ChatGPT in Excel and Sheets: an NZ ops guide

OpenAI put ChatGPT directly inside Excel and Google Sheets, and on 5 May 2026 it came out of beta to general availability on every plan, running on GPT‑5.5. You describe what you want in plain English in a sidebar, and it builds the sheet, writes the formulas, cleans the data, or explains an existing formula, without you leaving the file.

That matters more here than it looks. For most NZ small businesses the real operating system isn't the accounting package, it's a spreadsheet: the quoting calculator, the Friday cash‑flow tab, the job‑tracker one person keeps alive. ChatGPT just landed in the thing your ops manager opens forty times a day.

Who actually gets it

OpenAI shipped it across the board: Free and Go with limited usage, Plus and Pro under the same agentic usage limits as Codex, and Business, Enterprise, Edu and K‑12 as a free preview through 2 June 2026, after which usage runs on each plan's credits. It works in Excel on desktop and web, and in Google Sheets.

The headline is real. Building a working sheet from one sentence beats starting from a blank grid. What matters for a business is the fine print: whether you can trust the output, and what you're allowed to put in.

Task 1: build the quoting sheet

Start with the spreadsheet that costs you money when it's wrong. Open a blank sheet and brief it like a new staff member: "Build a quoting calculator. Inputs: labour hours, hourly rate, materials cost, margin percentage. Calculate a subtotal, add 15% GST, show a total. Put the inputs in a coloured block at the top and lock the formula cells." It lays out the labels, wires the arithmetic, and adds the GST line, the kind of build that used to be twenty minutes of fiddling with cell references.

The model builds the structure, you own the numbers. It doesn't know your real margin, your GST registration status, or your rates, so check the GST line computes 15% on the right base and confirm the margin lands where you meant it. The build is the time‑saver. The check is the job. It's the same discipline that makes ChatGPT worth using anywhere else in the business.

Task 2: clean a messy export

The second task is in almost every ops person's downloads folder: a messy CSV. Product names typed four ways, dates as text, stray spaces, a column that's half numbers and half "N/A". Point ChatGPT at the sheet and ask plainly: "Standardise the product names to title case, convert the date column to a proper date, strip leading and trailing spaces, and flag any row missing a price." Then: "Summarise total sales by product and by month." For a few hundred rows headed into MYOB, a CRM, or a board pack, that beats writing the formulas and costs less than a developer.

Two checks are specific to NZ data. Dates flip. New Zealand writes day/month/year, and a tool defaulting to the US month/day/year order can silently read 4/5/2026 as 4 May or 5 April. After any date conversion, spot‑check a row where the day is 13 or higher: if it converted cleanly the order was read right, if it errored or flipped it wasn't. Summaries drop rows. Ask it to show the row count it worked from and compare that to the source, because a summary that quietly lost 12 rows looks just as tidy as a correct one.

Where it falls over

OpenAI's launch notes don't list limitations, so treat these as normal behaviour of a language model writing formulas, to verify in your own copy.

Chained formulas are where the risk climbs. A single GST line or a SUMIF is well within range. A multi‑sheet model with nested lookups, array formulas, and conditional logic all referencing each other is not. The output looks complete and confident whether the logic is right or not, and a wrong formula buried in a model is harder to spot than a wrong number on a page.

It has no ground truth. ChatGPT produces something that reads correctly even when it isn't, because it can't see the real answer. In a spreadsheet that's a total formatted perfectly and computed on the wrong range. The fix is the Task 1 discipline: feed it inputs where you already know the output.

NZ formats trip it. If your raw data mixes $1,200, 1200 and 1,200.00, check the cleaned column is genuinely numeric and not text that only looks like a number, because a text column breaks every downstream SUM. Confirm GST is 15% and applied to the figures you meant, not folded into a rate it assumed.

None of this makes it a gimmick. It's a fast first‑drafter that needs a checker, which for an ops manager who already checks the load‑bearing spreadsheet by habit is a fair trade.

The data rule: what's safe to paste

Pasting a spreadsheet into ChatGPT sends it to a US‑hosted service. If that sheet holds personal information, customer names, emails, addresses, anything that identifies someone, you've made a cross‑border disclosure under the Privacy Act 2020, and Principle 12 keeps you accountable for how it's handled offshore.

Your plan changes the handling, not the accountability. OpenAI's policy is that on consumer plans (Free, Go, Plus, Pro) conversations can train its models unless you switch that off in Data Controls, so turn it off before any business data goes in, and keep customer personal information out either way. On Business, Enterprise and Edu, OpenAI says it doesn't train on your data by default. Better footing for internal data, but it's still offshore processing, so the cross‑border obligation still applies.

The team rule is one line. Structural and anonymous data is fine to paste: a quoting template with no customer attached, product lists, sample rows, a CSV with names and emails stripped. Anything with a real person in it, customer lists, contact details, anything you wouldn't email to a stranger, needs a deliberate decision first. The fastest way to get this wrong is to let everyone work it out on their own, which is the core of the AI team training I run for NZ businesses.

Your ChatGPT plan changes how your data is handled. It does not change the fact that, under the Privacy Act, you are the one accountable for it.

When it beats automation, and when it's a band-aid

Use it for variable, one‑off, judgement‑heavy work: a new calculator, a particular export, restructuring a tab before a board meeting. Different shape every time, a human reading the result. That's where it earns its place.

It's the wrong tool the moment the work repeats. If someone's pasting the same prompt to clean the same weekly export, or rebuilding the same Monday report, that's not a job for a person at a chat box, it's a job for an automation that runs on a schedule with a human approval step. Calling a recurring process "we use AI for it now" is a band‑aid.

And there's a ceiling. When the spreadsheet has quietly become the system, the quoting engine the whole business runs through, the job‑tracker three people depend on, the cash‑flow model the bank wants to see, no amount of in‑cell AI fixes critical operations living in a file with no audit trail, no permissions, and one keeper. That's the point to move it into a custom web app built for the job.

So use ChatGPT in Excel and Sheets for fast, checked, one‑off work, and treat the tasks you keep repeating in it as a shortlist of what to automate properly. If you want a second opinion on where your spreadsheets stop being a tool and start being a risk, book a call.