Data Privacy Checklist

An operational checklist for handling data in AI tools. Week 1's version covered your personal accounts; this one covers working with other people's data.

Classify before you paste

Class	Examples	Where it may go
Public	Published content, public filings	Any tool
Internal	Drafts, plans, non-sensitive analysis	Approved tool, training disabled
Confidential	Contracts, financials, strategy, staff records	Enterprise tier with a DPA, or local model
Restricted	Customer personal data, health, biometric, children's data, credentials, regulated records	Only with specific written approval, or never

When unsure, treat it as one class higher.

Before pasting anything

• I know which class this is
• The tool and tier are approved for that class
• Training on conversations is disabled
• If it contains another person's data, I have a lawful basis and, where needed, their consent
• If it is a client's data, their contract permits this
• I have anonymised anything that does not need to be identifiable

The anonymisation routine

Takes a minute and changes the risk profile entirely.

Replace before pasting:
  Names           -> [PERSON_1], [PERSON_2]
  Emails          -> [EMAIL_1]
  Phone numbers   -> [PHONE_1]
  Addresses       -> [ADDRESS_1]
  Account/ID no.  -> [ID_1]
  Company names   -> [COMPANY_A]
  Dates of birth  -> [DOB_1]

Keep the mapping in a local file. Substitute real values back yourself.

Caution: anonymisation is a mitigation, not a magic word. A small dataset can be re-identifiable from combinations — a role, a location and a date can identify one person. For anything at volume, or anything sensitive, get advice rather than relying on find-and-replace.

Account hygiene

• Work accounts for work, personal for personal, never mixed
• Two-factor authentication on every account
• Training disabled, re-checked monthly (settings reset after some updates)
• Retention period known and recorded per tool
• Connected integrations reviewed; unused ones revoked
• Shared conversation links audited — several tools make these public to anyone with the URL
• Mobile app settings checked separately; they do not always sync

When personal data is involved

The obligations differ by jurisdiction, but the questions are the same everywhere.

• Lawful basis. Why are you allowed to process this at all?
• Purpose. Is this use compatible with why it was collected?
• Minimisation. Are you sending more than necessary?
• Processor agreement. Is there a DPA with the AI vendor?
• Sub-processors. Where does the vendor send it, and to which countries?
• Transfers. If data leaves your jurisdiction, is that covered?
• Rights. If the person asks for deletion, can you comply — including in the AI tool?
• Automated decisions. If AI materially affects a decision about them, do they have a right to human review and an explanation?
• Records. Is this processing documented in your register?

That eighth point catches people out most often. Using AI to screen applications is usually an automated decision with specific obligations attached, whatever your intentions.

Vendor assessment

Before approving any tool for Confidential or above:

Question	Answer
Does it train on our data by default?
Can that be disabled contractually, not just in settings?
Retention period
Where is data processed?
Sub-processors listed?
DPA available and signed?
Can we delete data on request?
Security certifications
Admin controls and access revocation
What happens to our data if we leave?

Incident response

If data went somewhere it should not:

1. Stop. Do not paste anything further into that tool.
2. Record exactly what was sent, when, by whom, and to which tool.
3. Delete the conversation in-app and at account level.
4. Rotate any credential that was exposed. Assume it is public.
5. Escalate to [DPO / legal / manager] the same day.
6. Assess whether it is notifiable. Many jurisdictions have tight deadlines — often 72 hours.
7. Fix the cause, not just the instance. Most repeat incidents are the same gap twice.

The monthly five minutes

• Privacy settings still as expected
• Conversations no longer needed, deleted
• Integrations reviewed
• Anything pasted this month that should not have been — adjust habits, not just settings
• Team following the same rules

The test

Before pasting: if this conversation were published tomorrow with my name and my employer's name on it, what happens?

If the honest answer involves a phone call you do not want to have, anonymise it or use a different tool.