Someone in your team has almost certainly pasted client information into an AI tool already. A customer complaint, to help draft a reply. A list of overdue debtors, to summarise. A client’s financial statements, to pull out the key points. It’s quick and genuinely helpful.

It’s also a privacy decision. Under New Zealand law, that decision belongs to your business, not to whoever happened to be at the keyboard.

This article is general information, not legal advice. For your specific situation, talk to a privacy professional or a lawyer.

The Act applies to AI, even though it never mentions it

The Privacy Act 2020 doesn’t have a chapter on artificial intelligence. It doesn’t need one. The Office of the Privacy Commissioner (OPC) describes the Act as technology-neutral and principle-based, and says it applies to generative AI (OPC: generative artificial intelligence).

The Act is built around 13 information privacy principles, now joined by a new principle 3A on collecting information from other sources, which took effect on 1 May 2026. A handful of them matter most when AI tools are involved.

The principles that matter most

Purpose (principles 1 and 10). You collected client information for a particular purpose, and principle 10 limits using it for other purposes. Before running client data through an AI tool, ask whether that use fits with why you collected it.

Storage and security (principle 5). You need reasonable safeguards against loss, misuse and unauthorised access or disclosure. Pasting client details into a free consumer AI account with unclear retention terms is hard to square with that.

Accuracy (principle 8). You must take reasonable steps to check information is accurate before you use it. AI output can be wrong in convincing ways, which is why the OPC expects a person to review it before anyone acts on it.

Disclosure outside New Zealand (principle 12). Most AI tools process data offshore. Principle 12 limits disclosing personal information to overseas organisations unless certain conditions are met. Examples are the individual authorising it after being told it may not be protected to a comparable standard, or you reasonably believing the recipient must protect it to a comparable standard, for instance under an agreement with you.

There’s an important nuance. The OPC’s guidance on disclosing personal information outside New Zealand explains that where a provider only stores or processes information on your behalf, as your agent, the information is treated as still held by you. That changes if the provider uses it for its own purposes. Either way, you remain responsible for it.

So the question that matters most is this: does the AI provider use what you send for its own purposes, such as training its models? That’s why the provider’s terms are worth reading closely.

What the Privacy Commissioner expects

The OPC has published expectations for organisations considering generative AI. In summary:

  • Senior leadership should consider the risks and mitigations and explicitly approve the use.
  • Check whether the tool is necessary and proportionate, or whether another approach would do.
  • Carry out a privacy impact assessment before using it.
  • Be open with customers and clients about how, when and why you use it.
  • Engage with Māori about potential impacts.
  • Have procedures for accuracy, and for handling access and correction requests.
  • Have a person review outputs before acting on them.
  • Don’t enter personal or confidential information unless the provider has explicitly confirmed it won’t be retained or disclosed. Otherwise, strip out anything that could identify people first.

These are guidance rather than new law, but they show clearly how the Commissioner expects the Act to apply to AI.

A privacy impact assessment doesn’t need to be huge

The OPC describes a privacy impact assessment as a tool to identify and assess privacy risks and find ways to reduce them, particularly useful when introducing a new system or changing a process. It also publishes a toolkit to help.

For most small and mid-sized businesses, a privacy impact assessment for an AI tool can be a few pages. What information is involved, and whose? Why are we using it? Where does it go, who can see it and how long is it kept? What could go wrong, and what will we do about it?

Seven checks before your team uses AI with client data

  1. Decide which tools are approved. Use business versions with proper data protection terms, not personal accounts.
  2. Read the provider’s terms. Check retention, whether your data is used for training, where it’s processed and who at the provider can access it.
  3. Write a short AI use policy. Spell out what can and can’t be entered, with real examples from your work.
  4. De-identify before it leaves your hands. Replace names, IRD numbers, account numbers and addresses with placeholders before text reaches the model. That’s the idea behind Veil, a tool we’ve built in-house. It swaps identifying details for labels such as CLIENT_0042 before text goes to an AI model, and lets authorised staff restore the real values afterwards.
  5. Keep a person in the loop for anything that affects a client.
  6. Tell your clients. Update your privacy statement to explain how you use AI.
  7. Do a privacy impact assessment for anything beyond light, de-identified use, and record that leadership has approved it.

Where to start

Find out which AI tools your team is already using and what’s going into them. A short, no-blame survey usually does it. Then pick approved tools and write the policy. Our AI and agents work includes readiness and security reviews covering your Privacy Act obligations, and practical de-identification for client data. Get in touch if you’d like help putting it in place.