ChatGPT Now Builds Calculators in Chat. Check the Maths.

5 min read · 12 October 2026

This week: ChatGPT can now build you a calculator mid-chat. Check the maths before you quote from it.

Hello,

OpenAI launched Intelligent UI for ChatGPT on 7 October, alongside GPT-6. Instead of a wall of text, answers can now come back with tappable buttons, editable charts and calculators you can type numbers into. Free and Go users got it a day later. It looks impressive. It also looks finished, and that is exactly the risk.

Last week we looked at shadow AI and the tools your team may already be using. This week has a related thread: polished is not the same as correct. It runs through the new tools, and it runs through the scam research too. Structure, not AI noise. Let’s get into it.

The headline: a calculator that looks right is not a calculator that is right

According to TechCrunch’s coverage, Intelligent UI lets ChatGPT answers include interactive elements: buttons, task-specific calculators, and charts you can edit directly. OpenAI’s examples ranged from a savings calculator to a hiking map. You can also turn the visuals down if you find them noisy.

For an operator, that is genuinely useful. You could describe a quoting method, a cost-per-job estimate or a rostering rule and have a working calculator in minutes. I would use it for a first draft.

Here is the catch. The model decides the formula, the rounding and how GST is treated. You do not see those choices. A roundup of this week’s launches raised the same caution: check the calculation yourself before using it for capacity or pricing decisions. I agree. A wrong number inside a clean interface gets believed faster than a wrong number in a paragraph.

This is not a tool problem. It’s a systems problem. The tool will build whatever you describe. The system is the step where a human checks it against reality.

Tools worth trying this week

ChatGPT Intelligent UI (OpenAI)

Covered above. Good for sketching a quoting tool, a pricing comparison or a simple cash-flow view. Treat the output as a draft until you have tested it against a job where you already know the answer.

Audio uploads in ChatGPT (OpenAI)

From 6 October, paid plans and workspaces can upload audio and get a transcript, a recap and draft follow-ups. Free accounts do not get it, and availability depends on your plan, region and app version. Speaker labelling may not be reliable, so confirm who owns each action before you send anything. If you record customers or staff, check the consent rules that apply to you first.

Google skills, replacing Gems

Google announced on 30 September that reusable saved instructions are moving from Gems to skills. The Workspace rollout began on 5 October. Per the roundup I read, Gems will not be removed for business accounts before 1 March 2027, and skills do not yet sync between the Gemini app and Workspace. If you rely on a Gem, copy its instructions into a document now. It takes two minutes and saves you rebuilding later.

ChatGPT Scan on iPhone (OpenAI)

Launched on 1 October. Scan several pages from the ChatGPT camera and get one PDF. Handy for receipts, delivery dockets and handwritten job sheets. Handwritten numbers can be misread, so check figures and names against the original.

Claude Haiku 5.5 (Anthropic)

Released on 7 October as a low-cost model for narrow, high-volume work such as sorting enquiries or pulling details out of documents. I use Claude daily, so I will say plainly: cheaper does not mean more reliable. If you or your developer are considering it, test it on a small sample of real jobs first. API pricing starts at $0.10 per million input tokens for prompts up to 100,000 tokens, though most owners will feel this as cheaper tools rather than a bill they read.

If you are still working out where any of this fits, the pillar guide to implementing AI in a small business covers the order of operations. Workflow before automation. Always.

Quick win: the 15-minute known-job test

Before you trust any AI-built calculator, tool or template with a new quote, run it against a job you already know the answer to.

1. Pick one finished job. Choose a quote or invoice where you know the real hours, costs and final price.
2. Describe your method to the tool in plain English. Include your markup, GST treatment and any minimum charges.
3. Feed it that job’s inputs. Compare its answer with what you actually charged.
4. Chase every difference. Rounding, GST, a missed minimum charge. Write down where it went wrong and fix the instructions.
5. Save the verified version and the test job together. Re-run the test whenever the tool or your prices change.

While you are at it, keep real customer details out of the test. Use a past job with the names stripped. That one habit covers a lot of the risk in that 32 per cent.

Watch out for

An analysis by BizCover, reported by IT Brief on 25 August, found small business reports to Scamwatch fell 20.8 per cent between 2024 and 2025. That sounds like good news. It is not quite. Total losses still rose 3.4 per cent, and the average loss per report jumped 30.4 per cent. Fewer scams are landing, and the ones that do are costing more.

1. The payment change that arrives with the right details

Payment redirection is the classic: altered supplier bank details on a fake or hijacked invoice that slips through a rushed check. AI makes the message and the invoice look right. Do this: never change bank details because of an email. Confirm by phone, on a number you already had, before the next payment goes out.

2. The impersonator who knows your business

The research describes criminals posing as suppliers, executives or existing customers, using accurate business details, and in some cases deepfakes that copy a familiar face or voice. Do this: verify any unexpected payment request through a separate, known channel. Agree an internal check, such as a call-back or a pre-set question, for anything urgent.

3. The slow scam that spans several channels

Experts quoted in the same coverage warn that a scam can run across a social media ad, a text message, a login and finally a bank transfer. No single step looks wrong. And spelling mistakes and odd wording are becoming an unreliable warning sign. Do this: slow down high-value payments. Build in a pause, and connect the warning signs before money leaves.

Big picture: the cost of a miss is rising while the free help is shrinking

The Australian Signals Directorate reports the average cost of a cyber incident rose 14 per cent to $56,600 in 2024/25, according to The New Daily’s report on 6 October. The same article says the Council of Small Business Organisations Australia (COSBOA) wants the government to reinstate the free Cyber Wardens training program, which stopped when its funding was not extended in the latest federal budget. COSBOA is also asking for plain-language guidance on adopting AI safely.

I cannot tell you whether that funding will return. What I can say is that waiting for government guidance is not a plan. The businesses that stay out of trouble are the ones with a few written rules and a check built into the workflow, so that a polished calculator or a polished invoice has to prove itself before anyone acts on it.

Tie those rules to the type of data and the type of job, not to one product. Tools change monthly. A rule like “nothing goes out until it has been tested against a known job” holds up whichever assistant you use. If you want a second set of eyes on where to draw those lines, the AI automation for Wollongong businesses page explains how we work with Illawarra operators.

Workflow before automation. Always.

Book Your Free Strategy Call

Run the known-job test this week. Fifteen minutes now beats a wrong quote later. See you next week.
Ashley