AI & Automation
Is It Safe to Put Customer Data Into AI Tools?
It can be safe, if you decide what goes in before anyone pastes it. Under the PDPA, putting a customer's name, number, address or message into an AI tool is still using personal data, and the business that chose the tool stays responsible for it. The PDPC's Advisory Guidelines on Use of Personal Data in Generative AI, issued on 20 July 2026, say this directly to organisations that use generative AI systems.
The workable rule for a small business: keep identifying details out of general-purpose tools, use a business-grade plan with written terms when personal data genuinely has to go in, and put a person between the AI's output and the customer. The rest of this post explains where that comes from and how to apply it.
Does the PDPA apply when I paste customer data into an AI tool?
Yes. The PDPC's 2026 guidelines call an organisation that uses a generative AI system a “System Deployer” and say deployers bear primary responsibility for making sure the systems they choose can meet their PDPA obligations. They must safeguard the personal data in prompts, inputs and generated outputs, and in any internal data the tool uses.
The provider's role depends on what it does with your input. When it runs your prompt or hosts your data on your behalf, the guidelines treat it as a data intermediary with its own protection duties. When it collects prompts to develop its own models, it is an organisation in its own right. That second case is the reason to read the terms before you paste anything. The PDPC makes the same point about cloud storage: the organisation remains accountable even when a provider holds the data.
What do the PDPC's AI guidelines actually say?
There are two documents, and they cover different things. Neither is legally binding; both say the PDPA and its regulations prevail if there is any inconsistency, and that they are not legal advice.
The 2026 generative AI guidelines are the relevant ones for staff using chat tools. For deployers they ask for four things: be disciplined about the purpose and the amount of personal data a tool needs, safeguard the new data sources it creates, educate users on which personal data may be entered, and keep clear written policies. Their example is an HR chatbot whose owner sets guidelines on what personal data can go in prompts, scans prompts for common identifiers and limits access to prompt logs. They also say deployers should review safeguards regularly, especially where a tool can act on its own across files and systems.
The earlier Advisory Guidelines on Use of Personal Data in AI Recommendation and Decision Systems, issued 1 March 2024, concern systems that recommend, predict or decide. They matter if you use AI to score, rank or screen people, for example sorting enquiries. They expect consent or notification that explains the purpose, in proportion to the risk, and encourage pseudonymising or de-identifying personal data where possible.
What should never be pasted into an AI tool?
Anything you would not be comfortable explaining to the customer, and anything where a leak would hurt them. The PDPC's Key Concepts guidelines say security should fit the nature of the data and the possible harm from a breach, so the line is a judgement, not a list the PDPC publishes. A sensible one for a service business:
- Identity-document numbers, bank or payment details, passwords and access codes.
- A full name together with a phone number, address and the content of their messages.
- Site photos that show people, keys, alarm panels or door codes.
- Anything a customer told you in confidence, and staff records.
What is fine to paste?
Text that no longer identifies anyone. Anonymised data is outside the PDPA, but the PDPC warns that it still carries re-identification risk and asks for no serious possibility of re-identification. So replacing a name with “the client” is not enough if the unit number, street and a distinctive job detail remain.
Safe uses include rewriting your own template, tidying your own price list, drafting a chasing message with a placeholder where the name goes, and summarising a note after you have stripped it of names, numbers, addresses and anything unusual enough to point to one household. Add the real details yourself afterwards, outside the tool.
What controls should a small business put in place?
- Write a one-page rule. Name the approved tools and the categories of data that may be entered. The PDPC asks deployers to educate users on exactly this.
- Use a business plan with written terms. Read the terms for the specific plan, not the marketing page. Look for whether your inputs are used to train the provider's models, how long they are kept and how to delete them, where data is stored, who can access it, and what the provider does after a breach. The PDPC encourages providers to document and share their access controls, data residency, retention policies and breach procedures with deployers; whether your inputs train the model is a question to put to them directly. If you cannot find the answers, treat the tool as unapproved.
- Redact before pasting. A find-and-replace habit works for low volume, and a prompt scanner is worth considering for more.
- Put an approval gate on anything customer-facing. A person reads and sends it. We cover where to place the gate in putting AI in a workflow without handing it decisions.
- Control accounts. Business accounts under your admin, not staff members' personal ones, so you can see and delete what is stored.
- Treat a mistake as an incident. The PDPC defines a data breach as personal data accessed, disclosed or used without authorisation. If someone pastes a customer file into a personal account, tell your DPO, assess it and write down the outcome. The notification rules are in our PDPA checklist for small service businesses.
If you connect an AI tool to your inbox or shared drive so it can act, the risk changes. The PDPC says to restrict what it can reach and escalate high-risk tasks to a person, and points to IMDA's Model AI Governance Framework for Agentic AI for that case.
What if a customer asks what we did with their data?
You have to be able to answer. The 2026 guidelines say the access and correction obligations apply to personal data used in deploying generative AI, and that personal data under your control includes data held by a provider on your behalf. Keep a simple log of which tools received which categories of customer data. It costs a line per use and removes the guesswork when a request comes in.
When to skip AI on customer data altogether
If a template, a mail merge or a saved reply does the job, use that. If you cannot say what a tool does with your inputs, do not paste anything personal into it. And if you only use AI for customer messages a few times a week, redacting is cheaper than any enterprise plan. A business-grade tool earns its place when personal data must go in, regularly, and you have read and are content with its terms.
This is a practical starting point, not legal advice. The guidelines are advisory and will change, so check the PDPC pages linked above before you rely on any detail, and ask a lawyer about anything sensitive or unusual.
The automations we build at Ilumax Systems follow the same line. As our FAQ puts it, the workflow is designed around the level of oversight you need: important actions can require human approval, with fallbacks and records of what happened. They run on accounts in your business's name, so you own them. The model drafts, sorts and summarises; a person decides.
If a recurring administrative task is costing your business more time than it should, we offer a 20-minute workflow conversation — one task, and an honest answer about whether automation is appropriate. Get in touch.
More from the blog
Buying Software
Owning Your Code: Why Repository Ownership Matters
Owning your code means the repository, the accounts and the legal right all sit with you. What that looks like in practice, what goes wrong when a vendor holds it, and what to check before signing.
Security & Compliance
Bank-Detail Change Scams: The Controls That Work
"Please pay our new account" emails are a business email compromise scam that reaches firms paying subcontractors and receiving deposits. The control that stops most of it is a phone call to a number you already held.