Real-time discovery coaching should prompt the next evidence-seeking move. It should not turn one transcript fragment into a diagnosis, a pitch, or a script the rep must follow.
Discovery-specific coaching is narrower than general live call guidance. Its job is to notice when a configured evidence gap appears—impact is missing, the decision process is vague, the next step has no owner—and offer a short question or reminder while the moment is still useful.
Define the job of discovery coaching
Start from the team's visible discovery call plan. For each stage, define the minimum evidence needed to move on. Coaching may prompt for that evidence; it may not declare the evidence present.
| Observed state | Missing evidence | Permitted prompt |
|---|---|---|
| Buyer states a problem | Operational impact | “Ask what this changes for the team.” |
| Buyer gives an impact | Priority or timing | “Test why this matters now.” |
| Solution criteria emerge | Decision participants | “Ask who will evaluate this.” |
| Call is closing | Owned next step | “Confirm action, owner, and date.” |
Set the boundary between coaching and evaluation
Live coaching may remind the rep to seek evidence or slow down. It should not grade the rep, update qualification, or write CRM conclusions during the call. Those jobs need fuller context and separate review. Keep the live surface small and reversible: the rep can use, dismiss, or ignore a prompt.
Define prohibited prompts as clearly as permitted ones. Do not surface unverified claims about the buyer, private internal commentary, a full response the rep must read, or a commercial conclusion based on sentiment or one keyword.
Map prompts to evidence gaps
For each discovery stage, list required evidence, acceptable sources, counterevidence, and the point at which prompting would be too late. A problem prompt expires once impact is being discussed. A next-step prompt should not appear while the buyer is still explaining a constraint.
Write a discovery prompt contract
Every prompt needs a trigger, required context, missing evidence, permitted text, suppression rules, freshness window, and owner. Suppress when the evidence already appeared, the rep is in a different call stage, speaker identity is uncertain, or another prompt is active.
Keep prompts short enough to scan without abandoning listening. The rep decides whether to use them; dismissal and correction are useful feedback.
Add frequency and priority controls
Allow one active prompt at a time, set a cooldown, and rank prompts by the cost of the missing evidence. Merge prompts that seek the same fact. Suppress a lower-priority prompt when the buyer is speaking, the rep is handling a sensitive objection, or a higher-priority evidence gap remains open.
Store why each eligible prompt was delivered, suppressed, expired, or dismissed. Without suppression telemetry, a team cannot distinguish a quiet, accurate system from one that stopped processing.
Make the prompt packet auditable
Record call ID, participant state, source transcript window, trigger contract version, required context, missing evidence, confidence, prompt text, delivery time, expiry, rep action, and degraded-state flag. This packet supports QA without placing the full transcript in a coaching metric.
Handle transcript uncertainty
Zoom documents real-time participant-attributed transcript packets with timestamps and language metadata. It also notes that low-latency fragments may need post-processing because pauses can split sentences. A prompt engine should preserve that uncertainty.
- Wait for corroborating context before firing on ambiguous phrases.
- Use speaker attribution; the rep mentioning price is not the same as a buyer objection.
- Expire prompts when the conversation has moved on.
- Suppress content-dependent prompts when language identification or transcript continuity is unstable.
Use corroboration and negative tests
Test paraphrases and multi-turn context, not only trigger keywords. The phrase “we have budget” can be hypothetical, quoted from another stakeholder, or negated. Require the configured context before treating it as evidence. Seed false friends such as the rep repeating the buyer's words, background speech, and a competitor name in an unrelated story.
When a sentence arrives in disjointed fragments, wait for completion or suppress. A delayed but correct prompt can still be wrong for the current conversational moment.
Design for interruptions and degraded state
Zoom's RTMS documentation distinguishes signaling and media interruptions and their reconnection paths. The coaching UI should distinguish current, reconnecting, degraded, and unavailable states. A stale prompt presented as current is worse than no prompt.
Test stream interruption during a trigger, replayed transcript packets, participant reconnection, speaker changes, delayed fragments, language switches, and a prompt already dismissed before the delayed trigger arrives. Reconnection should not replay expired coaching.
Also respect meeting access and disclosure. Zoom documents participant notifications and host approval states when apps access meeting content. Follow applicable company policy and law in addition to the platform controls.
Define degraded behavior before reconnection
When transcript continuity breaks, stop content-dependent prompts, clear or mark active prompts stale, preserve the last verified sequence point, and show the rep that coaching is degraded. After reconnection, discard fragments and prompts whose freshness window elapsed. Do not replay missed coaching in a burst.
Test signaling interruption, media-only interruption, participant reconnection, speaker identity change, language switch, duplicated packet, reordered fragment, and recovery after the rep has moved to another stage. The expected result should specify what the rep sees.
Evaluate coaching without vanity metrics
Track prompt eligibility, delivery, acceptance, dismissal, false triggers, material corrections, expiry, degraded-state suppression, and whether the configured discovery evidence was ultimately captured. Review by prompt type and call stage.
More prompts are not better. A healthy system may suppress most possible prompts because evidence already exists or confidence is insufficient.
Calibrate with labeled call moments
Create a bounded evaluation set of permitted recordings or scripted roleplays with labeled eligible moments, non-eligible near matches, evidence already present, and degraded-stream cases. Have enablement reviewers agree on the expected prompt or suppression before comparing system output.
Report precision by prompt type, material false triggers, missed eligible moments, timing, dismissals, and evidence completion. Keep outcome claims separate: an accepted prompt does not prove a deal advanced, and a deal outcome does not prove a particular prompt caused it.
Use rollout gates
- Every enabled prompt has a current contract and owner.
- False friends and degraded-stream tests produce suppression.
- Disclosures and meeting-access states are visible and respected.
- Reps can dismiss prompts and report why without disrupting the call.
- Operators can disable one prompt type without stopping the entire coach.
Where Gangly fits
Gangly's documented Live Call Coach surfaces objection responses and context during Zoom or Google Meet calls while the rep remains in control. Evaluate its configured prompts, source coverage, disclosure, transcript uncertainty, degraded-state behavior, and current integration requirements. Repository facts do not establish prompt accuracy or outcome lift.