Should you paste it at all?
Start with your company’s rules. Many employers name the AI tools staff may use and what may go into them, and an approved business account usually comes with terms that keep your data out of model training. On a personal account, look in the data controls for the setting that lets the provider use your chats to improve its models, and decide whether you want it on.
Some material should not go into an AI tool even with names removed: passwords and access keys, anything a contract or NDA forbids sharing, legal advice you received, and health or financial details about a person when your tool is not approved for them. Anonymizing is for the large middle ground, such as a customer complaint, a tricky email thread or meeting notes, where the facts matter and the identities do not.
What to replace
Two kinds of detail identify someone. Direct identifiers name them outright; indirect ones narrow it down until only one person fits.
| Kind | Examples | Replace with |
|---|---|---|
| Names | People, customers, companies, suppliers, project code names | [PERSON_1], [CLIENT_1], [SUPPLIER_1] |
| Contact details | Emails, phone numbers, street addresses | [EMAIL_1], [PHONE_1], [ADDRESS_1] |
| Numbers that identify | Account, card, policy, order, employee and ID numbers | [ACCOUNT_1], [ORDER_1] |
| Links and system details | Web addresses with IDs in them, IP addresses, internal URLs | [LINK_1] |
| Indirect details | A job title in a small team, an unusual event, an exact date with a place | A general version: “a team lead”, “last spring” |
Keep what the task needs. If you want help with a refund dispute, the amount, the dates and the policy matter; the customer’s name does not. If the numbers themselves are confidential and the task is about wording, round them or scale them.
Use consistent placeholders
Give each person or thing its own numbered placeholder, and use the same one every time it appears. Consistency is what lets the AI follow the story: it can tell that [PERSON_1] wrote the first email and [PERSON_2] replied, and its answer will use the same labels, which makes swapping the real names back easy.
- Put roles in the label when they matter: [MANAGER_1], [CUSTOMER_1].
- Keep pronouns if they help the text read naturally; drop them if they would help identify someone.
- Use labels that cannot be mistaken for real text, in capitals and brackets, so nothing is swapped back in the wrong place.
An example, before and after
Hi Priya, Daniel Okafor from Brightwell Logistics called again about invoice INV-20931. He says the 3,400 pallet charge was never agreed. Call him on 0161 496 0142 or daniel.okafor@brightwell.example.
Hi [PERSON_1], [PERSON_2] from [CLIENT_1] called again about invoice [INVOICE_1]. He says the 3,400 pallet charge was never agreed. Call him on [PHONE_1] or [EMAIL_1].
The amount stays because the task, drafting a reply about a disputed charge, depends on it. Everything that points to a real person or company is gone.
What it does not fix
Removing names lowers the risk; it does not remove it. A detailed story can still point to one person: “the only engineer in our Leeds office who resigned in March” needs no name. Read the anonymized text as a stranger would and ask whether they could guess who it is about.
Anonymizing also does nothing for content that is confidential in itself, such as an unreleased product plan or a pending deal. If the problem is the secret, a placeholder does not help, and the question goes back to whether this belongs in the tool at all.
Putting the details back
Keep the list of which placeholder stands for whom on your own computer, never in the chat. When the answer comes back, swap the real details in, then read the result once more: the AI may have added a placeholder of its own, or written “[PERSON_1]” where a greeting needs a first name.
Key takeaways
- Check your company’s rules and your account’s data settings before you paste anything.
- Replace names, contact details and identifying numbers with numbered placeholders, the same one every time.
- Keep the facts the task needs, such as amounts and dates; generalize the details that only narrow down who it is.
- Anonymized text can still identify someone through context, so read it as a stranger would.
- Keep the placeholder list off the chat, and swap the real details back yourself.