Most New Zealand businesses now have one or two people using ChatGPT or Claude every day. That is a good thing — until a project kicks off, someone uploads a document with personal information in it, and the legal team finds out.
The Privacy Act 2020 does not ban AI. It does not even mention it. What it does is set rules about *personal information* — how you collect it, where it goes, who can see it, and what you do with it. Those rules apply equally well to a paper file and to a prompt sent to an offshore frontier model.
What the Privacy Act actually says about AI
Three principles do most of the work:
Principle 5 — storage and security. You must protect personal information against loss, misuse and unauthorised access. If you paste a patient note into ChatGPT, that note is now stored on someone else's servers, governed by their terms, with access controls you do not own. That is a Principle 5 problem.
Principle 11 — disclosure. You can only disclose personal information to a third party in narrow circumstances. Sending it to OpenAI or Anthropic via the public API is a disclosure, full stop, regardless of whether the model 'remembers' it.
Principle 12 — overseas disclosure. When personal information leaves New Zealand, you must check the recipient is subject to comparable privacy law. The big LLM providers are subject to US law, which is *not* comparable to New Zealand's by default. You can still disclose, but you have to do the work to justify it.
Together, these principles mean you cannot just blanket-ban or blanket-allow AI. You have to think about what data is in the prompt and where the prompt is going.
The Health Information Privacy Code adds another layer
If you are in healthcare, the Health Information Privacy Code (HIPC) overlays the Privacy Act with stricter rules around health information specifically. The principles look similar but the bar is higher — health data is treated as more sensitive, the disclosure tests are tighter, and the audit obligations are heavier.
And as of 2025–2026, the New Zealand Government has been rolling out a PII redaction standard that effectively codifies what 'safe' AI use looks like — particularly for the public sector and the agencies that supply it.
What businesses are actually doing
We see three patterns in the wild:
Pattern 1: blanket ban. IT blocks ChatGPT and Claude at the network. Result: shadow AI. Staff use the consumer apps on their phones with the same data. The audit trail is worse, not better.
Pattern 2: managed enterprise account. The business signs up for ChatGPT Enterprise or Claude Team, ticks the data-retention box and assumes that solves it. It helps — but the data still leaves the country, the trust line still passes through a US provider, and a Privacy Act review will still ask hard questions.
Pattern 3: self-hosted with PII redaction. A chat interface that looks and feels like ChatGPT runs inside the business's own environment. Every prompt is processed through a redaction layer before any frontier model sees a single token. The model returns answers; the original record never leaves.
Pattern 3 is what the Privacy Act, HIPC and the new NZ Government PII redaction standard quietly assume you will adopt for sensitive data. It is also the only pattern that lets the workforce keep using current-generation AI without grinding compliance to a halt.
What this looks like at Bridge Point
We have built two pieces of this exact stack:
Anonamoose is our PII redaction proxy. Point your existing OpenAI or Anthropic SDK at it instead of the public endpoint and identifying information is stripped before the request leaves your network. Multi-layer detection, audit logs, mappable to HIPAA Safe Harbor / GDPR / SOC 2 / ISO 27001 patterns. Read more about Anonamoose →
Polly is our self-hosted ChatGPT-equivalent stack for healthcare. It runs Anonamoose in front of frontier OpenAI and Anthropic models, deployed inside your environment, designed against the Privacy Act, HIPC and the new NZ Government PII redaction standard. We are the only NZ provider with this end-to-end stack. Read more about Polly →
The honest bottom line
AI is not the legal risk. *Sending the wrong data to the wrong place* is the legal risk. Once you separate those two things, the path forward gets a lot clearer:
- Audit what data your team would put into a prompt.
- Decide where prompts can go (in-tenancy, in-country, or strictly local).
- Put a redaction layer between staff and any frontier model that sits outside that boundary.
- Keep an audit trail you can show a regulator.
Done well, NZ businesses can use the same AI everyone else is using — without the Privacy Act tripping them up. Done badly, you get a shadow-AI problem the IT team cannot see.
If you would like a sober walk-through of what your stack would look like, come talk to us.