ChatGPT earns its keep on writing tasks you put off, not on facts you rely on. Use it for awkward emails, quote descriptions, plain-English explanations, first-draft safety paperwork and long supplier PDFs. It invents specs and prices, guesses at NZ regulations, and keeps what you paste unless you are on a business plan. OpenAI states that ChatGPT Business, Enterprise, Edu and its API do not use your inputs for training by default, checked 8 September 2026.
Five things it does well
1. The awkward email. Chasing an invoice that is 40 days out. Telling a client the job slipped a week because the joinery did not arrive. Declining work without burning the referral. These sit in your head for days because the tone is hard, and the writing is 20 minutes you spend at 9pm. Give it the facts and the tone you want, firm and unapologetic, or warm and regretful, and read what comes back. Fix the two lines that sound wrong. Send it.
2. Site notes into a quote description. You have four lines of shorthand from the walkthrough. It turns them into a scope paragraph, an inclusions list and an exclusions list, in the shape a client can read. The exclusions are the part worth having, because a model asked for exclusions will list the ones you keep forgetting to write down. The price stays yours. This one pairs with the quoting post, which is about why the rest of the quote process stays manual.
3. A compliance requirement, explained before you check it. You know the clause exists and you cannot parse it. Ask for it in plain words, get the shape of what it is asking, then go and read the source. Understanding first and verifying second is faster than reading a standard cold. The order matters, and trap two below is what happens when people stop after the first half.
4. A first-draft job safety analysis. Describe the task, the site and the hazards you have already spotted, and it produces a structured JSA with controls against each hazard. It will surface two or three you had not written down. It will also miss things and include controls that make no sense for your site. Treat the output as a checklist to argue with, then have the person doing the work sign off the version that survives.
5. Summarising a long supplier PDF. A 60-page installation manual, a warranty document, a set of tender conditions. Paste it in and ask what the payment terms are, or which fixings the warranty requires. This is the use with the best hit rate, because the answer is in the document you gave it rather than in its memory. Ask it to quote the line it is drawing from, and you can check the answer in ten seconds.
Three traps that cost real money
Trap one: it invents specs and prices. Ask for the flow rate of a particular pump model, the current price of a length of 20mm copper, or the coverage of a specific membrane, and you will get a confident, specific, plausible number. Sometimes it is right. There is no signal in the answer telling you which time this is. Nothing in the phrasing changes when the model has stopped knowing and started producing something that reads like knowing, which is the failure people mean when they say hallucination. Prices and specs come from the supplier, every time.
Trap two: it does not know New Zealand rules by default. Most of what it has read about building, electrical and gas work is American, British and Australian. Ask a New Zealand question without saying so and you can get an answer built on the wrong code, delivered with the same certainty as a right one. Even when you do say New Zealand, it will reach for overseas material to fill gaps. MBIE, the relevant Standard, the manufacturer's installation instructions and your regional council are the sources. ChatGPT is the thing that helps you understand them faster.
Trap three: it keeps what you paste. On a personal account, your conversations are used to improve the models unless you turn that off in Data Controls. That means a client's name, their address, what you charged them and what you paid for materials, all sitting in a system you have no control over. This is the trap with legal weight attached, and it is the one people trip on the fastest, because pasting the whole email thread is easier than editing it first.
The one rule about client data
No client names, addresses or financials go in unless the plan's data terms say your inputs are excluded from training.
OpenAI's policy page states it plainly: with ChatGPT Business, ChatGPT Enterprise, ChatGPT Edu and its API Platform offerings, by default it does not use provided inputs and outputs to train its models. That is the wording on openai.com/policies/how-your-data-is-used-to-improve-model-performance, checked 8 September 2026. Free, Plus and Pro accounts sit on the other side of that default and need training switched off by hand in Data Controls.
So the practical version. If you are on a personal plan, describe the job and leave the customer out: "a two-storey villa in a coastal zone" rather than the name and the street. If you are on a business plan, you have the terms, and I would still strip the surname and the dollar figures out of habit. A data term is a promise about training, and it is not a promise that the text stops existing.
ChatGPT Business costs USD 25 a user a month billed monthly, or USD 20 billed annually, since OpenAI's April 2026 price cut, with a two-seat minimum. If your firm sits in Microsoft 365 already, the comparison you want is Copilot Studio against ChatGPT Business, which walks through both meters side by side.
Worked example: what the subscription has to save
Take the quote descriptions, since that is the use with the clearest before and after. Say you write 8 quotes a week and each description takes 15 minutes of staring at a blank box.
- Now: 8 x 15 minutes = 120 minutes a week.
- With a first draft to edit: 8 x 5 minutes = 40 minutes a week.
- Saving: 80 minutes a week, or 61 hours across 46 working weeks.
- Cost: two seats on ChatGPT Business at USD 20 each billed annually = USD 480 a year, about NZD 826 at 1.72.
- Break-even: NZD 826 divided by 61 hours, so the time has to be worth about NZD 13.50 an hour.
Two seats of ChatGPT Business break even at about NZD 13.50 an hour of owner time saved, on USD 20 a seat billed annually and NZD 1.72 to the US dollar as at late July 2026. Against a charge-out rate, that clears on the first month's quotes. The number to be careful with is the 5 minutes: if the drafts need heavy rewriting because your prompts are lazy, that figure climbs and the saving goes with it. Which plan fits a bigger team is covered in ChatGPT Plus vs Team vs Enterprise.
The same arithmetic works for the awkward emails, and it is the same shape as the ROI calculator uses for larger automation. Hours out, dollars in, break-even in the middle.
When you do not need it
If you write two emails a week and quote off a price book you know by heart, a subscription buys you a novelty. Skip it. The people who get value here are the ones with a writing pile: quote descriptions, method statements, safety paperwork, client updates, tender responses.
Skip it too if the bottleneck is data rather than words. Retyping invoices into Xero, chasing timesheets, working out which jobs made money: those are integration and reporting problems, and a chat window makes none of them smaller. Stop retyping invoices into Xero is the more useful read for that pile.
And if what you want is something that runs on its own rather than something you talk to, that is a different purchase with a different meter. What an AI agent costs to run covers that side, and AI phone agents in NZ covers the version that answers your phone.
Getting more out of it than a chat window
If you have hit the ceiling of what typing into a box gets you, and the next step looks like connecting it to your job data, book a free 30-minute call. I will tell you which of your piles is worth automating and which one is fine as it is.