Privacy Pre-Flight: What Never Goes into a Prompt

Workflow area: Responsible Use · Tier: Foundation · Hands-on time: ~25 minutes

The most common AI incident in offices is not a hallucination — it is an employee pasting something confidential into a personal chatbot account because the deadline was close and the tool was open. The fix is not abstinence; almost every sensitive-shaped task has a safe version. The fix is a pre-flight: classify what you are holding, strip or fake what must not travel, then prompt freely. Ten seconds, every time, forever.

When to use — and when not to

• Use it before every paste into an external toolpersonal ChatGPT/Claude/Gemini accounts, free tiers, any tool your company has not approved for the data in your hand.
• Use it even inside enterprise tools for the highest tieran approved tenant protects data from the vendor, not from a bad idea; some content (below) should not be prompted anywhere.
• Use it when the task "needs" the sensitive partit almost never does. The drill's core skill is separating the shape of the problem, which AI helps with, from the specifics, which it does not need.
• Relax it for genuinely public contentpublished material, generic how-to questions and your own non-confidential drafts need no ceremony. The drill exists so you can be fast where it is safe.
• Never "just this once"deadline pressure is precisely when incidents happen. The safe version costs one minute; make it the reflex.

The recipe

1. 1.

   Learn the four never-paste categories. Commit these to memory — they cover nearly every office incident:

   • •
     People data — customer or employee names tied to anything sensitive: complaints, salaries, health, performance, ID numbers, contact details. Regulated (DPDP in India, GDPR for EU data) and the fastest way to turn a paste into a reportable breach.
   • •
     Money before disclosure — unpublished financials, forecasts, pricing under negotiation, deal terms, anything that moves a decision or a market if it leaks early.
   • •
     Keys and credentials — passwords, API keys, OTPs, account numbers, security configurations. No tool, no exception, not even "approved" ones; credentials go in vaults, never prompts.
   • •
     Other people's confidential material — client documents under NDA, partner data, vendor contracts, a colleague's draft shared in confidence. Not yours to paste; the NDA does not have an "but it was for a summary" clause.
2. 2.

   Run the ten-second triage on what you are holding. Three questions, in order: Is any of the four categories present? If no — proceed. If yes — does my company have an approved tool for this data class (M365 Copilot or Gemini for Workspace in the company tenant, an internal AI gateway)? Use that. If no approved tool — redact or synthesize (steps 3–4), or do the task without AI. That is the entire decision tree.

3. 3.

   Redact: same text, sensitive parts replaced. For tasks on a real document — summarize this complaint, draft a reply to this email — swap identifiers for consistent labels before pasting:

   Copy

   Original (never paste):
   "Ramesh Gupta (a/c 4402-8871, phone 98xxx) is threatening to
   escalate to the ombudsman over the ₹42,000 double-debit on his
   home-loan EMI on July 3."
   
   Redacted (safe to paste):
   "[CUSTOMER] ([ACCOUNT], [PHONE]) is threatening to escalate to
   the ombudsman over a [AMOUNT] double-debit on his loan EMI on
   [DATE]."
   
   Then prompt: "Draft an empathetic first response to this
   complaint. Keep the [PLACEHOLDERS] where personal details
   belong — I will fill them after."
   

   Keep labels consistent ([CUSTOMER] every time, not [NAME] then [CLIENT]) so the model tracks who is who — the reply comes back with placeholders positioned exactly where you re-insert the real details, outside the tool.

4. 4.

   Synthesize: same shape, fake specifics. For tasks about a pattern rather than a document — help with a salary-band analysis, a tricky HR situation, a formula over financial data — build a structurally identical fake:

   Copy

   "I have a spreadsheet of 40 employees with columns: level
   (L1-L5), department, current salary, proposed salary. I need a
   formula flagging anyone whose proposed raise exceeds 15% or
   whose new salary crosses their level's band maximum (bands in
   a second sheet). Here are 5 rows of FAKE data in the real
   structure: ..."
   

   The model solves the structure; the real numbers never leave your workbook. This one technique safely unlocks most "but my task IS the sensitive data" cases.

5. 5.

   Know the no-prompt zone, and make the drill visible. Some matters stay human even with an approved tool and redaction: active legal disputes, disciplinary and termination decisions, anything under regulator or auditor scrutiny, M&A. For everything else — run the triage aloud with your team for a week ("category 1, using the tenant tool", "redacting, then pasting") until it is muscle memory, and put the four categories on a sticky note where the pasting happens.

Make it yours

• ChatGPT / Claude personal accounts: assume anything pasted leaves your control — check your settings (training opt-outs, temporary chats) but behave as if the postcard rule applies regardless; settings change, the habit should not.
• M365 Copilot: commercial data protection means prompts stay in your tenant and are not used to train models — which safely upgrades categories 2-lite and internal drafts, but does not repeal category 3 (credentials) or the no-prompt zone, and Copilot sees whatever you can open, so over-shared files are now over-summarized files.
• Gemini for Workspace: same logic — enterprise terms cover Workspace data, personal Gmail-account Gemini does not. The account you are logged into is the privacy decision.
• No policy at your company? Adopt this drill as your personal floor, and send your manager the four categories as a proposed team norm — being the person who introduced the safety habit is a better look than being the person in the incident report.

Quality check

• You can recite the four categoriespeople data, undisclosed money, credentials, others' confidential material — without the page.
• The triage happens before the pastetool choice and redaction are decided in the ten seconds before Ctrl+V, not regretted after.
• Redactions are consistent and reversiblestable placeholders, real values re-inserted only outside the AI tool.
• Synthetic data matches structure, not valuesclose enough that the answer transfers, fake enough that a leak means nothing.
• The no-prompt zone is respectedlegal, disciplinary, regulated and M&A matters stay with humans regardless of tooling.