## TL;DR
Never let a raw transcript write directly to the CRM. Require a confidence check on the transcription, flag low-confidence and negation-heavy segments for review, and spot-check high-stakes fields like intent and plan interest before they drive follow-ups. One flipped "not" can reroute a whole sequence.

```
agent transcribed a call wrong and logged 'interested in enterprise plan'  -  the lead actually said 'not interested in any plan'
```

## Steps

1. Put a verification layer between the transcript and the CRM: extract the call outcome (interested, not interested, follow-up date, plan mentioned) as structured fields, each with the transcript's confidence score attached.
   Expected: every logged outcome carries a confidence value, and you can see which calls were borderline.

2. Flag for human review any call where the outcome hinges on a negation ("not interested", "do not call back", "no budget") or where confidence falls below your threshold.
   Expected: negation-bearing calls land in a review queue instead of auto-logging, with the relevant sentence highlighted.

3. Spot-check a weekly sample of auto-logged calls by listening to the audio for the key 30 seconds, and track the error rate.
   Expected: a running accuracy log exists, and a rising error rate triggers a review of the transcription setup.

4. Fix the records that are already wrong: find calls logged as positive where the transcript contains negations near the intent sentence, and correct the dispositions.
   Expected: a cleanup pass corrects the flipped records, and the affected leads are pulled out of the wrong follow-ups.

5. Improve the input: check microphone and call-audio quality for the worst-offending reps or lines, since bad audio is the most common cause of dropped negations.
   Expected: the lowest-quality audio sources are identified, and error rates fall after the fix.

## Use this when
- transcripts mishear negations and flip the call outcome
- CRM dispositions contradict what the lead actually said
- transcription feeds follow-up logic with no verification step

## Not for this skill when
- the transcript is right but the agent chose the wrong disposition anyway - that is a judgment/prompt problem
- the call never connected or the recording failed - that is a telephony problem
- you need transcripts for coaching, not for driving automation - lower stakes, lighter checks

## Variant phrasings
- transcription dropped the "not" and logged false interest
- call outcome flipped by a misheard negation
- agent logged enterprise interest the lead never expressed
- transcript-to-CRM pipeline with no verification

## Why it happens
Transcription models are weakest exactly where it matters most: short negation words in noisy call audio. "Not interested in any plan" compresses easily into "interested in enterprise plan" when the "not" gets swallowed. And because the pipeline treats the transcript as ground truth, the error compounds - the CRM, the sequence, and the AE all act on the flipped version.

## Edge cases
- Accents, crosstalk, and speaker overlap raise error rates - weight review toward calls with those conditions.
- The lead says "not right now" (timing) vs "not interested" (verdict) - the extraction must distinguish deferral from rejection, since the follow-up differs completely.
- Corrected dispositions should also fix downstream automation - a lead moved to "not interested" must exit the active sequence, not just carry a new label.
- Recording consent still applies - none of this transcription pipeline matters if the call was not supposed to be recorded; see the consent skill.

## Provenance

Resolved from the public thread: https://vectle.com/posts/pst_WngGA_LvzlxABFotmSUciA