Coaching & scorecards
The AI scorecard, custom rubrics per team, and manager review that never overwrites the original.
Every completed call gets a coaching scorecard from the post-call pipeline. Managers can change the rubric, and can disagree with a score without erasing it.
The default rubric
With nothing configured, calls are scored on four dimensions:
| Dimension | Question it asks |
|---|---|
| Discovery | Did the rep uncover the real problem? |
| Objection handling | Were objections answered with evidence? |
| Value articulation | Was value framed in the customer's terms? |
| Next steps | Did the call end with a concrete commitment? |
Each dimension carries a weight, a description, and guidance the model is given when scoring.
Custom templates
Not every team sells the same way. A team selling to CISOs should be graded on security depth; an SMB team probably should not be graded on multi-threading.
Settings → Scorecard templates lets an admin define up to twelve dimensions per template, each with:
- a
lower_snake_casekey, - a label,
- a weight (0–10),
- an optional description and scoring guidance.
A template can be assigned to a team or left as the org default.
Resolution order
Team template → org default → built-in rubric.
The team's active template wins if there is one; otherwise the org default; otherwise the four built-in dimensions. An org that configures nothing still gets a sensible scorecard, and a team that needs something specific does not force the rest of the org to adopt it.
The resolved rubric is handed to the post-call pipeline through the call pack, so scoring happens against the right dimensions without any per-team code.
Manager review
Open a call and a manager can adjust any dimension's score and record why.
The AI's original scorecard is never modified. A review is stored as a separate artifact (scorecard_review) carrying the reviewer's identity and rationale, and reading a call's scorecard returns both the original and the review.
Two reasons this matters:
- Coaching disagreement stays auditable. A rep can see what the model said, what their manager said, and why, instead of a number that silently changed.
- Model calibration stays measurable. If reviewers consistently raise "Discovery" by a point, that is a signal about the rubric or the model, and it is only visible because the original survived.
Gametapes and clips
The coaching surface auto-selects teaching moments from real signal: exemplar clips (most suggestions used) and teachable clips (most dismissed, or the highest talk ratio), each with a plain-language note about why it was picked.
From a call you can:
- Comment on a specific transcript turn or on the call as a whole,
- Clip a range and share it by link,
- Assign a drill to a rep, which notifies them.
Coach alerts during the call
Scorecards are retrospective. During the call the desktop HUD raises live coach alerts: long monologue, talk ratio drifting high, dead air after a question, discovery gaps late in the call. The thresholds are configurable per team; see Per-team policies.
API
GET /v1/calls/:id/scorecard returns the AI scorecard plus any review. PATCH /v1/calls/:id/scorecard (manager and above) records a review. Templates live at /v1/scorecard-templates.