Find Hidden Identifiers in a CRM Export

A practical AI workflow for checking a renewal handoff before any upload

Deleting the Contact and Email columns from a CRM export can leave the same details in a free-text note. A quick summary request may then carry customer information into an AI tool that was never approved to receive it. The useful first step is to inspect the actual copy you plan to share, cell by cell.

This is a constructed teaching case. Every account, row, status, and note below is fictional. No real CRM export or customer data was uploaded for this demonstration, and no separate live model response was captured.

The job: a short renewal handoff

Imagine a team needs a two-line handoff: which renewals still need an owner decision, grouped by month? Its fictional export has four rows. In the working copy, the only fields needed for that question are a temporary row label, renewal month, and decision status.

The original fictional export also has Contact, Email, and Notes fields. One Notes cell says, "Send update to contact@example.invalid after the owner decides." The example.invalid address is fictional; the .invalid domain is reserved as intentionally invalid. Deleting the Email column while keeping Notes would still copy that address into the material sent to the assistant. In a real export, a note could contain an actual address, phone number, customer name, or other identifying clue.

The note is not needed to count decision statuses. Neither are the Contact and Email fields. A copy built for this one handoff can omit all three. Look at every row of each field you keep; a column heading alone does not tell you what a free-text cell contains. Check hidden columns, comments, and attachments if they are part of the file you intend to send. A temporary label can itself link back to a person or account in your organization, so do not assume replacing names with codes makes real data anonymous.

Approval comes before minimization

A smaller extract is not automatically an approved upload. First confirm that this particular AI tool and this category of customer data are permitted under your organization's rules. If the answer is unclear, ask the data owner or use a workflow that keeps the data outside the tool. The Australian Cyber Security Centre and partner agencies explicitly advise against entering sensitive information into AI tools without approval. Their small-business guidance also recommends defining what data cannot be shared, reviewing the tool's data handling, and removing or changing personal details where appropriate.

A "do not train on my data" switch is not a substitute for that approval. Product settings, account ownership, retention, and administrator access can differ. Check the actual tool and workspace controls before sharing anything. Even a carefully reduced copy can remain sensitive or linkable. NIST's AI risk guidance treats privacy and reidentification as risks to assess, not problems solved by deleting one obvious field.

Copy this verification rule

Confirm tool and data approval before upload.

Keep needed fields; scan every retained note cell.

Codes are not proof of anonymity.

Check each claim; the owner approves the handoff.

Unclear? Stop and ask the data owner.

Work the example, then check the answer

If the tool and data were approved, the request could say: "Using only the four fictional rows below, list the rows marked Needs owner decision by renewal month. Do not invent a contact, next step, or completion status. Show the row labels that support each line." The prompt would contain only the three needed fields shown in the table, not the note or contact details.

For this constructed illustration, the checked answer is: "November: A1 needs an owner decision. December: A3 needs an owner decision." That is two rows in total, one in each month. You can verify it without AI: A1 and A3 carry the exact status; A2 and A4 do not. This demonstrates a checkable handoff, not a measured AI result or a privacy guarantee.

If an actual assistant adds "contact the customer," declares a renewal complete, or mentions a person absent from the permitted copy, remove the unsupported claim. OpenAI's own guidance warns that ChatGPT can give confident but incorrect answers; the person responsible for the handoff still needs to check the output against the allowed source and decide what can be sent.

Try this next: Before your next authorized AI-assisted handoff, write the one question the draft must answer. Make a separate working copy with only the fields needed for it, inspect every retained free-text cell, and confirm tool and data approval. Then check the draft line by line against that copy before anyone relies on it.

Sources

• Australian Cyber Security Centre, Artificial intelligence for small business, 14 January 2026.
• Australian Cyber Security Centre and partner agencies, Engaging with artificial intelligence, 24 January 2024.
• NIST AI Risk Management Framework Playbook, Measure 2.10.
• OpenAI, Data Controls in ChatGPT.
• OpenAI, Does ChatGPT tell the truth?.
• RFC 2606, Reserved Top Level DNS Names.

The CRM case and answer are constructed. No real customer records or live model output were used. Product settings and organizational policies must be checked for the tool you actually use.