Using AI in a NZ Business: What the Privacy Commissioner Expects

Short answer: New Zealand has no AI-specific law for businesses, but the Privacy Act 2020 applies to every AI tool that touches personal information, and the Office of the Privacy Commissioner expects you to approve use at senior level, do a privacy impact assessment first, tell customers when AI affects them, keep a human reviewing outputs, and not put personal or confidential information into tools that may retain or reuse it. Two newer rules matter in 2026: IPP3A (in force since May 2026) on telling people when you collect their information from someone else, and the Biometric Processing Privacy Code, which fully applies to existing users since 3 August 2026 and covers things like voice identification and emotion detection.

What the Commissioner expects, in brief

The OPC's page on generative AI lists what it expects of any agency, public or private, before adopting a tool:

Senior leadership approval, after considering the risks and mitigations.
A necessity and proportionality check: is AI needed here, or would another approach do?
A privacy impact assessment (PIA) before use, including due diligence on how the provider has designed privacy into the tool.
Transparency: if the tool will affect customers and their information, tell them how, when and why, in plain language, before you collect their information.
Engagement with Māori about impacts on their communities and information.
Procedures for accuracy, access and correction.
Human review before acting on AI output.
No personal or confidential information in a tool unless you've confirmed it isn't retained or disclosed by the provider.

The OPC's separate guidance on AI and the Information Privacy Principles (September 2023) adds a blunt line worth remembering: if in doubt, don't use AI tools to handle personal information.

Principle by principle: common small-business uses

The table maps the IPPs the OPC highlights to tools small businesses actually use. The examples are illustrative.

Principle	What it asks	Where AI tools trip it
IPP1–4: collection	Collect only what you need, from the person where possible, tell them why, and collect fairly	An AI note-taker recording every client call by default; a chatbot asking for more detail than the enquiry needs
IPP3A: indirect collection (from May 2026)	If you collect someone's information from another source, notify them unless an exception applies	AI "lead enrichment" tools that pull a prospect's details from other sources into your CRM
IPP5: security	Protect information, including prompts and training data	Staff pasting customer lists into a free chatbot on a personal account
IPP6–7: access and correction	People can see and correct what you hold	AI-generated summaries in client files you can't find, export or correct
IPP8: accuracy	Check information is accurate before using it	Acting on an AI summary of a phone call without checking it; biased screening of CVs
IPP10: use limits	Use information for the purpose it was collected for	Feeding old customer emails into a tool to train a sales assistant
IPP11–12: disclosure, including overseas	Limits on sharing, and extra rules for sending information offshore	A provider that can use your inputs to improve its own models, which looks more like disclosure than storage on your behalf

On the last row: the OPC's guidance on working with third-party providers is the place to check whether your AI vendor is acting only on your behalf, or using your data for its own purposes. Settings that switch off training on your data matter here.

Voice, faces and the Biometric Processing Privacy Code

The Biometric Processing Privacy Code 2025 came into force on 3 November 2025; agencies already using biometrics had until 3 August 2026 to comply, so it now applies to everyone. The OPC lists voice, face, fingerprints, keystroke patterns and gait as examples of biometric information.

The code covers "biometric processing", which it defines as comparing or analysing biometric information by means of:

Identification: working out who someone is.
Verification: confirming someone is who they claim to be (for example, voice ID on your phone line).
Categorisation: analysing biometric information to infer things like mood, emotion, intention, attention or demographic category.

Reading those definitions, an AI phone agent that only turns speech into text to take a booking doesn't obviously fall within them. Adding caller voice recognition, or "sentiment" scoring based on how someone sounds, is much closer to the line. The code excludes detecting a readily apparent expression, but emotion inference is named specifically. If a tool you're considering does any of this, read the code and the OPC's biometrics guidance before switching it on, because it brings extra proportionality, notice and safeguard requirements. The code was amended in March 2026 to reflect IPP3A.

A staff AI policy outline

A one- or two-page policy covers most small businesses. Suggested headings:

Approved tools. Named tools and the account type staff must use (business accounts with training on your data switched off). Personal accounts not allowed for work data.
What never goes in. Health information, bank details, passwords, client files, anything under a confidentiality agreement, staff HR records, unless a tool has been specifically approved for that use.
What can go in. Public information, your own marketing copy, de-identified examples.
Human review. AI output is a draft. Someone checks facts, figures and tone before anything goes to a customer or into a client record.
Customer transparency. Where customers deal with AI (chatbots, phone agents, AI-written replies), say so, and offer a human route.
Recording and transcription. When AI note-takers or call transcription may be used, how people are told, and how long recordings and transcripts are kept.
Biometrics. No voice ID, face recognition or emotion analysis without a documented assessment under the biometrics code.
New tools. A short approval process: what data, which provider, where it's stored, can it train on our data, PIA done.
Incidents. What to do if personal information goes somewhere it shouldn't, including whether it's a notifiable privacy breach.
Review date. At least yearly, and whenever a tool changes how it handles data.

The OPC publishes a Privacy Impact Assessment toolkit you can use for step 8.

This article summarises the OPC's published guidance as of November 2026; it isn't legal advice for your business.

FAQ

Do I need a privacy impact assessment for a simple chatbot on my website?

The OPC expects one before using generative AI with personal information. For a small chatbot it can be short, but it should cover what the bot collects, where it's stored, whether the provider trains on it, and how customers can reach a person.

Is it a privacy breach if staff paste a client email into a public AI chatbot?

It may be, depending on what the email contains and how the tool handles it. The OPC's expectation is clear: don't input personal or confidential information unless you've confirmed it isn't retained or disclosed by the provider.

Does IPP3A apply when my AI tool looks up public information about a customer?

Collecting personal information from a source other than the person can trigger IPP3A's notification duty, subject to its exceptions. Check the OPC's IPP3A guidance before using enrichment features.

Are AI meeting recorders covered by the biometrics code?

Transcription on its own isn't one of the code's three processing types. Features that identify speakers by voice or analyse emotion may be, so review them against the code.

If you're weighing an AI assistant or phone agent and want it set up with privacy settings that match this guidance, we'd be glad to hear what you're planning.