Most tasks AI is good at, such as summarizing support conversations, organizing order notes and classifying inquiries, involve customer information. If names, phone numbers, addresses and account numbers go straight to an external AI service, reviews such as outsourcing of processing or cross-border transfer follow, and the original text can easily end up somewhere in the logs. That's why it's safer to make "mask before sending" the default when designing AI features.
Why mask "before sending"?
In August 2025, Korea's Personal Information Protection Commission (PIPC) published guidance on processing personal information for developing and using generative AI. It covers the privacy issues, legal standards and safeguards to look at in each stage of the generative AI lifecycle. In May 2026, a privacy guide for users of generative AI services followed, covering key questions such as whether your input is used for training and what to watch out for when entering work materials.
The direction running through both is simple: don't pass along personal information you don't need in the first place. The AI doesn't need a customer's real phone number to summarize a support conversation.
Mask in 3 steps: rules, lists, human review
- Use rules (regular expressions) for values with a fixed shape: values with a consistent format, such as phone numbers, emails and account numbers, are found by pattern and replaced with tags such as
[PHONE1]and[EMAIL1]. (Regular expression: a way to find text by its character pattern.) - Use lists for names and addresses: names and addresses vary in shape, so rules alone miss them. Match them against the customer name and shipping address lists already in your order data and replace them with
[CUSTOMER1]and[ADDRESS1]. - Human review: when handling a new type of document for the first time, have a person look at a few masked results and add any missed patterns to the rules.
| Value | How to mask | Replacement tag |
|---|---|---|
| Phone number, email | Rules (regex) | [PHONE1], [EMAIL1] |
| Account, card number | Rules + digit check | [ACCOUNT1] |
| Name, address | Match against customer list | [CUSTOMER1], [ADDRESS1] |
How to restore values after the answer comes back
Keep the pairs of masked and original values (the mapping table) only on your own server. The AI returns an answer such as Please notify [CUSTOMER1] of the shipping delay, and the original name is restored from the mapping table only right before it's shown on screen. This way, only tags remain on the external AI side.
Delete the mapping table once a request finishes or set a short retention period, and keep the people and systems that can access it to a minimum.
Want to check this against your own data?AI Readiness Assessment · we review your material and map candidates and orderContact usDon't leave the original text in logs either
The place most often missed is logs. If you store full AI requests for debugging, the original text before masking can pile up as is.
- Store only the masked text in AI request and response logs
- Don't put entire request bodies in error logs
- Check the data retention and training policies of the AI service before signing a contract
- Review, together with your privacy policy, whether using an external AI service counts as outsourced processing or cross-border transfer
References
- PIPC and Privacy Portal, Guidance on processing personal information for developing and using generative AI (Aug 2025): https://www.privacy.go.kr/front/bbs/bbsView.do?bbsNo=BBSMSTR_000000000049&bbscttNo=20836
- PIPC, Privacy guide for users of generative AI services (May 2026): https://www.pipc.go.kr/np/cop/bbs/selectBoardArticle.do?bbsId=BS074&mCode=C020010000&nttId=12084
We review your current system and data, then outline the migration order and scope.
We build web apps, mobile apps and business systems, then add and run the AI they need. We build and operate our own meeting notes, contract and company analysis AI.
See our projects