The Privacy Act and AI: What NZ Businesses Need to Know
If you put an AI agent on your phone line or your website, it starts collecting personal information the moment somebody types their name into it. Names, phone numbers, addresses, and depending on your industry, details about somebody's health or finances. That is personal information under New Zealand law, and the Privacy Act 2020 applies to you whether you employ two people or two hundred. Anyone looking into how the Privacy Act and AI fit together in New Zealand runs into the same handful of practical questions, and they are answerable without a law degree.
This article is general information, not legal advice. I build AI systems, I am not a lawyer, and nothing here should be relied on for a decision that carries real risk. For anything authoritative, go to the Office of the Privacy Commissioner at privacy.org.nz, which publishes guidance written specifically for businesses and has material on AI tools.
What I can usefully offer is the practitioner's view. These are the points that come up when a New Zealand business puts an AI agent in front of customers, and the questions I ask before I build anything.
How the Privacy Act applies to AI in New Zealand, whatever your size#
The most common misunderstanding I run into is the belief that privacy law is for big companies. That comes from overseas, where some regimes have thresholds based on revenue or headcount. New Zealand does not work that way. The Privacy Act covers agencies broadly, and a two-person plumbing business collecting customer addresses is inside its scope.
The Act is built around a set of information privacy principles. Rather than quoting principle numbers at you, it is more useful to group them by the questions they force you to answer, because that is how they show up in practice when you configure a system.
They ask you to be clear about why you are collecting information, how you are protecting it, how long you are keeping it, and who else ends up with it. An AI agent touches all four, and most of the work of doing this properly happens during setup rather than afterwards.
Collect what you need for the job, and say why#
The principles start with a fairly simple idea. Collect personal information because you have a lawful purpose connected to what you do, collect it from the person where you reasonably can, and be open about what you are doing with it.
For an AI agent this translates into concrete configuration choices. If the agent's job is to book a plumbing visit, it needs a name, a contact number and an address. It does not need a date of birth, and it should not be asking open-ended questions that invite people to volunteer things you have no use for.
I see agents built the other way around, gathering everything they can in case it is useful later. That creates a liability with no matching benefit. The version of the agent that asks four questions is both more compliant and more pleasant to talk to, which is a rare case of the rules pointing in the same direction as good design.
Being open matters too. Your customers should be able to find out, without effort, that a conversation with your website assistant is recorded and stored, and what happens to it. A short line in the agent's opening message and a plain paragraph in your privacy statement covers most of this. The practical side of what to feed an agent, and what to leave out of it, is in how to train an AI agent on your business.
Security and retention are where most setups are weakest#
You have to take reasonable steps to protect personal information against loss and misuse. What counts as reasonable scales with how sensitive the information is, so a clinic holding health details is held to a higher standard than a landscaper holding street addresses.
In an AI setup the weak points are usually mundane. Transcripts sitting in a vendor dashboard that four staff can log into with a shared password. Exports downloaded to somebody's laptop for a report and never deleted. An integration that quietly copies every conversation into a spreadsheet.
Retention is the other half. The principle is that you should not keep personal information for longer than you need it for the purpose you collected it. Most AI platforms default to keeping conversation history indefinitely because that suits them, not you. Find the retention setting during setup and choose a window that matches your actual needs, then write down why you chose it.
There is also a notifiable breach scheme in the Act. If a privacy breach has caused, or is likely to cause, serious harm, you have to notify the Privacy Commissioner and the affected people. That obligation does not disappear because the breach happened at your vendor rather than in your office. Ask any vendor what its breach notification process is and how quickly it would tell you. The broader vendor question list I use is in AI, customer data and privacy.
Sending information offshore is the point most businesses miss#
Almost every AI tool available to a New Zealand business processes data overseas. The moment your customer's details go into a system hosted in the United States or Europe, you are dealing with the cross-border part of the Act, which sets expectations about the protections that apply where the information ends up.
You do not need to panic about this, and you do not need to keep everything onshore. What you do need is to know the answers to a few things before you sign up.
- Where is conversation data stored, and in which countries is it processed? A vendor should be able to tell you this without checking.
- Does the vendor's contract commit it to protecting the information, and does it allow it to be passed on further?
- Is your data used to train the vendor's models, and can you switch that off? On business plans this is usually off by default. On consumer plans it frequently is not.
- What happens to your data when you cancel, and how long does deletion take?
If your business handles genuinely sensitive information, get advice rather than working it out from a blog. Health information is covered by its own code with additional requirements, which is why clinics need a more careful approach than most, discussed in AI for clinics and allied health practices.
Access, correction and the human in the loop#
People have the right to ask what personal information you hold about them and to ask you to correct it. That right covers AI conversation transcripts as much as it covers your invoicing system. In practice this means you need to be able to search your AI platform for a person's name or number and produce what it holds. Test that you can actually do it before you need to, because some platforms make it surprisingly hard.
The other thing worth building in deliberately is a route to a human. There are decisions an automated system should not be making alone, and there are conversations where a person is going to be upset if they cannot escape the bot. Both the compliance argument and the customer experience argument land in the same place. Getting AI to pass a customer to a human properly covers how to set the triggers so the handoff feels seamless.
Keep the AI out of decisions that materially affect someone, such as whether they get credit or a service. Use it to capture, answer and book, and leave judgement calls with a person.
A short setup routine that keeps you tidy#
Before an AI agent goes live, do four things.
- Write down what personal information it will collect and why each field is needed. Delete any field you cannot justify in a sentence.
- Set the retention window in the platform, and record the setting and the reason somewhere you will find again.
- Confirm in writing where the vendor stores data and whether your conversations train its models.
- Update your privacy statement so it reflects that an automated assistant handles enquiries and stores what it is told.
None of this takes more than an afternoon, and it is far cheaper than untangling it after a complaint. If you are weighing up the wider set of New Zealand-specific constraints on AI, from currency to vendor support hours, AI for New Zealand small business covers them. For the local picture on cost and which industries move first, start with AI solutions for Auckland businesses.
Then go and read the Privacy Commissioner's own guidance at privacy.org.nz. It is written in plain language, it is free, and it is the source that actually governs what you do.
Common questions
Does the Privacy Act 2020 apply to my small business?
Yes. New Zealand's Privacy Act applies to every agency handling personal information, including sole traders and small businesses. There is no exemption based on turnover or staff numbers, which is a common assumption people carry over from other countries.
Can I use an overseas AI tool with New Zealand customer data?
Generally yes, but sending personal information to an overseas provider is treated as a disclosure and carries obligations around the protections that apply at the other end. Check where the vendor stores data and what its contract says before you start, and seek advice if the information is sensitive.
Do I have to tell customers they are talking to an AI?
Being upfront is the safer position and it is what the Office of the Privacy Commissioner encourages in its guidance on automated tools. Beyond privacy, telling people plainly avoids the trust damage that comes when a customer works it out for themselves partway through a conversation.