Using ChatGPT with confidential documents: the risks and the private alternative
Your staff are almost certainly already using ChatGPT or a similar tool at work — to summarise a contract, draft a reply to a customer, or find a clause in a 200-page policy manual. It saves real time. The question for management is not whether people use AI, but what they paste into it and where that text goes.
This guide sets out, in plain terms, what happens to a document you put into a public AI chatbot, which Malaysian rules apply, and what the options are if you want the benefit without the exposure.
What happens to what you paste in
When someone types or uploads a document into a public AI chatbot, the text is sent to the provider's servers, processed there and usually kept in the chat history. Whether it is also used to improve the provider's models depends on the plan and the settings:
- Personal accounts (ChatGPT Free, Plus and Pro): according to OpenAI's help centre, conversations may be used for model training by default. A user can switch this off under Settings → Data Controls (“Improve the model for everyone”), but the switch only applies from then on and is per user.
- Company plans (ChatGPT Business, Enterprise, Edu) and the API: OpenAI says inputs and outputs are not used for training by default.
- Other tools (Gemini, Copilot, Claude, DeepSeek and others) each have their own terms, which differ between free and paid plans and change from time to time.
The practical risk is not that a model will recite your contract to a stranger tomorrow. It is that confidential material ends up outside your control: on servers abroad, in a personal account the company cannot see or delete, held for as long as the provider's terms say.
The documents that cause the most trouble
- HR files — salaries, disciplinary records, medical certificates, IC numbers.
- Customer data — names, phone numbers, account details, complaints.
- Contracts and tenders — pricing, terms under NDA, bid documents.
- Legal and board papers — advice, litigation, minutes.
- Technical material — source code, credentials, network diagrams.
Most of these contain personal data, which brings in the PDPA, or information covered by a confidentiality obligation to someone else.
Where the PDPA comes in
If a document contains personal data — anything that identifies a customer, employee or supplier contact — the person who pastes it into an AI tool is processing that data on the company's behalf. Three parts of the Personal Data Protection Act 2010 matter:
- Purpose (section 6): personal data may only be processed for the purpose it was collected for. Feeding a customer complaint file into a third-party AI service to draft a reply may stretch that.
- Security (section 9): the company must take practical steps to protect personal data from loss, misuse and unauthorised disclosure. Uncontrolled use of personal AI accounts is hard to square with that.
- Transfers abroad (section 129): since 1 April 2025 there is no Ministerial whitelist. Personal data may go to another country only if one of the conditions applies — for example, that country has a law substantially similar to the PDPA or ensures an adequate level of protection, or the company has taken reasonable precautions such as contractual clauses. The Commissioner's 2025 guideline expects a documented transfer impact assessment.
Breaches of the PDPA principles can carry a fine of up to RM1 million and/or three years' imprisonment, and a personal data breach must be notified to the Commissioner within 72 hours.
What a workplace AI policy should say
Banning AI outright rarely works — people use it on their phones instead. A short, clear policy works better:
- Which tools are approved, and on which plan (a company-managed account, not personal ones).
- What must never be pasted in — IC numbers, salaries, medical data, customer lists, anything under an NDA, passwords.
- What may be pasted in — public information, your own drafts with names removed.
- Who to ask when unsure, and how to report a mistake quickly (it may be a notifiable breach).
- Checking the answer — AI output must be verified against the source before it goes to a customer, a regulator or a court.
The private alternative: AI that stays on your server
The reason staff paste documents into ChatGPT is that they want answers from their own documents — the HR handbook, the SOPs, past contracts, policy manuals. That job can be done without sending anything outside the organisation.
A private document AI such as KnowledgeCore is installed on your own server, or hosted for you by TechnoNet in Malaysia. It reads your PDFs, Word and Excel files, and answers questions with the source document and section cited, so the answer can be checked. The language model runs on that server: there are no calls to OpenAI, Anthropic or any other cloud AI provider, and no telemetry.
| Public AI chatbot | Private document AI | |
|---|---|---|
| Where documents go | Provider's servers, often abroad | Your server, or a Malaysian server you choose |
| Used for training | Depends on plan and settings | No |
| Who can see the history | The individual user's account | Your administrators, with role-based access |
| Answers cite the source | Sometimes | Always, document and section |
| Cross-border transfer (s.129) | Needs assessment | Not triggered when hosted in Malaysia |
KnowledgeCore scored 98% on a 43-question academic-policy benchmark and 47 of 47 on a published PDPA compliance benchmark. Plans start from RM 299 a month.
Questions
Is it illegal to use ChatGPT at work in Malaysia?
No. There is no law against using AI tools at work. The legal risk comes from what is put into them — personal data (PDPA), information under an NDA, or material subject to professional confidentiality.
Does ChatGPT use my company's documents for training?
On personal Free, Plus and Pro accounts, OpenAI may use conversations for training unless the user switches it off in Data Controls. On ChatGPT Business, Enterprise, Edu and the API, OpenAI says data is not used for training by default. Check the current terms of whichever tool and plan you use.
Is turning off 'Improve the model for everyone' enough?
It stops future conversations being used for training on that account, but the text is still sent to and stored on the provider's servers, the setting is per person, and the company cannot enforce it on staff's personal accounts.
Does sending documents to an overseas AI service count as a cross-border transfer under the PDPA?
If the documents contain personal data, yes. Since 1 April 2025, section 129 allows the transfer only if one of its conditions is met, and the Commissioner's guideline expects the company to document its assessment.
Can a private AI work without internet access?
Yes. KnowledgeCore can be installed on a server with no internet access after installation — the model, the search and the answers all run locally.
What does a private document AI cost?
KnowledgeCore plans start from RM 299 a month. Larger deployments on your own hardware are quoted after a short assessment of your documents.
Ask questions of your own documents — without them leaving your organisation.
Talk to us