A model reads a call transcript and proposes CRM updates. Some are right, some are invented, and some are plausible but wrong. Configure the guardrails, move cards between lanes, then reveal which updates were wrong and see what actually reached the CRM.
Fictional calls, accounts and fields. The AI output is a fixed mock, not a live model, and human review is assumed to catch errors. The scenarios are composite and built to show failure modes, so the counts illustrate trade-offs and are not measurements of any real system.
Call
Guardrails
Call transcript
Click an update to see where its evidence appears.
Guardrails
AI-proposed CRM updates
What reached the CRM
0Updates applied automatically
?Wrong updates written to the CRM
?Correct updates held back
Reveal the ground truth to see how many wrong updates slipped through and how many correct ones were held back.
Ground every value in evidence
A value without a verifiable quote is a guess. Verifying the quote against the transcript catches invented values, at the cost of blocking honest paraphrases.
Plausible is the dangerous case
The worst errors cite real quotes and still draw the wrong conclusion. Only field policy and human review catch those.
Every guardrail has a price
Stricter rules hold back more correct updates. Design the thresholds against the cost of a wrong value in each field.
In a real environment
Data sources
Call transcripts can be requested from Gong by call IDs or by date range.
Field types and picklist values for the schema check come from Salesforce describe, filtering out inactive values.
Protected fields, overwrite policy and thresholds are decisions of the organization.
Limits and prerequisites
Gong API credentials are issued by a technical administrator.
Gong documents default limits of 3 calls per second and 10,000 calls per day for the company.
Salesforce API requests are pooled per org within a 24-hour period. For Enterprise Edition the allowance is 100,000 plus a per-license amount and any purchased add-ons, and exceeding it returns HTTP 403 REQUEST_LIMIT_EXCEEDED.
The model in this demo is a fixed mock. A real version needs a language model that returns structured output, and its cost and latency must be sized for the call volume.
How it would be validated
Ground truth in this demo is known by construction. In a real org it comes from people reviewing a sampled set of proposed updates.
Track the two numbers shown here: wrong updates that reached the CRM and correct updates that were held back.
Strict quote matching would also block honest paraphrases and transcription errors. Test how often before choosing it.
What this demo does not claim: the counts are produced by three fictional calls and do not predict the behavior of a real model.