ChatGPT automation for NZ small business: where it works, where it fails

Most NZ small businesses I talk to are already running ChatGPT in six or seven places, and a couple of those uses are quietly creating problems the owner hasn't spotted yet. ChatGPT earns its keep on narrow, high-frequency drafting where a person reads every output before it leaves the building, and it fails the moment the model has to know a fact nobody gave it.

Nobody planned any of it. One person opened ChatGPT, found it useful, and the habit spread until it's woven through the business, with no line drawn between the parts that are safe and the parts that aren't. This is where I draw it.

The drafting jobs that pay for themselves

The wins are consistent, and underneath they all look the same: narrow, high-frequency drafting with a person still in the loop. Four show up in almost every business.

• First-draft customer replies. Say your admin person spends the first hour of the day on email. Drafting the replies in ChatGPT and editing them down before sending turns that hour into a fraction of it, same voice, same accuracy. The point is she still reads every one.
• Meeting summaries and action lists. Paste a transcript or your own bullet notes, get back a clean summary and a list of who owes what. It's one of the few jobs ChatGPT takes over that nobody enjoyed doing anyway.
• Tidying data. Product names spelled four ways, addresses in no fixed format, dates stored as text. For a 200-row CSV that needs cleaning before it goes into MYOB or a CRM, ChatGPT is faster than writing a formula and cheaper than a developer. At scale it drifts, so you still check it.
• Turning technical into plain English. A bookkeeper decoding an IRD letter, an admin lead making sense of a Vodafone contract, an owner writing a job ad that doesn't read like a recruiter wrote it. ChatGPT is good at this and the cost of a wrong word is low.

What they share: a human owns the output, and nothing leaves the building unread. The value is the speed of the draft, not handing over the judgement.

Where it falls over

The failure modes all look like wins right up until they don't. Here they are in the order they'll cost you.

Quote and pricing drift is the most expensive one I see. Someone asks ChatGPT to draft a quote from a brief, it fills in numbers that look plausible, the quote goes out. Now a customer is holding you to a price you'd never have given them. The model has no idea what your margins, lead times, or stock levels are. It fills the gap anyway, and it sounds sure.

Hallucinated facts on a customer email are the same problem in a cheaper spot. A customer asks about delivery times, ChatGPT invents a date, the busy admin person sends it, and two days later the parcel that was "promised" hasn't turned up. The model doesn't know your delivery times. It will guess. It will sound certain.

Context resets every session. On a shared account or the free web version, the assistant that helped with yesterday's proposal remembers none of it. People retype the same background every morning, which is not the time win they think it is, and the tone drifts, so Monday's email reads nothing like Friday's.

Sensitive data walks out the door. Customer lists, supplier pricing, half a contract, personal information, all pasted in to "just summarise it." Under principle 12 of the Privacy Act 2020 you're accountable for personal information you send offshore, and the Office of the Privacy Commissioner's test is that you need reasonable grounds the receiver will protect it to a New Zealand standard. Consumer ChatGPT is a US service that can train on what you paste unless you've turned that off, so a pasted customer list is a real exposure worth talking about before it happens. I go deeper in AI agent risk and governance for NZ small businesses.

Nobody can explain what happened. Once a job runs through a chat box, the prompt was a little different every time and nothing was logged. When a quote comes back wrong or a reply mentions a feature you don't offer, there's no trail to follow. It's a black box you built by accident.

Drafting, not deciding

After enough of these, the line is obvious: ChatGPT belongs on the drafting side of your work, not the deciding side. Drafts get reviewed, decisions get acted on, and every failure above is what happens when a decision slips onto the drafting side.

Two things keep it there. Bounded input: "summarise these notes" works because everything the model needs is in front of it; "write a quote for this customer" doesn't, because the facts that matter live in your head and your systems. One owner per workflow: not "everyone uses it," but one person who owns how the team uses ChatGPT for a specific job, notices when the output starts drifting, and updates the prompt when something in the business changes.

Get those right and the time back is real, usually enough that the admin team feels it by the end of the week, with no new spend and no integration project. Get them wrong and the failures above turn up on their own, usually when a customer or 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.

Some jobs shouldn't be at a chat box at all

The other mistake is pointing ChatGPT at work that was never a good fit for a chat box. If a task is genuinely repetitive, the same shape of email every day, the same report every Monday, the same invoice follow-up every week, you don't want a person pasting a prompt to make it happen. You want it running on a schedule with a human approval step at the end.

That's what Claude Managed Agents do: the repeatable work runs in the background, and a person signs off before anything goes out. ChatGPT is the right tool for the variable, judgement-heavy drafting that turns up in a new shape dozens of times a day. It's the wrong tool for the work that should just be running. Most businesses have a mix of both, and the expensive mistake is treating them the same, or defaulting to ChatGPT for everything because it was the first AI tool the team tried. I've written up that failure in why most NZ SMB AI projects fail.

The fix is a one-page list, not a policy

You don't need a policy document or enterprise licensing to sort this out. You need to draw a few lines, and you can do it in an afternoon.

Write a one-page list of what's OK to paste in: a Xero how-to, yes; a customer's name and email, no; a supplier contract, no; an internal draft, yes. Then pick the three or four jobs you actually want the team using ChatGPT for, high value and low risk, rather than "everything." And name who reads the output before it goes out for each one, which is obvious in a five-person team and the exact thing that breaks in a fifteen-person one.

If you want help drawing those lines, working out which workflows suit ChatGPT and which need Claude Managed Agents running on a schedule instead, that's the work I do. You can read how I work on the ChatGPT consulting page, or just get in touch.