Choose one of three approaches. The steps for that option appear below.
Choose an option above. On a phone, swipe the tabs left or right to see all options →
No AI: one person redacts, another checks.
Make a working copy
Redact by hand
Independent read
Decide on the copy
Use: You
Have ready: the original transcript. Make a working copy and list what counts as personal data.
Original untouched.
Use: You
Redact every listed type in the copy by hand, using labels.
Every listed type handled.
Use: The second person
Read the redacted copy independently and search for @ and long digit runs.
Cannot identify the customer.
Use: The support lead
Read it start to finish once more; give a clear go/no-go note for sharing internally.
Check nothing has been sent or published (Option A has no 'still to confirm' marker anywhere on the page).
The technical colleague runs Presidio on a local copy; its data types supply the starting list. You read the result.
Make a working copy
Run Presidio locally
Read and mark misses
Independent read
Decide on the copy
Use: You
Make a working copy of the original transcript and swap the customer's name and contact details for codes by hand.
Original untouched; the customer's name and contact details are swapped for codes; only the copy goes forward.
Use: The technical colleague
Run Presidio on the copy on a local machine, replacing found items with labels; use its personal-data types as the starting list, noting uncovered types.
Covered types match names, contacts, numbers, addresses, payments. The transcript stays on that machine; nothing is uploaded or sent out.
What to paste into the tool
Instructions for the technical colleague, not a prompt to paste into Presidio: Verify the installed setup, supported languages and personal-data types, and whether any component sends text outside the local machine. If offline handling cannot be confirmed, stop before loading the transcript. Configure detection and replacement of found items with labels such as [NAME]. Run on the working copy only, leave other text unchanged, and list all unsupported or unchecked types.
Use: You
Read the whole redacted copy against the original and mark anything Presidio missed.
No personal item remains; meaning still clear.
Use: The second person
Read the corrected copy independently.
Cannot identify the customer from the copy.
Use: The support lead
Read it start to finish once more; give a clear go/no-go note for sharing internally.
Check nothing has been sent or published, and anything not known yet is still marked.
One message per row; Google Workspace with Gemini (Docs, Sheets, Slides) flags rows likely to hold personal data.
Check permission, build sheet
Flag risky rows
Redact flagged rows
Independent read
Decide on the copy
Use: You
Have ready: transcript and the owner's confirmed permission for Google Sheets and Gemini; without it, stop. Copy into a sheet, one message per row.
Permission confirmed before upload, original unchanged, every message appears exactly once. Obvious names, contact details and account numbers replaced by placeholders before Gemini.
Use: Google Workspace with Gemini (Docs, Sheets, Slides)
Flag each row yes, no or unsure for personal data and name the type.
What to paste into the tool
For each message row below, return its row number, a yes/no/unsure flag for personal data, and the data type. Look for names, contact details, order or account numbers, addresses, payment details, and other details that could identify the customer. Do not repeat message text or personal values. If uncertain, use unsure. Message rows from the owner-approved labelled working copy, where obvious names and contacts are already replaced by neutral labels: [paste rows with row numbers]
Use: You
Redact every yes and unsure row by hand, and scan the no rows.
No personal item remains in any row, including no rows.
Use: The second person
Read the redacted sheet independently.
Cannot identify the customer.
Use: The support lead
Read it start to finish once more; give a clear go/no-go note for sharing internally.
Check nothing has been sent or published, and anything not known yet is still marked.
Best when: One short transcript, or no tool is approved
What it does: AI does nothing; you do it all
Strong at: No text goes to an AI tool
Limit: Slow; misses are likely on long text
Free plan: No AI tool, so no plan needed
Best when: Many transcripts and a technical colleague is available
What it does: Presidio detects and labels; you read the whole copy
Strong at: Detects and replaces personal data on your own machine
Limit: Detection may miss items; needs setup by someone technical
Free plan: Yes (open source, MIT Licence)
Best when: Long transcripts and a Workspace tool is already approved
What it does: Google Workspace with Gemini (Docs, Sheets, Slides) flags rows; you redact them
Strong at: Categorises table data row by row
Limit: Sends text to an online tool; needs the owner's approval first
Free plan: Paid plan needed (Business Standard or higher)
Before you paste into any AI tool, remove any confidential information. Put it back at the end.
Search the copy for @, digit runs and common name words.
Original stays unchanged and unshared.
Check the result against your original inputs. Approve it before sharing or acting.
Plan, price, language coverage and data-handling terms for each tool are not listed on the pages we checked.
Presidio was read from its GitHub repository page only; the older documentation site page said "This page has moved" and was not used.