How NZ SMBs are actually using ChatGPT — and where it falls over

Most of the NZ small business owners I talk to have been using ChatGPT for a year or two now. Not in a structured way — usually one person on the team opened it up, found it useful, and the habit spread. By the time I get a call, there are already six or seven places ChatGPT has crept into the business: drafting customer replies, summarising meetings, cleaning up spreadsheets, writing job descriptions, helping the bookkeeper understand a Xero error message.

Some of that is genuinely good. Some of it is quietly creating problems the owner has not seen yet. This post is what I keep saying to clients: where ChatGPT actually earns its place in a small business, and where it falls over hard enough that it stops being worth the speed.

The use cases that actually pay back

The pattern I see in real engagements is pretty consistent. The wins come from narrow, high-frequency drafting work where a person is still in the loop before anything goes out the door.

The four things I see working most often:

1. First-draft customer replies. A trades business with a flat-out admin person who used to spend half her morning on email now drafts replies in ChatGPT, edits them in 30 seconds, and sends. Same voice, same accuracy, less keyboard time. The key is she still reads every one.
2. Meeting summaries and action lists. Paste a transcript or your own bullet notes, get back a clean summary plus a list of who-owes-what. This is one of the few places I see ChatGPT replace a task that nobody enjoyed doing in the first place.
3. Cleaning up data. Inconsistent product names, addresses formatted four different ways, dates as text. ChatGPT will not always get it right at scale, but for a 200-row CSV that needs tidying before it goes into MYOB or a CRM, it is faster than writing a formula and easier than handing it to a developer.
4. Translating something technical into plain English. A bookkeeper trying to understand an IRD letter, an admin lead trying to make sense of a Vodafone contract, an owner trying to write a job ad without sounding like a recruitment agency. ChatGPT is good at this and the cost of a wrong word is low.

What these have in common: a human owns the output. Nothing leaves the building without someone reading it. The value is in the speed of the draft, not in delegating judgement.

Where it falls over

This is the part most owners do not see coming, because the failure modes look like wins right up until they do not.

Hallucinated facts on a customer email. The most common one I clean up. Someone asks ChatGPT to draft a reply to a customer query about delivery times, ChatGPT confidently invents a number, the admin person sends it because they are busy. Two days later the customer is annoyed because the parcel did not arrive when "you told me it would". The model does not know your delivery times. It will guess. It will sound certain.

Lost context across sessions. If your team is using a shared ChatGPT account or just the free web interface, every session starts fresh. The "AI that helped you with the proposal yesterday" has no memory of yesterday. People forget this and end up retyping the same context every morning, which is not the time win they think they are getting. It also means tone drifts — Monday's email sounds nothing like Friday's.

Pasting in things that should not leave the building. Customer lists. Supplier pricing. Half a contract. Personal information. Most owners have not had the conversation with their team about what is and is not OK to paste into a US-hosted LLM. Under the NZ Privacy Act, sending personal information offshore is something you are accountable for, and "I just pasted it into ChatGPT to summarise" is a real exposure. I cover this in more depth in AI agent risk and governance for NZ small businesses.

The "ChatGPT did it" black box. Once a workflow runs through ChatGPT, nobody can quite explain what happened. Why did the quote come back at that price? Why did the customer reply mention a feature you do not offer? When something goes wrong, you cannot trace it because the prompt was a bit different each time and nothing was logged.

Quote and pricing drift. I see this one in services businesses. Someone asks ChatGPT to draft a quote based on a brief, ChatGPT fills in numbers that look plausible, the quote goes out. Now you have a customer expecting a price you would never have given them. This is the single most expensive failure mode I see and it always comes from a team treating the model as if it knows things it cannot know.

The pattern that separates the wins from the failures

After enough of these engagements, the pattern is obvious. ChatGPT works in your business when three things are true:

1. The task is drafting, not deciding. Drafts get reviewed. Decisions get acted on. Keep ChatGPT on the drafting side of the line.
2. The input is bounded. "Summarise these meeting notes" works. "Write a quote for this customer" does not, because the model has no idea what your margins, lead times, or stock levels actually are.
3. There is one owner per workflow. Not "everyone uses it." One person who owns how the team uses ChatGPT for that specific job, who notices when the output drifts, and who updates the prompt when something changes in the business.

When those three are true, the time savings are real. I see admin teams getting back two to four hours a week with no extra spend and no integration project. That is a useful outcome and it is genuinely available to almost any NZ SMB right now.

When they are not true, you get the failure modes above and you usually do not notice until a customer or a supplier tells you.

ChatGPT works in a small business when the task is drafting, not deciding. Keep it on the drafting side of the line and the wins are real.

What to do about the gap between "we use ChatGPT" and "ChatGPT works for us"

If your team is already using it informally, you do not need to roll out a policy document or buy enterprise licensing. You need three short conversations.

The first is about what is OK to paste in. Write down a one-page list. Customer names and emails — no. Generic question about a Xero workflow — yes. Supplier contract — no. A draft of an internal blog post — yes. The list does not have to be perfect. It just has to exist.

The second is about which tasks you actually want the team using it for. Not "everything", because that is how the failure modes creep in. Pick three or four specific jobs where the value is high and the risk is low. Drafting customer replies is usually one. Cleaning up data is often another. Be deliberate.

The third is about review. For each of those jobs, who reads the output before it goes out? In a five-person team, this is usually obvious. In a fifteen-person team it is not, and that is exactly when things start going wrong.

If you want to do this properly — figuring out which workflows in your business are actually a good fit, where the boundaries should be, and what to use beyond ChatGPT for the ones it cannot handle — that is the kind of thing I help with. The starting point for most owners is figuring out what to automate first in a small NZ business before adding more tools to the stack.

When ChatGPT is the wrong tool

There is one more thing worth saying. Some of the workflows that get pushed at ChatGPT should not be there at all.

If the task is genuinely repetitive — the same shape of email going out every day, the same report being generated every Monday, the same invoice follow-up running every week — you do not want a person sitting at a chat box pasting a prompt. You want that work to run on a schedule with a human approval step at the end. That is a different category of tool, and it is where most of the durable productivity comes from in the businesses I work with.

ChatGPT is great for the variable, judgement-heavy drafting work that comes up dozens of times a day in different shapes. It is the wrong tool for the work that should just be running in the background. Most small businesses I see have a mix of both, and the mistake is treating them the same way.

The other failure mode I see often enough to flag: teams who chose ChatGPT for everything because it was the first AI tool they tried. There is a whole separate class of failure that comes from using one tool where you needed three, and I have written about that pattern in why most NZ SMB AI projects fail.

If you want help separating the work that suits ChatGPT from the work that needs something more reliable — and getting a clear plan for both — that is the kind of consulting I do. You can read more about how I work on the ChatGPT consulting page or just get in touch.