Workflow automation

Staff pasting customer data into AI tools: an educational starter policy

Workflow automationBy Fiona Wang8 November 20269 min read
Illustration for Singapore B2B provider matching and regulatory benchmarks
Public sources cited 2026 S$ Fee Benchmarks 1–5 Free Quotes
Illustration of an open book whose pages turn into signal waves

In short: This is a starter your leadership must adapt and then give to a qualified data protection officer or legal adviser before anyone adopts it. It is not an official PDPC document, it has not been professionally reviewed, and using it does not make an AI tool or your company PDPA-compliant.

A self-described e-commerce business asked how companies with compliance or privacy duties were handling staff who paste customer data, pricing and strategy into ChatGPT, and whether anyone had set rules, switched off access, or added other guardrails. That is one company’s question, unverified, not a Singapore survey. It is a good question. The answer is not a slogan that a tool is “PDPA compliant”.

Read this before the template. This page is an educational starter for discussion. It is not legal advice. It is not an official publication of the PDPC or IMDA. It has not been reviewed by a lawyer or a data protection officer for your organisation, and a model rereading it is not that review. Do not adopt it as your policy until you have adapted it and a qualified DPO or legal adviser has reviewed the version you mean to use. Following it does not mean you comply with the PDPA, and it does not make any AI product compliant. FYAdvisers is not a law firm.

Two kinds of harm get mixed together. One is confidentiality of prices and strategy, which is a business problem even when no personal data is involved. The other is personal data, which sits under the PDPA when the Act applies. A rule that only says “don’t paste secrets” misses identity numbers and customer mail. A rule that only recites the PDPA misses the price list. Write both.

Technology & cybersecurity illustration for Staff pasting customer data into AI tools
1Tell us the needScope, budget, timing2Concierge callWe confirm requirements3Registry checksACRA, licences, fit41–5 introductionsCompare quotes free5You decideDecline any; nothing owed
How FYAdvisers matching works
Checklist before you hireEntity is live on ACRA BizFile+ (check the UEN)Certifications that fit the job: CREST/OSCP testers, CSA, ISO 27001Named lead and who actually deliversWritten scope, deliverables and retest or support termsPDPA handling, NDA and data locationFixed price vs day-rate, and change-request rules
Singapore B2B Vetting & Concierge WorkflowHow we verify and match 1 to 5 providers for your brief
01
Registry & Licence CheckACRA, MAS, CSRO, BCA, MOM & ISO status verified
02
Scope & 2026 Fee AuditItemised S$ benchmarks with hidden traps flagged
03
1 to 5 Vetted QuotesCompare proposals from matched Singapore firms
04
100% Free for BuyersZero obligation; providers fund the service

What you can rely on, and what you cannot

These points were checked against primary or vendor pages. They are narrow.

Someone must be responsible. Section 11(3) of the Personal Data Protection Act 2012 requires an organisation to designate one or more individuals responsible for ensuring the organisation complies with the Act. They may delegate (section 11(4)). Business contact information of at least one such person must be available to the public (section 11(5)). Designating them does not remove the organisation’s own duties (section 11(6)). How to buy that help in is discussed in the outsourced DPO guide. Naming a person is not the same as having a policy.

A vendor may be a data intermediary, and may not. The PDPC’s Guide to Managing Data Intermediaries (2020) discusses organisations that process personal data on another organisation’s behalf under contract. It puts responsibility on the organisation that decides the purposes to set the scope and the security requirements, and it expects the contract to say what the obligations are. Whether a given AI provider is a data intermediary depends on that arrangement. This article does not decide it for any product.

The AI advisory guidelines have a scope. The PDPC’s Advisory Guidelines on the use of personal data in AI recommendation and decision systems, dated 1 March 2024, are about AI systems that recommend or decide using personal data. They are not a general permission or ban on typing into a chat box. Read them if a model will score, rank or decide about a person. Do not stretch them to every experiment, and do not ignore them when the experiment is actually a decision.

OpenAI’s business page is not every ChatGPT plan. OpenAI’s business-data page, checked for this article, says that by default OpenAI does not use data from ChatGPT Enterprise, ChatGPT Business, ChatGPT Edu, ChatGPT for Healthcare, ChatGPT for Teachers, or the API, including inputs and outputs, to train or improve models, unless you opt in. It says qualifying organisations can configure retention, including a zero-retention option on the API. It lists MFA, roles and SSO among controls, with SSO and richer admin features described on the business and enterprise tiers in the way that page sets out. Data residency, including Singapore at rest, is described for eligible enterprise, edu, healthcare and API customers, not as a feature of every login. Consumer free and Plus behaviour was not verified and is not stated here. A personal account is not “basically enterprise”.

Protection and retention still apply. The PDPC’s Advisory Guidelines on Key Concepts (17 May 2022) describe a protection obligation and a retention limitation. Reasonable security arrangements, and not keeping personal data once the purpose and any legal or business need are over, are organisational duties. They are not switched off because the text was pasted into a tool. The retention guide is the operational split between a business record and a prompt log. The PDPA training page is broader than this starter.

The starter, with notes in italics

Copy the headings. Replace the blanks. Delete what does not fit. Then get it reviewed before you issue it.

1. Purpose. This policy tells staff which AI tools they may use for company work and what they must not put into them. Add the date and the person who owns the policy. That person is not automatically your DPO unless you designate them.

2. Approved tools. Staff may use only the workspace accounts listed here: ________. Personal ChatGPT, personal Claude, personal Gemini and browser extensions not on this list are not approved for company work. Name the plan, not just the brand. An approved brand on a consumer login is not the control OpenAI describes for Business and Enterprise.

3. Never paste. National registration numbers, passport or work-pass numbers, bank account numbers, card numbers, passwords, health information, and a full export of customers. Add categories that matter in your trade. Do not assume this list is complete for you.

4. Allowed only after you remove identifiers. Short examples for a draft, with names, emails, phone numbers and account numbers taken out, and only in an approved tool. If you cannot remove them, do not paste. “Anonymised” is a conclusion, not a mood. If the person is still identifiable, you have not done it.

5. Customer messages. A model may draft a reply from an approved tool. A person sends it. The approval ladder is the operational rule. Do not write “the AI may reply”.

6. Connecting AI to company systems. Any automation or agent that sends company data to a model needs written approval from ________ before it is switched on, including what fields are sent and how to turn it off. See the agent guide for the questions to ask. Approval here is internal governance. It is not a legal basis.

7. Retention and training. For each approved tool, record whether the vendor trains on your content by default, how long prompts are kept, and who at the vendor can see them. Recheck when the plan changes. OpenAI’s sentences above apply only to the products that page names.

8. New tools. Staff do not sign the company up to a new AI product. They ask ________. Include a personal-phone app. That is how the unapproved tool arrives.

9. If something was pasted that should not have been. Tell ________ the same day. Say which tool, what kind of data, and whether it was a personal or a workspace login. Do not forward the data again in the report. This is an internal route. It is not a PDPC notification procedure. Breach notification has its own rules. Your adviser should say if they apply.

10. Review. The owner rereads this when a tool or a plan changes, and at least once a year. A yearly date does not cure a bad tool choice.

Eight questions before a tool goes on the approved list

  1. Is the login a business workspace the company administers, or a personal account?
  2. Does the vendor train on inputs by default on this plan? If the answer is only published for a different plan, you do not have an answer.
  3. Can you set retention, and what is the default?
  4. Are there admin roles, MFA, and a way to remove a leaver? OpenAI lists these for the business products on its page. Confirm they are on the subscription you pay for.
  5. Where is content stored, and does your customer contract allow that?
  6. Is there a data-processing addendum, and has someone read it?
  7. Who at the vendor can access content, and for what support reason?
  8. How do you delete a conversation or a file, and does deletion include backups?

If you cannot answer 1, 2 and 3, it is not approved. The n8n guide asks the same kind of question when the model is a node in a workflow you host. Self-hosting the workflow does not answer the question for the model you call.

Making the rule one people follow

A ban with no approved tool pushes staff onto personal accounts, which is the outcome you do not want. Give them the workspace login, show a before-and-after of a pasted complaint with the name removed, and use the rollout approach: ask what they are using the chat box for, replace that job with the approved path, and log the exception. Do not start with punishment. Do not start with a promise that the policy makes you compliant. It does not.

A redaction you can practise before the policy is signed

Use a real complaint only after identifiers are removed, and only in a tool you already pay for as a workspace. Practise on this invented message:

Hi, this is Aisha Rahman, phone 8123 4567, order 44021. The unit arrived cracked. Please refund the card ending 4242.

A version that the starter would allow, if the tool is on the approved list and your adviser agrees the leftovers are acceptable:

A customer reports a cracked unit on an order and asks for a refund. Draft a short reply that asks for a photo and does not promise a refund amount.

What was removed: name, phone, order id, card fragment. What remains is the issue. If your business is small enough that “cracked unit this morning” still identifies the person, you have not finished. That judgement is yours. The template cannot make it.

Map the starter onto the Act only as far as this page checked. Section 11 tells you to designate someone. It does not tell you which AI tool to allow. The protection obligation tells you to make reasonable security arrangements. It does not list approved vendors. The 2024 AI guidelines apply when the system recommends or decides about people. A draft reply is not automatically that system. A score that rejects a job applicant is closer to it. When you are unsure which side a use sits on, stop and ask the designated person before the staff rollout, not after.

Illustrative scenario

Illustrative scenario, not a client and not a legal outcome. A customer-service lead pastes a complaint email, including a name, a phone number and an order id, into a personal chat account to “make it shorter”. The starter, once adapted, says that paste is not allowed, and that a shorter draft may be made in the company’s workspace after the name and phone number are removed. The order id stays only if the approved tool’s plan and the company’s adviser say it may. The lead sends the reply themselves. The scenario does not say the company avoided a breach, because this page cannot say that.

What this starter refuses to do

It does not set a fine, name a safe vendor, or declare a use lawful. It does not replace section 11, a contract with a vendor, or the advisory guidelines when you are actually making decisions about people. High-risk uses, including hiring scores, health data and anything you would not want read aloud, need advice before a template.

The wider organisation list, including the PDPC’s advisory guidelines and IMDA’s voluntary framework, is the SME AI governance checklist. Staff courses are corporate AI training.

If a workflow would send customer text to a model, describe the fields. The operational question can be discussed through the workflow service or WhatsApp +65 9071 3976 (chat). The legal question cannot be closed on that chat.

Frequently asked questions

Is pasting customer data into ChatGPT a PDPA breach?

There is no blanket answer. It depends on what was pasted, why, whether a legal basis applies, and what the tool does with the text. Free and paid consumer plans were not verified for this article. Do not treat a vendor’s enterprise promise as a description of a personal login. Ask a qualified adviser about your facts.

Do we have to appoint a data protection officer?

The Personal Data Protection Act 2012, section 11(3), says an organisation must designate one or more individuals to be responsible for ensuring it complies with the Act. Section 11(4) allows that person to delegate. Section 11(5) requires the business contact information of at least one designated or delegated individual to be made public. Section 11(6) says the designation does not relieve the organisation of its obligations. The outsourced DPO guide explains how some SMEs buy that role in. This FAQ is not a finding that your current arrangement is enough.

What about AI steps inside an automation tool?

They are in scope of the same questions: what text is sent, on which plan, and whether a person checks the output. The n8n hosting guide and the billing explainer are the operational pages. They are not a legal clearance to connect a model.

Can we adopt the template on this page as our policy tomorrow?

No. It is an educational starter. Adapt it to your tools and your data. Have a qualified DPO or legal adviser review the version you intend to use. FYAdvisers has not had this template reviewed for your organisation, and rereading it with software is not that review.

Chat on WhatsAppConcierge · +65 9071 3976
Get Free QuotesWhatsApp Us