ChatGPT is one click away, which is exactly why confidential information can leave a company in seconds. Someone pastes in a customer list, a complaint thread, an unreleased offer. The answer that comes back looks harmless. The input was the problem.
Teams need the rule written down before that happens. Staff should know which account is approved, what may go into it, and who can decide an exception. Training settings are part of the check. They do not settle your duties under the law or under client contracts.
We are firm on one point. A personal ChatGPT account is the wrong place for customer data, even with model training switched off.
Personal data needs an approved purpose
Under Singapore's Personal Data Protection Act, personal data means data, true or not, about a person who can be identified from that data, or from that data combined with other information the organisation holds or is likely to get.
The Personal Data Protection Commission's GenAI guidelines say an organisation using a GenAI service still carries its PDPA obligations. Define the purpose. Limit how much personal data goes in. Assess the provider's safeguards. Tell staff which categories are permitted.
So no CRM exports, mailing lists, survey responses tied to named people or support conversations in an unapproved account. Be stricter again with children's information, health details, credit records and identity documents. If the task can run on anonymised or made-up examples, strip the identifiers and run it that way.
A short input rule prevents a large data problem.
Confidential work can be sensitive without personal data
The PDPA is not the only reason to stop. A media plan or a launch schedule may hold no personal data at all. It can still be confidential under an employment policy or a client agreement.
Keep unreleased pricing, campaign plans, passwords, API keys and private client documents out of personal AI accounts. Same for pitch decks under a confidentiality agreement, and internal performance reports. Redacting helps only when what remains cannot point back to the client, the campaign or the person through context.
Then give people a safe route for low-risk work: fictional customer records, public campaign material, an empty template. Real data enters an AI system after the company has approved the tool and the purpose, not before.
Access control still applies inside an approved account. Work email addresses only. Remove leavers quickly. Restrict projects that hold sensitive material. Decide how long chats and uploaded files may stay. A business workspace is a company system, so it needs an owner and the usual information-security controls.
Consumer and business accounts use data differently
OpenAI says content from individual services such as ChatGPT and Codex may be used to improve its models unless the user opts out, and its training and privacy guide explains that choice. Turning off "Improve the model for everyone" stops new chats being used for training.
ChatGPT Business, Enterprise and the API do not use business inputs and outputs for training by default, unless the organisation opts in. Useful, but that default does not make a given use PDPA-compliant. Access, retention, overseas transfers, security, purpose and the promises you made to customers are still yours to answer for.
Check the contract and the admin settings before you roll anything out. Write down which features are approved, because connectors and file uploads can expose far more than a plain text prompt. Staff should also know whether a feedback button sends the conversation off for model improvement. A safe default changes the moment a user picks another option.
Temporary Chat is a limited control too. OpenAI says Temporary Chats are not used for training, but a copy may be kept for up to 30 days and reviewed for abuse. "Not used for training" is not the same as "not stored".
Paste this rule into the staff handbook
Handbook rule: Do not paste personal data, confidential client material, credentials or unreleased commercial information into a personal AI account. Use company-approved AI workspaces only. Include personal data only when there is an approved purpose, the minimum required fields are used, and the company has reviewed the provider's data handling. Remove names and other identifiers whenever the task can be completed without them. If you are unsure, stop and ask the data protection officer or designated approver.
A rule holds only when the safe option is easy. Give staff the approved workspace, show them how to strip identifiers, hand them sample prompts built on dummy data. If approval takes three days, someone will open a personal account instead.
Back the paragraph with an approved-tools list and a named contact. Revisit it when account settings or company data practices change. This article is general operational guidance, not legal advice. For the formal position, read the PDPC guidance and get professional advice for your own circumstances.