Salesforce live sales coaching is not one feature connection. It is a controlled path from meeting media to transcript, from transcript to a trigger, from Salesforce to the correct account context, and from approved knowledge to a prompt the rep can use or ignore. Buy and test those stages separately.
Direct answer: Salesforce can supply opportunity, account, contact, activity, and approved knowledge context. A meeting platform or other authorized capture path supplies the live conversation. Keep those planes separate, bind them with stable identity, and do not let a live inference become CRM truth automatically.
Define the Salesforce coaching boundary
Start with the exact decision the prompt may help. Suitable jobs include retrieving an approved answer, showing a current battlecard, reminding the rep of a required question, or surfacing a known opportunity fact. “Improve every call” is not testable.
Salesforce Trailhead describes Agentforce Sales Coach around deal-specific practice, role play, and feedback before meetings. Salesforce Architects separately publishes a live architecture pattern in which Zoom RTMS transcript segments enter Salesforce, trigger classification and knowledge retrieval, and deliver guidance. The distinction matters: pre-call practice and live buyer-call assistance have different media, latency, privacy, and attention requirements.
Write a coaching contract for each trigger: observable phrase or state, exclusions, Salesforce evidence allowed, approved response, maximum frequency, useful timing window, abstain condition, severity if wrong, content owner, and version.
Separate CRM context from live media
A reliable design has four planes. The media plane supplies authorized audio or transcript events. The identity plane maps meeting, participants, account, contact, opportunity, and owner. The knowledge plane retrieves approved facts and battlecards. The action plane renders a prompt and records the rep’s response.
| Plane | Authoritative input | Stop condition |
|---|---|---|
| Media | Eligible live session and transcript stream | Capture unavailable, paused, or unauthorized |
| Identity | Meeting and Salesforce record IDs | Two opportunities or accounts remain plausible |
| Knowledge | Approved, owned, current content | No supported answer or the item is stale |
| Action | Bounded prompt contract | The buyer has moved on or prompt frequency is exceeded |
Do not join only on a meeting title, company name, or email address. Forwarded invitations, consultants, subsidiaries, duplicate contacts, and several open opportunities make those keys ambiguous.
Build the evidence packet
Every prompt should carry an internal evidence packet even if the rep sees only a short card: meeting and opportunity IDs, transcript span, speaker attribution, trigger label, Salesforce fields consulted, retrieved content ID and version, generated wording, confidence or abstention state, and display time.
The packet makes corrections possible. If the prompt used an old competitor card, the enablement owner can fix the content. If it matched the wrong opportunity, RevOps can fix identity logic. If the trigger fired on the rep naming a competitor, the classification set needs a speaker-context example.
Test prompt timing and errors
Measure the chain, not a blended “AI accuracy” score. Label eligible and ineligible moments across competitor names, abbreviations, homonyms, negation, overlapping speech, weak audio, pricing, security, legal questions, and silence.
- Trigger precision: correct eligible prompts divided by all prompts shown.
- Trigger recall: detected eligible moments divided by all labeled eligible moments.
- Evidence precision: supported factual statements divided by factual statements shown.
- Useful timing: correct prompts visible before the moment closed.
- Critical errors: wrong account, unauthorized context, unsupported commitment, or protected-data disclosure.
Report raw counts beside every rate. A rare cross-account disclosure cannot disappear inside a high average.
Control Salesforce writeback
Live prompts are assistance, not authoritative opportunity updates. Salesforce documents that sObject resources can operate on individual records; technical access does not grant business authority. Keep prompt logs and proposed follow-up separate from stage, amount, close date, forecast category, consent, and suppression.
After the call, show source evidence, proposed note, destination, and protected field suggestions to the rep. Write only the accepted values, use an idempotency key, retrieve the resulting record, and reconcile it with the approved payload. Route ambiguous or partial outcomes to an exception queue.
Run a bounded pilot
- Freeze trigger definitions, content versions, Salesforce mappings, and permissions.
- Run labeled recordings or role plays offline.
- Run shadow mode without showing prompts.
- Expose only low-consequence prompts to a small approved group.
- Inject wrong-record, stale-content, token-revocation, transcript-loss, and duplicate-event failures.
- Expand only when error, attention, and recovery gates pass.
Compare the live path with a post-call or no-prompt condition where practical. Count rep dismissals and distraction, not just card delivery.
Where Gangly fits
Gangly repository facts describe Salesforce account context and reviewed CRM updates plus live guidance on supported Zoom or Google Meet calls. The rep drives the conversation, and consequential CRM changes remain reviewable. Evaluate Live Call Coach with the same media, identity, evidence, latency, and recovery tests. This draft does not claim independent accuracy or outcome lift.