# Our voice agent filled every CRM field correctly yet none of the payments went through — retained record

> Retained reddit record from reddit:AI_Agents.

Canonical: https://highsignal.app/data/records/ef5e2f5c4b15038d

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:AI_Agents)  published 2026-10-12 14:08 ingested 2026-10-13 02:36

[original source ↗](https://www.reddit.com/r/AI_Agents/comments/1x42unw/our_voice_agent_filled_every_crm_field_correctly/)

## retained text



```
Our voice agent was running collections calls with confidence scores above 90 percent on basically everything but then someone on ops pulled a random batch of CRM entries and listened to the actual recordings and it was BAD.

Agent captures an account number, gets one digit wrong, confidence is still high bc the model is confident about what it heard it just heard the wrong thing like customer says “fifteen twenty” meaning 1520 and the model logs it as two separate fields. Customer gives their name and the model just picks a common name that sounds close bc that’s what it does when it’s not sure, picks the most probable word not the right one.

and nothing downstream catches it bc the CRM entry looks fine. Payment bounces three days later bc the account number was wrong and nobody traces it back to the agent and we were literally celebrating field completion rate while half the fields were wrong lol.

now we run a validation layer between the agent and the CRM, IFSC codes get regex matched, account numbers get length and checksum checked, names get held if character level confidence drops below a threshold plus anything that fails goes to a human while the caller is still on the line instead of discovering it when the payment fails three days later.

This adds about a second per field though and on a collections call where someone’s already pissed that second is not nothing and we’re still only catching maybe 70 percent of bad entries bc some wrong values are valid formats that just happen to belong to someone else and no validation catches that.

we’re patching over something that should be fixed at the model level not after it, gnani and a few others do entity extraction inside the voice stack itself instead of transcribe then validate which would skip this entire mess but idk anyone doing it that way ??
```



document fetched 2026-10-13 00:19 · canonical https://reddit.com/r/AI_Agents/comments/1x42unw/our_voice_agent_filled_every_crm_field_correctly
