How to keep card numbers, IDs and health details out of chat transcripts: practical redaction steps, plus what AI SNS CHAT does and does not do.
Customers type far more into a chat window than you ask for: card numbers, home addresses, ID photos, medical details. Once those messages sit in a transcript, they become data you have to protect, search, export and eventually delete. This guide explains where personal data hides in website chat, WhatsApp and Messenger conversations, and how to redact it before, during and after the conversation. It also states plainly what AI SNS CHAT does and does not do, so you can plan around the gaps.
Why chat transcripts become a privacy problem
A support chat feels informal, so people treat it like a conversation with a friend. They paste an order confirmation with a full address, photograph a passport "to prove it's me", or explain a diagnosis to justify a refund. Nobody asked for any of it, but it is now stored.
Three things make chat riskier than a web form:
- Free text has no boundaries. A form field limits what can be entered. A chat box accepts anything.
- Transcripts travel. They get exported to spreadsheets, forwarded to colleagues, pasted into tickets and used as training examples.
- Retention is often invisible. Many teams cannot say how long conversations are kept or who can open them.
The goal is not to make customers stop sharing. It is to make sure sensitive data is not collected needlessly, is masked where it must be seen, and is deleted when it is no longer needed.
What counts as personal data in a chat
Different regulations draw the lines differently, so treat this as a working list rather than a legal definition. Your own legal counsel decides what applies to you.
| Category | Typical examples in chat | Usual handling |
|---|---|---|
| Direct identifiers | Full name, email, phone, home address | Collect only when needed; mask in exports |
| Financial | Card numbers, bank details, CVV | Never collect in chat; redirect to a hosted payment page |
| Government IDs | Passport, national ID, tax numbers | Never collect in chat; use a verified secure process |
| Health and sensitive | Diagnoses, medications, religion, union membership | Do not request; escalate to a trained human |
| Credentials | Passwords, one-time codes, API keys | Never collect; warn the customer and mask immediately |
| Indirect identifiers | Order numbers, device IDs, IP addresses | Treat as personal when combined with other data |
Three moments to redact: before, during, after
Redaction is not one switch. It happens at three moments, and a sound policy covers all three.
- 01Prevent: do not ask
- 02Mask: hide what appears
- 03Limit: restrict who sees it
- 04Delete: remove on schedule
- Prevent. Design the conversation so sensitive data is never requested.
- Mask. When it appears anyway, hide it in the places people and systems read it.
- Limit. Control who can open transcripts, export them or train on them.
- Delete. Remove conversations when the retention reason has expired.
Most of the value comes from the first step. Data you never collect cannot leak.
Prevent first: design the conversation not to collect it
Write the guardrail into the agent's instructions
An AI agent follows its persona and guardrails, so use them. A short, specific rule works better than a long policy. For example:
- "Never ask for card numbers, passwords, one-time codes or government ID numbers."
- "If a customer types any of these, do not repeat them. Tell them not to share this in chat and explain the secure alternative."
- "For health or legal matters, offer a human handoff instead of discussing details."
Repeating the rule back to the customer matters. A line such as "Please don't send card details here; I'll send you our secure payment link" teaches the habit and shows you care.
Ask for the minimum, and say why
If you need an email address to follow up, ask for the email address, not "your contact details". Define each field and the moment it is requested. A contact field described as "ask for the phone number only when the customer requests a callback" prevents the agent from collecting it as a reflex.
Move sensitive steps out of the chat
Payments, identity verification and anything involving documents belong on a hosted page or in a controlled human process. The chat's job is to point the customer there, not to carry the data.
Mask what still slips through
Even with good design, customers paste things. Typical patterns worth masking in any transcript view or export:
- Sequences of 13 to 19 digits that look like card numbers
- Email addresses and phone numbers
- Street addresses and postcodes
- ID-style strings (passport, national or tax numbers)
- Tokens, keys and passwords
Two practical rules apply to masking:
- Mask in the copy people read, not just the copy machines store. If an exported spreadsheet contains the full number, the problem has only moved.
- Prefer irreversible masking for secrets. A password or card number has no legitimate reason to be recovered later.
A common failure is masking the display but leaving the original in logs, backups or analytics. Ask where else the text goes before declaring a job done.
Retention, access and deletion
Redaction is easier when there is less data to redact. Retention is the lever many teams forget.
Decide how long you keep chats
Pick a retention period tied to a real purpose: dispute windows, warranty periods, quality review. "Forever" is a choice with a cost, not a neutral default. As a hypothetical example, if a shop receives 300 chats a month and keeps them indefinitely, it holds 3,600 transcripts after one year, each a possible exposure, even though the useful life of most may be weeks.
Limit who can open transcripts
Give each team member the least access that lets them work. A viewer who reviews analytics does not need to read raw conversations containing addresses. Review the member list quarterly and remove people who changed roles.
Keep a record of who changed what
An activity trail showing who created, edited, deleted or uploaded something makes incident reviews faster and discourages casual copying.
Have a deletion routine
When a customer asks you to erase their conversation, you need to know where it lives: the chat log, any exported file, any lead or contact record created from it, and any note or example derived from it. Write that list down once so every request follows the same steps.
A quick audit routine you can run monthly
- Search recent transcripts for digit strings, "password", "passport" and "ID".
- Check the answers your agent gave when those appeared. Did it repeat the data or redirect?
- Open one exported file and confirm it matches what you intended to share.
- Review who has access to chat logs and exports.
- Confirm the retention setting still matches your written policy.
- Update the agent's guardrail wording if a new pattern appeared.
Thirty minutes a month finds most drift before it becomes an incident.
Doing this in AI SNS CHAT
It is worth being precise here. The product facts do not describe a customer-facing switch that automatically masks every personal detail a visitor types into a conversation. What AI SNS CHAT does provide are the controls around that gap, plus some protections built into the platform.
Built into the platform
- Knowledge-base passages that could be shown to strangers are scrubbed of personal data, and secrets are redacted from the platform's logs.
- Secrets are encrypted with AES-256-GCM and passwords are hashed with bcrypt.
- Per-agent prompt-injection and abuse filters sit in the agent's security page.
- Practices are designed with reference to GDPR and APPI. The site states clearly that it is not SOC 2, ISO 27001 or HIPAA certified, and it lists PrivacyMark No. 20002676. See /security for the current wording.
- Customers sign a My Number (特定個人情報) memorandum before creating an agent.
Controls you configure
- Persona and guardrails. Open the agent and add the "never ask for, never repeat" rules described above. The agent's instructions are where you tell it to redirect card, ID and credential data to a secure path.
- Custom attributes and forms. Each contact field has a description that tells the agent when to ask for it. Write those descriptions narrowly so the agent collects only what you need. Leads from the "Collect leads" Action appear in Activity → Leads and create a contact, so they are part of your deletion checklist.
- Human handoff. For health, legal or identity matters, route to a person using Helpdesk handoff, which needs Standard or higher.
- Chat logs. In Activity → Chat logs you can filter, flag, bulk select, resolve or delete conversations, and export them. Use bulk delete for transcripts that contain data you should not hold, and remember that exports are copies you now manage.
- Roles and permissions. Owner, Co-owner, Member, Support and Viewer roles, with per-member overrides, let you keep raw conversations away from people who do not need them.
- Activity log. Workspace settings → Activity, for owners and admins, records every create, edit, delete or upload with name, email, feature and time.
- Retention. A conversation retention window exists as an opt-in limit. The default is to keep conversations forever, so confirm what is set for your account rather than assuming.
- Account security. Turn on two-factor authentication (TOTP). Admins can require it, and repeated failed logins lock an account for a period.
If automatic masking of visitor messages is a hard requirement, plan to cover it with guardrails, short retention and review of the chat logs rather than relying on a feature that is not documented. Plans and limits are on /pricing, and the full feature list is on /features.
Frequently asked questions
Does AI SNS CHAT automatically mask card numbers that customers type?
Not as a documented customer setting. The platform scrubs personal data from knowledge-base passages shown to strangers and redacts secrets from its logs. For visitor messages, use agent guardrails, short retention and deletion in Chat logs.
How do I stop an agent from asking for sensitive details?
Add explicit rules to its persona and guardrails, and write narrow descriptions for contact fields so it asks only when needed. Then test the behaviour in Playground with realistic customer messages before going live.
Can I delete a conversation that contains personal data?
Yes. In Activity → Chat logs you can select one or many conversations and delete them. Check for related leads, contacts and exports too, because those are separate copies.
Is AI SNS CHAT certified for GDPR or HIPAA?
The site says its practices are designed with reference to GDPR and APPI, and it holds PrivacyMark No. 20002676. It states that it is not SOC 2, ISO 27001 or HIPAA certified, so check the details on /security against your own obligations.


