Can I put customer information into ChatGPT?

This is the question we are asked most often, and it almost always arrives in the same shape: someone has already been using ChatGPT for a few weeks, has found it genuinely useful, and has started to wonder — quietly, and usually too late to be comfortable about it — whether they were allowed to.

The short answer is that there is no law in New Zealand that stops you. There is also nothing that waves you through. The Privacy Act asks you to be deliberate, and most of the trouble we see comes not from a bad decision but from no decision at all.

There is no small-business exemption

This surprises people, so it is worth saying plainly. The Privacy Act 2020 applies to an “agency”, and an agency is more or less anyone who holds personal information about other people. There is no staff-number threshold and no turnover threshold. A sole trader with a customer list is covered in the same way a bank is.

Australia has a small-business exemption. We do not, and a good deal of the advice circulating online is written for somewhere else. So the useful question is never “does this apply to me”. It is “what does being careful look like at my size”, which is a much more answerable thing.

Three questions, before you paste

1. Does this actually need to be personal information?

Most of the work people want help with does not need a name attached. Drafting a reply to a difficult complaint, tidying up a quote, working out how to word an awkward reminder about an overdue invoice — none of that needs the customer's name, address, or account number in order to come back useful.

Strip it out and you have moved the whole question off the table. This is the single highest-value habit in the list, and it costs about four seconds. If you find yourself arguing that the tool needs the name to do a good job, that is worth testing rather than assuming; it usually does not.

2. Where does it go, and who can read it afterwards?

The answer depends on which product you are using and which plan you are on, and it changes. A personal free account and a business account from the same vendor can have genuinely different terms about whether your inputs are used to improve the model, how long conversations are retained, and who inside the vendor can look at them.

So do not take anyone's word for it, including ours and including a blog post from last year. Open the settings for the account you are actually using and read what it says about training and retention. If you are on a business plan, the answer is usually more protective than most people assume — but it is worth knowing rather than hoping.

3. Would the customer be surprised?

This is not a legal test, but it is a good one, and it catches most of what the legal tests are trying to catch. If you would be uncomfortable explaining to the person what you did with their information, that discomfort is information. Stop and do it the other way.

What actually gets people into trouble

In practice, the problems we are called in to clean up are rarely subtle. They are almost always one of these.

• Pasting the whole spreadsheet. Someone wants an answer about one customer and uploads the entire export because it was quicker than filtering it.
• A personal free account doing work. Nobody decided this; it is just the account they already had, on their own phone, with no record of what went through it.
• The genuinely sensitive categories. Health information, financial hardship, anything involving children, anything that could identify someone in a small community. These deserve a deliberate answer, not a default.
• Letting the output go straight out. The draft reply that gets sent without anybody reading it, carrying a confident detail the tool invented.
• Nobody knowing what is allowed. Staff are making the call individually, each one slightly differently, and none of them has been told what the rules are.

Notice that four of those five are not really about AI. They are about a business doing something new without saying out loud how it is to be done.

What good practice looks like at this size

You do not need a compliance programme. You need about a page, and for people to have read it.

• Name the tools people may use for work, and the accounts they must use them on. One line each.
• Say that identifying details come out by default, and name the handful of jobs where they may stay in and why.
• List the categories that are off limits without asking first — health, financial hardship, anything about a child, anything a competitor would pay for.
• Require a person to read anything before it reaches a customer. This is where most reputational risk actually sits.
• Write down who to tell if something goes wrong, and do it before something goes wrong.

That last one matters more than it sounds. New Zealand has a notifiable breach regime: if a privacy breach has caused, or is likely to cause, serious harm, you are required to notify the Office of the Privacy Commissioner and the people affected. The time to find out how that works is not the afternoon it happens.

The part people get backwards

Most owners frame this as a question about whether to let the tools in. By the time they ask it, the tools are already in. Someone in the business is using one this week, on their own account, with their own idea of what is reasonable — and the business has no view on it, no record of it, and no way to answer a customer who asks.

A written position does not slow anybody down. It ends the quiet guessing, which is the actual risk, and it means that when someone does ask — a customer, an insurer, a larger client running you through their procurement checklist — you have an answer rather than a sinking feeling.

So: can you put customer information into ChatGPT? Often you do not need to. Sometimes you reasonably can. The thing you cannot do is leave it to each person to work out on their own and hope they all arrive at the same place.