Press on anything and hear where it came from
Accurate enough to draft, never to trust blind. Every field should show its sentence, so reps check it in seconds and their corrections become training data.
Jared Hawkins-Young · Founder, Lost Star TechnologyOctober 8, 20265 min readPart 7 of 13Key takeaways
- Reps adopt a system they can check. Every field should open to the signal that produced it and the timestamped sentence, with a play button.
- The model will misread things, so every output is built to be corrected. A correction stores the quote, the old value, the new value and who changed it.
- A correction made repeatedly becomes a schema change, reviewed by a person, so the next run gets it right. It is the only labelled training data the system gets for free.
- Measure accuracy in your own data: fields approved unedited, corrections per 100 extractions, and fields that survive to close unedited.
AI data entry from sales calls is accurate enough to draft and never accurate enough to trust blind, so every field it writes has to show the sentence it came from. The number that matters is not a vendor's accuracy claim. It is how many fields your reps approve unedited, and whether that number rises as their corrections feed back. A rep who can press a value and hear the buyer say it will trust the record. A rep who cannot will quietly route around it.
Why does traceability decide whether reps use it?
A sales team reacts badly to being told what to do by software, and the software is not always right. Those two facts together decide adoption. A rep who cannot check a field has two choices: trust it blind or ignore it. In practice that means ignoring it and keeping a spreadsheet on the side. A rep who can press the field and hear Dana say it has a third choice: check it in ten seconds and move on.
The starting point is low. Salesforce's 2024 State of Sales survey of 5,500 sales professionals found only 35% completely trust the accuracy of their organization's data. It is a vendor-run survey, so read it as a direction rather than a measurement. Adding an AI that writes more fields does not raise that number. Adding a source to every field can, because trust in a record is built one checked value at a time.
What does a rep see when they press a field?
Every value opens the same way: the field, the signal that produced it, how it was classified, and the moment in the recording, ready to play. The first card is an open issue from the ledger in part 5. The extraction routed it to Legal, because Dana tied the HubSpot scope question to the security review. Marcus listens to nine seconds and reroutes it to his solutions engineer. It was a scope question, not a legal one. It takes him four seconds, and the issue is kept rather than deleted.
The second card is the pain field. Marcus presses it, hears Dana say two deals slipped last quarter before anyone saw them, and marks the line for the proposal, word for word. Part 8 opens on it.
What happens to a correction?
- 1Rep corrects a row"Scope, not legal"
- 2Stored as a labelQuote, old value, new value, who
- 3Grouped into a patternThe same fix, made repeatedly
- 4Schema updatedThe next run gets it right
The model will misread things. The design assumes it will, so every output is built to be corrected, not just accepted or deleted. When Marcus reroutes the scope issue, the system stores the quote, the old value, the new value and who changed it. That is a labelled example, and nobody was asked to label anything. Marcus was going to fix the row anyway.
One correction changes one record. The same correction made again and again is a pattern: integration scope questions keep getting sent to Legal. A pattern becomes a schema change, which a person reviews before it ships, and the next call gets it right. A correction is the only labelled training data this system will ever get for free, and it arrives exactly where your deals differ from everyone else's.
Our view: a model that starts rough and learns from corrections beats a more accurate one that cannot be told it is wrong. And reps only keep correcting something they believe is listening, which is why the curve has to visibly fall. The same discipline runs through outbound part 12, where a drop in replies is traced through the data before anyone rewrites the message. A wrong field is usually a schema or a source problem, and the correction tells you which.
How accurate is AI CRM data entry, then?
A single accuracy figure is the wrong question. It depends on the audio, on how your buyers talk, and on how specific your fields are. Measure it in your own data instead, with numbers you can collect from the first week:
| Measure | What it tells you | What a bad number means |
|---|---|---|
| Fields approved unedited, per rep and field | Where the extraction already works | That field's definition is wrong for your deals |
| Corrections per 100 extractions, by week | Whether corrections are feeding back | Edits are not reaching the schema |
| Fields that survive to close unedited | The real accuracy figure | Values were approved on arrival but were wrong |
| Recommendations dismissed, per rep | Whether the arguments in part 3 hold | One rep: talk to them. Every rep: fix the rule |
The gap between fields filled without a person and fields that survive to close unedited is the honest number. Part 12 covers how to baseline it before anything ships, so the comparison flatters nobody. The dismissal rates come from the arguments in part 3.
Traceability also survives a resignation. When a rep leaves, the next one inherits a record they can press on. They hear the buyer, not a predecessor's summary of the buyer, and the account does not leave in anyone's head.
Who this is not for: a team whose recordings are deleted after a few days by policy. Without the audio the play button goes dark, and a field with a quote but no recording is weaker evidence. Keep at least the timestamped transcript for as long as your retention policy and your buyers' consent allow.
Next in the series
Part 8, The proposal writes itself, and every claim shows its source, applies the same mechanism to the document a rep sends: a proposal that opens on the buyer's sentence and can be defended line by line. The series page lists every part in order.
Common questions
- How accurate is AI at filling CRM fields from sales calls?
- Accurate enough to draft, not accurate enough to write without review. Accuracy depends on audio quality, how your buyers talk and how specific your fields are, so measure it in your own data: the share of fields reps approve unedited, and the share that survive to close unedited.
- What happens when a rep corrects a field?
- The correction is stored with the quote, the old value, the new value and who changed it. When the same correction recurs it becomes a pattern, a person reviews a schema change, and the next run gets it right.
- Do we need to keep call recordings for this to work?
- The play button needs the audio. Without it, a field can still show the transcript sentence and its timestamp, which is enough to check most values. Keep whichever your retention policy and your buyers' consent allow.
Sources
Written by Jared Hawkins-Young from the system he builds and runs. Diagrams are drawn from the same specs as the software. Published October 8, 2026. Illustrative numbers are labelled as such; every other figure carries its source.