Competitor mention detection in sales calls is not a keyword-search demo. A useful live system must identify the right entity, speaker, context, and opportunity in time to retrieve current approved guidance without distracting the rep or exposing unsupported claims.
Direct answer: freeze a competitor alias registry, label eligible and ineligible transcript moments, test entity detection and context classification separately, measure complete latency, and allow a prompt only when identity, evidence, timing, and content gates pass.
Define the live detection job
Choose the output before the model: flag the moment for later review, tag the opportunity after approval, show a short live battlecard, or prompt one discovery question. These jobs have different error costs and latency requirements.
Zoom documents live transcript data through RTMS and its stream lifecycle. Salesforce publishes an architecture example that receives transcript segments, classifies them, retrieves knowledge, and delivers guidance. These documents establish possible mechanics, not detection accuracy, usefulness, or compliance.
Keep this job distinct from the human conversation skill in handling a competitor mention. Software should retrieve and remind; the rep diagnoses what the buyer means.
Build a competitor alias registry
For each competitor, maintain canonical name, products, abbreviations, common speech-to-text variants, former names, subsidiaries, excluded homonyms, owner, approved content version, and review date. Do not let a model invent a new competitor mapping during a live call without marking it unknown.
Include difficult cases: short names that are ordinary words, products shared across vendors, acronyms, possessives, domain names, partial names, regional pronunciations, and a buyer correcting the transcript. Separate exact aliases from fuzzy candidates.
Label context, speaker, and negation
A labeled test set should distinguish:
| Context | Example meaning | Live action |
|---|---|---|
| Buyer uses competitor | Incumbent may be present | Ask an approved discovery question |
| Buyer compares options | Active comparison is stated | Surface current differentiation evidence |
| Buyer rejects competitor | Negation or exclusion | Do not show an incumbent assumption |
| Rep names competitor | Seller introduced the entity | Usually no detection prompt |
| Historical story | Past relationship, not current deal | Record context; avoid live urgency |
| Homonym | Name refers to another entity or ordinary word | Abstain |
| Unknown name | Potential unregistered competitor | Flag for later review, no live claims |
Label speaker, transcript span, entity, context, polarity, deal relevance, prompt eligibility, and expected content. Cross-talk and weak diarization must produce explicit uncertainty.
Measure detection and useful latency
Separate entity precision and recall from context precision and recall. Then measure the pipeline: media delivery, transcription, entity detection, context classification, CRM lookup, content retrieval, rendering, and rep reading time.
Useful latency ends when the rep can still use the prompt, not when a server responds. Report median and tail values by context. Also report late correct prompts, duplicate prompts, prompt frequency, dismissals, and no-prompt compliance.
Keep critical errors separate: wrong account, another customer’s battlecard, unsupported security or pricing claim, stale content, unauthorized capture, and a prompt that reveals protected information.
Control the surfaced response
A live card should be short: detected entity and uncertainty, one approved distinction or fact with source, one discovery question, content owner and freshness, and an easy dismiss control. Link to the full battlecard for later rather than covering the call window with a document.
Do not generate discounts, legal positions, roadmap promises, security answers, or comparative superlatives from the transcript alone. Retrieve approved content, preserve its version, and abstain when no supported response exists.
Test failures and drift
- Run offline role plays and authorized recordings against the frozen labels.
- Run shadow mode and inspect false positives before showing cards.
- Force transcript interruption, wrong speaker, late join, duplicate segment, stale battlecard, revoked CRM access, and kill switch.
- Monitor new competitor names, aliases, transcription variants, content expiry, and rep corrections.
- Re-test every material registry, model, prompt, or content change.
Stop live display on cross-account disclosure, unauthorized capture, repeated distraction, or unavailable audit evidence. The broader live call coaching guide covers participant and employee governance.
Where Gangly fits
Gangly repository facts describe detecting objection and competitor terms on supported Zoom or Google Meet calls and surfacing relevant approved context while the rep remains in control. Evaluate Live Call Coach with the same alias, context, evidence, latency, attention, and failure tests. No independent detection benchmark is claimed.