark whose quote is not in the call is dropped, never stored, and the call card says which claim lost its mark and why. Expand a call to see its marks with both quotes.

Won vs lost, and the five-call threshold. Each row shows how many judged calls the claim was stated on, the mix of reactions, and the buyer quote that best shows it landing. It also splits how often buyers picked the claim up (repeated or asked) on deals you won and deals you lost. A call counts toward an outcome when its account has a closed deal and the call came before the close. That split stays hidden until the claim was stated on at least five decided calls on each side; until then the row shows how many decided calls there are so far. Every number is printed as a count over its total, so you can see how thin the evidence is.

Marks only exist for calls you chose to judge. Nothing judges old calls in the background.

Benchmarks for one call, your own baseline for a trend

The anchors behind each signal come from published conversation-intelligence research, the largest of which analyzed hundreds of thousands of recorded calls. They define a green zone for a single call's card, and that is all they do.

Trends work differently. The dimension view, and the list of every signal beneath it, compares your newer calls to your own earlier ones, never to other people, because the durable finding in that research is that behavioral stability under pressure marks the strong rep, not the absolute level. A number that is fine for one seller and poor for another is not a coaching signal; a change in your own numbers is.

One signal is personal even on a single call: filler rate. Once four or more of your calls have a measured rate, new calls are scored against your own rolling median rather than the generic band, so a naturally chatty style is not marked down for being yours, and a drift above your own norm is. Under four calls the published cold-start band applies.

Where the calls come from

Call Coaching reads the transcripts your connected tools already bring in, the same calls as the weekly behavioral read. Nothing new to install and nothing to paste. Calls are scored on the weekly pass, and the panel shows them newest first with a 7, 30 or 90 day window or all time.

Every source that keeps a transcript counts: Gong, the meeting bot, calls your own tools push through the Coach API, and the Zoom, Fathom, Google Meet and Teams connectors. Each card says which one the call came from. Before November 2026 only Gong calls were scored, so your counts can rise once the other sources are read. Every one of these counts on your own page. Your manager's team view counts only calls the platform fetched from a recorder itself, so a call you push yourself through the Coach API, or a browser whisper session, is yours alone.

A score needs to know which speaker is you. Gong and the Coach API can say so with an internal participant list, Fathom marks who is on your side, and Zoom names the host. Otherwise we match the name in your login email, so sam.rivera@ finds a speaker called Sam Rivera. A call where no speaker can be matched to you is left out rather than scored as if you never spoke, and a transcript with no speaker names at all, like a browser whisper session, is left out the same way.

The AI coaching pass runs on calls captured as call records: Gong, the meeting bot and Coach API pushes. Calls scored straight from a Zoom, Fathom, Meet or Teams transcript show their transcript-only score without the button.