Data protection and AI: a practical checklist

Key takeaways

• Data protection laws apply when an AI system handles personal data, including chats, order records and CVs.
• Know what data the system sees, why, where it goes and how long it stays.
• Sensitive data, such as health information, needs particular care and usually explicit consent or another specific ground.
• Your vendor is a processor: contracts, access limits, logs and a breach process are part of the answer.

Many countries now have a comprehensive data protection law: the EU and UK GDPR, Thailand's and Singapore's Personal Data Protection Acts (PDPA), India's Digital Personal Data Protection Act (DPDP), California's CCPA and CPRA, and many more. They differ in detail but share a core. Any time an AI system reads a customer message, an order record, an invoice with a name on it or a CV, it is handling personal data, and that core applies. For the EU position see the European Commission's data protection pages, and for guidance on AI specifically the UK ICO's AI guidance.

1. Map what the AI will see

Before choosing a tool, list the data it would touch. Typical examples are chat messages, names and phone numbers, delivery addresses, order history, payment references, ID documents, employee records and CVs. Note which items are personal data and which are sensitive. Many problems begin with a system that can read far more than its task requires.

• Which systems will the AI read from, and which will it write to?
• Which fields does it actually need? Remove or mask the rest.
• Does the data include children's data or anything about health, religion, biometrics or similar special categories?

2. Decide your lawful basis and your purpose

Most laws require a lawful basis for collecting and using personal data, such as consent, performance of a contract, a legal obligation or legitimate interests. For routine customer service the basis is often the contract or the customer's own request, but it depends on the facts and the country. Whatever the basis, the purpose should be specific, and the data should not be reused for unrelated purposes without a fresh basis.

• Write down the purpose of the AI feature in one sentence.
• Confirm that the lawful basis covers using the data this way.
• Do not reuse customer data to train models unless you have a clear basis and have told people.

3. Handle sensitive data with extra care

Most regimes treat certain categories as sensitive, including health data and biometric data. These generally require explicit consent or another specific ground. If you run a clinic, a pharmacy or a wellness business, this affects what an AI system may read and store. Many organizations keep clinical questions out of the AI's scope and route them to staff. Our healthcare page describes that approach.

4. Assess the risk before you start

Some laws require, and good practice recommends, a written impact assessment before high-risk processing begins. An AI system that handles sensitive data, profiles people or makes decisions about them is a likely candidate. A short assessment of purpose, data, risks and safeguards is worth doing even where it is not mandatory, and it gives your vendor and your auditors something concrete to review.

5. Tell people what is happening

• Update your privacy notice to say that AI is used, for what, and who processes the data.
• Make it clear in the conversation that the customer is talking to an AI system, and how to reach a person.
• Do not ask customers to send passwords, card numbers or other secrets in chat.

6. Treat your AI vendor as a processor

When a vendor processes personal data on your behalf, it is generally acting as a data processor and you remain the controller. That puts obligations on both sides and calls for a written agreement. Questions to ask:

• Is there a data processing agreement that limits use to your instructions?
• Is customer data used to train models? The answer should be no, in writing.
• Which sub-processors, such as model providers and hosting, handle the data, and where?
• How is access scoped and logged? Can you see who or what accessed which record?
• Is data masked before it reaches a model provider?
• How are staff access, confidentiality and security incidents handled?

Our own approach is described on the security page: least-privilege access, workspace isolation, human approval for sensitive actions, audit trails and no training on customer data.

7. Check cross-border transfers

Many AI services process data outside the customer's country. Most laws place conditions on sending personal data abroad, generally requiring that the destination has adequate protection or that another permitted mechanism applies. Ask your vendor where each step runs, and record the answer. If a model provider is in another country, that is a transfer to account for.

8. Set retention and deletion

• How long are conversations and documents kept, and why?
• Can the vendor delete a customer's data on request, including from logs?
• What happens to your data when the contract ends?

9. Prepare for people's rights

People generally have rights over their personal data, including access, correction, objection, erasure and portability, in the circumstances each law describes. Your systems should let you find the data you hold about a person, including what an AI system recorded, and act on a request in good time.

10. Plan for incidents

Data protection laws expect security measures and prompt notification of certain breaches. The GDPR and Thailand's PDPA, for example, generally require notifying the regulator within 72 hours of becoming aware of a reportable breach, and telling affected people where the risk is high; other laws set their own periods. Agree with your vendor how an incident is detected, who tells whom and how quickly.

11. Keep a person in the loop where it matters

Decisions with significant consequences for individuals, such as hiring, credit or health, should not be left to an automated system alone, and some laws restrict fully automated decisions of this kind. We keep people deciding in these areas by design: see autonomy is a dial, not a switch and our recruitment screening page.

A short vendor checklist

• A written processing agreement and a clear list of sub-processors.
• No training on your customers' data.
• Masking and least-privilege access.
• A complete, searchable audit trail.
• Clear answers on where data is processed and stored.
• A defined incident process and deletion on request.
• A way to switch AI features off, for the workspace or for individuals.

Where to go from here

If you are planning AI for customer messages or documents, start by mapping the data, then choose a vendor that can answer the questions above plainly. For how we design for many languages and regions, read designing AI for multilingual customers and our page on multilingual AI automation, or speak with Primues.

Sources

• European Commission: Data protection
• UK ICO: Guidance on AI and data protection