AI governance

Can Staff Paste Customer Data Into ChatGPT? A One-Page Rule

Singapore's PDPC expects you to tell staff which personal data may go into AI tools. Use the traffic light below, then copy a one-page rule your team can follow on Monday.

Line drawing of a branching connector, a data cylinder, an orange key and an open padlock joined by thin lines.

TL;DRThe 6-minute read in 60 seconds

Only if the data and the tool match. Keep customer details out of personal chat accounts. Put only needed fields into your checked business tool. NRIC, bank and health data need your data protection officer's written yes.

  1. Singapore's data regulator wants you to tell staff which personal data may go into AI tools. It said so on 20 July 2026.
  2. One firm had no AI rule for staff. A script one worker wrote with AI showed 95,364 members' email addresses to other members.
  3. Sort your data into four classes, then check each paste against the tool that gets it. The rule below fits on one page.

Your staff can paste customer details into an AI tool only when the data and the tool match. Never a personal chat account; only needed fields in a checked business tool; NRIC, bank or health data only with your data protection officer's written approval. This is general information, not legal advice.

At a glance: why your staff need a written AI rule

20 JulPDPC: tell staff whichpersonal data may go intoAI tools (para 9.3, 2026)95,364people whose emails onestaff member's AI-writtenscript exposed (April 2026)3 daysto tell PDPC once a breachis assessed as notifiable;500+ people makes it so
20 JulPDPC: tell staff whichpersonal data may go intoAI tools (para 9.3, 2026)95,364people's emails exposedby one AI-written script(April 2026)3 daysto tell PDPC once a breachis notifiable; 500+ peoplemakes it notifiable
Sources: PDPC guidelines, undertaking notice and breach guide, linked in this post.

Can staff paste customer details into ChatGPT?

No for personal accounts. For business tools, only under your written rule: in 2026 Singapore's regulator asked businesses to tell staff which personal data may go in.

Timeline

Official positions on your staff's use of AI tools, 2026

Select a date. The Australian letter applies only to APRA-regulated firms.

  1. 30 Apr 2026

    Australia: policy alone falls short

    APRA review of large banks, insurers and super funds: staff AI use outside approved controls is a concern; many relied mainly on policy.

  2. 20 Jul 2026

    Singapore: tell staff what may go in

    You bear primary responsibility for AI systems you use (paragraph 9.1) and should tell your users which personal data to input (9.3). Advisory; the PDPA binds.

  3. 21 Sep 2026

    Singapore: no AI policy, 95,364 exposed

    Bee Cheng Hiang's undertaking: an AI-written script exposed member emails, and the firm had no staff policy on generative AI.

  4. 30 Sep 2026

    Singapore: the Shadow AI advisory

    On unapproved AI tools used to build scripts: keep an inventory of your AI tools, and tell your staff which data may go in.

This includes educating their end users, whether internal or external, on the specific types of personal data that should be input into their systems (e.g. only pre-defined categories of personal data) and how personal data will be processed following collection.

Personal Data Protection Commission, Advisory Guidelines on Use of Personal Data in Generative AI, paragraph 9.3, issued 20 July 2026.

For the regulator's side, see our PDPC guidelines explainer.

What counts as customer data?

Four classes cover almost everything your staff will paste; when your paste mixes classes, you use the highest.

Four data classes for your staff rule

Class What it covers Examples
Public Your business published it; nobody identifiable Your website prices, brochures, FAQs
Internal Your business information, no person identifiable Your draft proposal, process notes, sales totals
Personal Anyone identifiable, alone or with your other data Name, phone, email, order, chat log, photo
Regulated ID numbers, money, health, logins, vulnerable people NRIC, passport; bank, salary, medical details; logins
Source: PDPC Key Concepts guidelines (revised 29 April 2026), page 12 and paragraph 20.15, read 11 November 2026. Only Personal is a PDPA term; the other labels are ours.

Your Regulated class is stricter than the breach rules on purpose: treat ID, money, health, login and vulnerable-person data as high risk even where a leak might not be notifiable.

Which tool may take which data?

Each paste needs two answers from you: the data class and the tool that receives it.

Traffic-light classifier

Can you paste this into that AI tool?

Your files and screenshots count as pastes. Colours are our rule, built on PDPC sources; PDPC has not graded them. Nothing you enter leaves your browser.

Your traffic light by data class and tool type
Data classPersonal or free chatChecked business planIn-house, approved for this use
PublicGreenGreenGreen
InternalRedGreenGreen
PersonalRedAmber: needed fields onlyGreen: approved use, needed fields only
RegulatedRedRed, unless the DPO approves in writingAmber: only if approved for it

Not sure which tool you have, or nobody has checked your plan contract? Red until someone checks the contract, except for Public data.

Illustrative: to summarise 40 complaint emails, your coordinator pastes only complaint text and order dates into your checked business plan. That is still Personal, so Amber: strip names and phones first.

Where does a pasted customer record go?

Paste into a third-party AI tool and your record goes to the provider's servers, where it is processed, stored and, on some accounts, used for training.

Where a pasted customer record goes

Each step, then your control at that step (green).

1. PastedName, phone, order,complaint textCONTROLClassify it first:traffic light2. ProcessedOn the provider'sservers, maybe abroadCONTROLApproved toolunder a contract3. StoredChat history,logs, backupsCONTROLContract setsstorage, deletion4. Trained on?Only if the accountallows trainingCONTROLTraining off,in writing5. Sent onReply, email orscript for customersCONTROLSecond reader;code by a coderYour business answers for the tools it chose, at every step: PDPC guidelines para 9.1; PDPA s24.
1. PastedName, phone, order, complaint textClassify it first: traffic light2. ProcessedOn the provider's servers, maybe abroadApproved tool under a contract3. StoredChat history, logs, backupsContract sets storage, deletion4. Trained on?Only if the account allows trainingTraining off, in writing5. Sent onReply, email or script for customersSecond reader; code by a coderYou answer for the tools youchose: para 9.1; PDPA s24.
Our diagram, built from PDPC's generative AI guidelines (paras 7.1, 7.4, 7.5, 9.1) and the Bee Cheng Hiang undertaking.

What you get back can be code. Bee Cheng Hiang tested an AI-written email script only against activity logs; no customer data went into the AI tool, yet the code exposed emails. Its undertaking with PDPC commits it to independent review of such code (full case).

What does PDPC expect you to write down?

PDPC asks you for six safeguards; your rule covers the staff side, and you still own your tool list and contracts.

What PDPC asks for, and the line in your staff rule that answers it

PDPC asks for Where Your rule line
Tell your staff which personal data may go in Guidelines para 9.3 Your four classes and the traffic light
Written AI safeguard policies (encouraged) Para 9.4; advisory This one-page rule
An inventory of AI tools in use Advisory, 30 Sep 2026 Approved tools line, plus a list of tools in use
A set purpose for personal data Para 9.2; PDPA s18 In-house tools: approved use only
Reasonable security for prompts, outputs Para 9.3; PDPA s24 No personal accounts; checked plan contract
Technical review of AI-written code before it runs Advisory, 30 Sep 2026 Code review before any script runs
Sources, read 11 November 2026: PDPC generative AI guidelines; PDPC Shadow AI advisory; PDPA obligations.

In Australia, APRA's 30 April 2026 letter found many large firms relying "primarily on policy direction" rather than "enforceable technical restrictions". Our advice for any firm: pair your rule with a managed business account, and block personal AI accounts on work devices where you can.

Where Origin Pi stands

Fill in three boxes and copy your one-page rule; the colours match the classifier.

One-page staff rule

Customer and staff data in AI tools

Empty boxes keep their [BRACKETS]. First confirm your plan contract rules out training on inputs.

[COMPANY] staff rule: customer and staff data in AI tools

Approved tools: [TOOL] (our business plan, training on our inputs off) and our in-house tools with an owner and a set purpose. Personal chat accounts take PUBLIC data only.

If you are unsure which tool it is, or nobody checked its contract, treat it as a personal account.

Before you paste, pick the class. A mix or a doubt takes the higher class.

Files, screenshots and recordings count as pastes. AI note-takers only if approved and announced.

PUBLIC (we published it, it names nobody): personal account GREEN. [TOOL] GREEN. In-house GREEN.

INTERNAL (our business information, nobody identifiable): personal account RED. [TOOL] GREEN. In-house GREEN.

PERSONAL (anyone identifiable from our records, even masked: name, phone, email, order, chat, photo, voice): personal account RED. [TOOL] AMBER, needed fields only. In-house GREEN, approved use and needed fields only.

REGULATED (ID numbers, bank, card or salary, health or insurance, logins, children or vulnerable people): personal account RED. [TOOL] RED unless [OWNER] approves in writing. In-house AMBER, only if approved for this data.

A second person reads AI-written messages before customers see them. Someone who reads code reviews AI-written scripts before they run.

Pasted something wrong? Tell [OWNER] the same day: tool, time, what you pasted.

Questions: [OWNER], our data protection officer. We review this rule when our AI tools change.

Origin Pi's template, built on the PDPC sources above; PDPC has not approved it, and it is not legal advice. Nothing you type leaves your browser.

Origin Pi's AI governance work sets up the tool list and prompt logs behind this rule; AI compliance maps it to your PDPA duties. If a paste has gone wrong, use our AI incident response plan. For a second reader on your rule, talk to us.

Take the data with you

Cite this post if you reuse the data.

Frequently asked questions

Can employees put customer data into ChatGPT under the PDPA?
Our rule says no for personal or free accounts; the PDPA names no tools. It makes you protect the personal data you control (section 24), and PDPC calls unapproved staff AI tools Shadow AI. Use a tool you approved after checking its contract.
Is a business AI plan enough to paste customer data?
A business plan is the right place to start once someone checks it. Our check: the contract rules out training on your inputs and says where data is kept and deleted. Then tell your staff which fields may go in (PDPC paragraph 9.3).
Is pasting customer data into an AI tool a data breach?
Whether your paste is a breach depends on your tool and the facts. PDPC's breach guide requires you to assess it within 30 calendar days. If it is notifiable, for example because it affects 500 or more people, you must tell PDPC within 3 calendar days of deciding so.
Do I need a written AI policy for my staff?
PDPC encourages you to write AI safeguard policies (guidelines paragraph 9.4). The Bee Cheng Hiang undertaking of September 2026 records that the firm had no such policies for staff when an AI-written script exposed emails. Your first version fits on one page.
Does removing names make customer data safe to paste?
Only if nobody can link it back. The PDPA counts data as personal if a person can be identified from it alone or with other information your business has or is likely to have access to. Order numbers often identify a customer, so masked records usually stay Personal (PDPC Key Concepts guidelines, page 12).

Next step

Ready to make your business agent-ready?

20 minutes on your sector, your systems, and where this applies. No deck, no templates.

Talk to us →