Is It Safe to Upload Customer Data to ChatGPT or Claude? What Indonesia's UU PDP Says
Customer data is personal data, so Indonesia's UU PDP applies when you paste it into AI. The relevant articles, personal versus business plans, and how to do it safely.
Uploading customer data to ChatGPT or Claude can be safe, but it isn't only about the service's security features. Customer data is personal data, so once you paste it into an AI service, Indonesia's Personal Data Protection Law (UU PDP) applies. Your company decides why the data is processed, so the responsibility stays with your company, not with OpenAI or Anthropic. The safest approach: use a business account, send as little as possible, and mask customer identities before pasting anything.
This article is a summary for orientation, not legal advice. For high-stakes decisions, consult a lawyer.
What UU PDP says
Law No. 27 of 2022 was enacted on October 17, 2022 and gave a two-year adjustment period (Article 74). The articles most relevant when customer data goes into an AI service:
| Article | What it says | What it means for you |
|---|---|---|
| Article 1 | The controller decides the purpose and controls processing; the processor processes on the controller's behalf | Your company is the controller; the AI service can act as a processor |
| Article 20 | The controller must have a legal basis for processing, such as consent, performing a contract, or legitimate interest | It has to be clear why that customer data may be processed for this purpose |
| Article 35 | The controller must protect and ensure the security of the data it processes | Choosing the service and setting access are part of your obligation |
| Article 51(3) | Processing by a processor remains the controller's responsibility | If the AI service has a problem, the responsibility still sits with your company |
| Article 56 | Transfers abroad need a destination country with equal or higher protection, or adequate and binding safeguards, or the data subject's consent | AI services usually run on servers outside Indonesia, so this applies |
| Article 46 | After a data protection failure, the controller must notify data subjects and the authority within 3 x 24 hours | A leak through an AI account counts too |
| Article 57 | Administrative sanctions, including fines of up to 2% of annual revenue | The risk is real for businesses |
UU PDP also separates general personal data, such as full name and gender, from specific personal data (Article 4), such as health data, biometrics, children's data, and personal financial data. Specific data calls for extra care.
Personal and business plans differ
| Personal plans | Business plans | |
|---|---|---|
| ChatGPT | Content can be used to train models unless you turn it off | Business and Enterprise don't use business data to train models by default |
| Claude | Chats can be used for training, mainly if you allow it in settings | The Team plan doesn't train models on your content by default |
| Company control | None, the account belongs to the employee | Admin, SSO, and access settings |
Sources: the ChatGPT pricing page, Anthropic's Privacy Center, and the Claude pricing page. For OpenAI, retention details and the Data Processing Addendum are covered in Want to Summarize Your Company's Financial Data in ChatGPT? Is It Dangerous?.
From UU PDP's point of view, the biggest difference isn't ChatGPT versus Claude, but whether the data goes into an account the company manages under a clear agreement, or into an employee's personal account that nobody can oversee.
How to use AI with customer data safely
- Ask first whether the AI needs to know who they are. For segmentation, trend analysis, or summarizing complaints, a customer ID, city, and purchase history are usually enough. Names, phone numbers, and addresses can go.
- Mask before pasting. Replace names with codes like C001, and keep the lookup table in your internal system, not in the AI service.
- Send as few columns as possible. Remove columns that aren't needed for your question.
- Don't paste specific data, such as national ID numbers (NIK), health data, or customers' personal financial data, into third-party AI services.
- Use a business account, and make sure its training and retention settings match company policy.
- Record the purpose and legal basis. For example "complaint analysis to improve service". This record helps when customers or auditors ask questions.
- Delete files and chats containing customer data once the work is done, following your company's retention rules.
An example of this in practice is Can AI Learn Your Sales Style From WhatsApp Chat History?, where customer chats are masked before being used to build a style guide.
How to sort common requests
| Need | Treatment |
|---|---|
| Order counts per city or per product | Fine, because it's aggregate numbers |
| Summarizing this month's customer complaints | Fine, after removing names and contact details |
| Segmenting customers by purchase history | Use masked IDs and a business account |
| Sending a full customer list with names, phones, and addresses | Avoid, unless there's a clear legal basis and a service with a binding agreement |
| Customer ID cards, health data, or bank details | Don't paste into third-party AI services |
One more thing, automated decisions
If AI is used to make decisions about customers, such as rejecting an application or setting a price, another right comes into play. Article 10 of UU PDP gives data subjects the right to object to decisions based solely on automated processing. An example from hiring is in Screening CVs With AI, Watch Out for Bias Getting Automated Too.
Wrapping up
The question isn't "can we use AI or not", but "what data really needs to be sent, to whose account, and on what basis". Once the data is masked, minimized, and goes into a business account, AI can be used with customer data at much lower risk.
If you'd like to set up a safe workflow for using AI with customer data, including automatic masking before the data is sent, tell me what you need.
The photos in this article are from Unsplash, by Maksym Kaharlytskyi (main photo) and Mahen Rin, used under the Unsplash License.
Is your team still doing something by hand?
Tell me about it, and let's find the fix together.
More notes
ChatGPT Work vs Dots, When to Use Which?
Work finishes the tasks you start, a dot keeps work moving. How ChatGPT Work and Dots differ, which plans you need, and example tasks for each.
Read the note
Can AI Learn Your Sales Style From WhatsApp Chat History?
Yes, and you usually don't need to train your own model. How to export WhatsApp chats, turn them into a style guide, a ready-to-use prompt, and the customer data risks.
Read the note