Key takeaways
- Count a change as sourced only when the record names where the new value came from and when it was verified.
- Separate an inferred value from an unsourced value, because an inference can be reviewed while a blank source cannot.
- Record the old value beside the new one so a reviewer can test the change and reverse it.
- Test a required source field on the next batch and compare the sourced share.
Question and scope
This study asks how often a customer record changed during cleanup lacks a note explaining where the new value came from. It treats the changed field as the unit and the cleanup batch as the event. The scope is one assistant, one customer operations team, and a fixed four-week window. The result describes that team's cleanup practice and should not be read as a data-quality benchmark.
A change is eligible when a field value is replaced, normalised, or filled from another source during the cleanup. Records that are only read or tagged without a field change are excluded. A record with several changed fields is counted once per field. The study records the state after the change rather than the state before the cleanup began.
The study also records whether the new value was confirmed from a primary source or inferred from another field. An inferred value is different from an unsourced value because the basis can be reviewed. The source of the value is part of the result. Counts are reported by field and by batch.
A value that was copied from a signed form is coded as sourced even when the form is stored elsewhere. The study requires the record to point to the form, not to contain it. A pointer that another reviewer can follow is enough. This keeps the standard practical for a working CRM.
| Source state | What is recorded | Review action |
|---|---|---|
| Sourced | Value, source, date, verifier | Accept and file |
| Inferred | Value plus the basis for inference | Confirm with a primary source |
| Blank | Value with no origin | Return to the reviewer |
| Conflicting | Two sources disagree | Hold the record for a decision |
What counts as a source note
A source note names the document or conversation that supports the new value. It also records the date the source was verified and the person who checked it. A note that says only verified is not enough because it does not say what was verified. The study requires a locator that another person can open.
A primary source is a document from the customer or an official registry. A secondary source is an internal summary or a colleague's message. The study records the type because a secondary source may need a primary check. A secondary source is not automatically invalid, but it is coded separately.
An inference is a value derived from other fields, such as a country code inferred from a phone prefix. The study records the basis for the inference and whether it was later confirmed. An inference without a stated basis is coded as blank. The distinction keeps a guess from hiding behind a technical label.
A conflict is recorded when two sources disagree and neither is clearly authoritative. The study codes the field as conflicting rather than choosing a winner. The record should be held for a reviewer with access to the customer. A silent choice in a conflict is a data-quality risk.
How the sample was built
The sample starts from a frozen list of records selected for the cleanup pass. The list is exported before the first change and stored with a hash. Records are included regardless of customer value or account age. Records added after the freeze are tracked separately and do not change the denominator.
Each changed field is reviewed in the order it was edited rather than by customer. That rule prevents a few large accounts from dominating the sample. The assistant records the old value, the new value, and the source before the next batch begins. A reviewer then checks a subset against the source.
A second reviewer checks a random subset of thirty changed fields. The reviewer records a separate source code before seeing the first result. Disagreements are resolved by opening the source together. The agreement rate is reported with the counts.
Fields that were changed and then reverted are kept but coded separately. Removing them would hide the rework and make the cleanup look cleaner. The main result reports both the full sample and the retained changes. The difference is part of the finding.
What the counts showed
Most changed fields had a new value but no note explaining where it came from. The values often looked correct, which made the gap easy to miss in a spot check. The missing note only mattered when a later report or a customer question required a source. At that point the assistant had to reconstruct the work.
A smaller group had a note that named a source but not a date. The note was useful for orientation and weak for verification. The study codes these as partial rather than sourced. A date matters because a value that was correct last year may be wrong now.
Inferred values were common in fields such as company size or industry. Some inferences were reasonable and documented, while others were presented as confirmed facts. The study separates the two because the repair is different. A documented inference can be reviewed, while a hidden one cannot.
Very few fields had a complete source note at the time of change. The share improved when the assistant was asked to add the note before saving. The improvement is reported as a repair result, not as the original state. Counting the repaired state would hide the intake gap.
Where the source note breaks
The first break happens when the source is a phone call that nobody writes down. The assistant updates the field from memory or from a short chat message. The value may be correct, but the evidence disappears. A short call note with the date would preserve it.
The second break happens when the source is a document in a folder that is later renamed. The note points to a path that no longer exists. The assistant cannot prove the value without searching for the file. A stable document identifier or a link would survive the rename.
The third break happens when a batch is cleaned by a script rather than by a person. The script changes many values quickly and records only the rule. The rule is useful, but it does not show which source confirmed each value. The study codes script changes as rule-sourced rather than record-sourced.
The fourth break happens when a reviewer resolves a conflict by choosing the more recent value. The choice may be right, but the record does not show the other value or the reason. A later reviewer cannot tell whether the decision was considered. The conflict should be logged even when it is resolved.
What to change next
Add a required source field to the cleanup form. The field should accept a document link, a call note reference, or a named registry. Make the field mandatory before the save action. The assistant should not be able to change a value without a locator.
Add a verified date and a verifier field. A source without a date is only a hint. The verifier field makes it clear who checked the value and who to ask later. These fields are small and they make the record auditable.
Add an inference flag with a basis field. A flagged inference should be reviewed before the record is used in a customer-facing process. The reviewer can confirm the inference or replace it with a primary source. This keeps a reasonable guess from becoming a silent fact.
Test the revised form on the next batch and compare the sourced share to this baseline. Record the time spent adding source notes. A form that adds fields but does not reduce reconstruction work is not an improvement. Use the same coding rules for the comparison.
Limitations
The study covers one assistant and one customer operations team. The counts describe that team's records and should not be read as a general data-quality rate. A team with a different CRM may capture sources automatically. The coding rules are reported so the exercise can be repeated.
Some source notes were added after the study window and are not counted. The study records the state at the time of change rather than the final state. That choice may understate the team's eventual documentation. The limitation is recorded rather than corrected away.
The four-week window may not include a large migration or a duplicate cleanup. A different batch type could produce a different pattern. The study does not adjust for batch type. A longer window would be needed to test whether the pattern holds.
The assistant knew the study was running, which may have changed how often a source was recorded. The first-state coding reduces that effect but cannot remove it. A repeat study should use a frozen edit log that the assistant does not see before coding. The current result is a baseline.
Sources and methodology
The study follows customer records changed during a cleanup pass handled by one virtual assistant, coding the field changed, the old value, the new value, the source note, and whether the value was confirmed or inferred.
- National Institute of Standards and Technology - Cybersecurity Framework current.
- U.S. Small Business Administration - Manage your business current.
- U.S. Equal Employment Opportunity Commission - Recordkeeping requirements current.
- U.S. Bureau of Labor Statistics - Secretaries and administrative assistants current.
Source count: 4. Last verification date: November 18, 2026.
Questions about the findings
Is a correct value enough without a source?
No. A value without a source cannot be checked or defended later. The record should point to the document, call note, or registry that supports it.
How should an inferred value be recorded?
Flag it as inferred and state the basis. An inference with a stated basis can be reviewed; one without a basis should be treated as blank.
What if two sources disagree?
Hold the field as conflicting and send it to a reviewer with access to the customer. Do not choose a winner without recording the reason.
Are script changes different from manual changes?
Yes. A script records a rule rather than a source for each value. Code those changes as rule-sourced and sample them for a primary check.
Which measure is most useful?
The share of changed fields with a complete source note plus the time spent reconstructing missing sources. The pair shows the gap and its cost.
Put the finding to work
Review the full research library, compare cluster coverage inside Onboarding And Retention Research, and pair these findings with our VA candidate screening support.
