Useful reads
Five things to get in writing before AI touches customer data
The questions a partner, a practice manager or an IT lead should ask any AI vendor, with the Australian and New Zealand rule changes that make them urgent.
The question that stops most AI projects in a small firm or clinic is not “does it work?” It is one sentence from a partner or a practice owner: is it safe with client data?
It is the right question. It is also answerable, in writing, in an afternoon. Here are the five things to get on paper before any AI tool, ours or anyone else’s, touches a customer record. If a vendor cannot answer them inside a day, the product is not ready for a business that holds other people’s information.
If a vendor cannot answer them inside a day, the product is not ready for a business that holds other people’s information.
1. Which tools touch which data
A plain register. One row per tool, one column each for the data it can see, the data it stores, and how long. Most businesses have never written this down for the software they already use, let alone the AI layered on top, and the act of writing it usually finds two things nobody knew about.
This register is also what lets you answer a customer, an auditor or a regulator without a scramble.
2. Where the thinking happens
Ask where inference runs: the actual place the model reads your text and writes its answer. For most of 2025 the honest answer was “the United States”. That has changed. Every major model can now be run in Australia, and most in New Zealand, through the big cloud providers. It costs a little more and it is worth it for anything with personal information.
Ask the question precisely. “Data stored in Australia” and “inference run in Australia” are different promises, and some vendors offer the first while quietly doing the second elsewhere.
3. The sub-processor list
Every AI product sits on other people’s infrastructure: a model vendor, a cloud, a logging service, a payments company. You are entitled to the list. A vendor who cannot produce one either does not know or does not want you to.
In Australia, sending personal information to an overseas contractor is a disclosure your business stays accountable for. In New Zealand, information a vendor holds as your agent is treated as held by you. Either way, the sub-processor list is your list, legally, so you should have it.
4. Who approved what, in a log you can read
If an AI system can send, file or change anything, there must be a record of each action and who said yes to it. Not a transcript. A log: this message went out, under this rule, approved by this person, at this time.
And the approvals must not be defeatable by convenience. The 2025 incident record is largely a list of what happens when an “always allow” button exists: a hidden comment in a web page, a support ticket with instructions in it, a token that could do far more than the job needed. We removed the button from our own crew. Ask your vendor whether theirs has one.
5. The words your privacy policy will need
Two dates matter here.
Australia, from 10 December 2026: an APP entity that arranges for a computer program to use a person’s personal information to make a decision, or do something substantially and directly related to one, that could reasonably be expected to significantly affect the person’s rights or interests must say in its privacy policy what information is used and which decisions or actions are involved. Booking triage may be such a decision. Eligibility checks may be. Recruitment screening often will be. Infringement notices apply.
New Zealand, since 1 May 2026: a new notification duty applies when personal information is collected from someone other than the person it is about. A tool that enriches or receives customer data from a third party on your behalf can trigger it for you.
Neither is hard to comply with. Both require you to know, in plain English, which decisions your tools touch. That is the register from point one, used twice.
What we hand over before any work starts
All five, as a written pack: the register, the inference location, the sub-processor list, the approval log format, and the privacy wording for your policy. For a clinic or a firm it is the document that lets the partners say yes without a committee. We include it because nobody else does, and because it is cheaper to write once than to be asked for it in a hurry.
Not legal advice
This is an operator’s checklist, not legal advice. The dates and duties above are as we understand them in October 2026; your own lawyer or the privacy regulators’ guidance should be your reference for the detail. What we can promise is that every system we set up answers all five questions in writing before it touches a record, and that the answers are kept current as the rules change. That is part of the Full Pot retainer.
