Why ChatGPT gets your quotes wrong: hallucinations, old prices and no live data

ChatGPT can't price a building job reliably because it doesn't know your labour cost, your overheads, your merchant's prices this week or what's behind the plasterboard. It fills those gaps with plausible numbers and states them confidently. Use it to lay out a quote and word it; take every price from your own figures.

Why does ChatGPT make up numbers?

Because it's built to give an answer. OpenAI describes hallucinations as "plausible but false statements generated by language models", and says standard training and evaluation "reward guessing over acknowledging uncertainty". Ask for the cost of a re-roof and you'll get a figure, not "I don't know your rates".

OpenAI's help centre is plain about it: ChatGPT can produce incorrect facts and dates, fabricated quotes and references, and overconfident answers to complex questions. It advises treating the output as a "first draft, not a final source" and verifying important information.

A property developer with two directors and four or five live sites had used paid ChatGPT for 18 months before speaking to us. "We've noticed hallucinations," he said. His fix was to build his own Excel sheet and feed prices in by hand.

Why are its material prices out of date?

Without its search tool, ChatGPT answers from training data that stops at a cut-off date. With search turned on, it can pull current web pages, but those are retail list prices, not your trade account price, your discount or this week's increase on timber.

Even a correct web price isn't your price. A groundworks firm we spoke to cut its material costs by about 30% by switching suppliers on £300,000 to £400,000 a year of materials. No AI chat tool knows which merchant you use or what you've negotiated. How to price a job covers the costs that get missed even when the materials are right.

What doesn't ChatGPT know about your job?

Anything that isn't in the chat. That includes the things that decide whether a job makes money.

Work out the first two once and they're yours for the year: see overheads per man-day.

What do small guesses cost on a real job?

More than they look. Take a bathroom refit priced properly from the firm's own numbers, then change three inputs by amounts an AI chat tool could easily get wrong.

Worked example

£8,700 bathroom refit: your numbers vs three plausible guesses

Two people for five days on a full refit, supply and fit.

Labour and overheads2 × 5 days × £262 a man-day	£2,620
Materials at your merchant pricetiles, suite, boards, sundries	£4,800
Cost£2,620 + £4,800	£7,420
Price at a 15% margin£7,420 ÷ 0.85, to the nearest £	£8,729
Same job if the guesses were: materials 8% low, one day short, £20 a day low£4,416 + 2 × 4 × £242	£6,352
Price from the guessed cost at 15%£6,352 ÷ 0.85, to the nearest £	£7,473
What the job costs you in factas above	£7,420
Profit at the guessed price£7,473 − £7,420	£53

The overrun rule bites hardest on short jobs. A six-person roofing firm we work with put it like this: "When they run over by one day. That's all the profit gone."

How can you test an AI chat tool on your own work?

Take a job you've finished, where you know the real cost, and ask the AI chat tool to price it from the same description you had at quote stage. Compare line by line. It takes five minutes and tells you more than any review.

1. Pick a finished job.

   Something you priced in the last three months, with the final labour days and supplier invoices to hand.

2. Give it the description you had.

   The enquiry, your site notes and the photos, nothing you learned later.

3. Ask for a priced breakdown.

   Labour days, materials by line, margin. Ask it to show where each price came from.

4. Compare with your real figures.

   Note where it was close, where it was off, and whether it said it was guessing.

5. Ask for sources.

   Open any links it gives. OpenAI warns it can cite sources that don't exist.

What you'll usually find is that the layout, wording and list of items are useful and the numbers are not yours. Keep the first and replace the second.

How should you use AI for quotes instead?

Feed it your numbers, and use it for the writing. Paste your rate card, labour cost per day and the merchant quote into the chat, and ask it to build the breakdown from those figures only and flag anything it had to assume. Then check the sums yourself. Better still, keep your rates in a tool that applies them every time so you aren't pasting them in.

A North East builder who uses ChatGPT for everything learned the other side of this: a multi-option quote listed every option and the client "nearly had a heart attack" at the combined total. Ask the tool to show options as alternatives, each with its own total, and check the layout before it goes.

N

Where Nexley fits

Nexley is the AI back office for UK trade firms: five AI assistants run your tenders, quotes, buying, scheduling and invoicing. It connects to the tools you already work in, you run it all from WhatsApp, and you approve everything. £799 a month.

Nexley drafts quotes from your own price list, labour rates and the prices on your recent merchant invoices, rather than from general knowledge. You send a voice note or photos over WhatsApp, it returns a priced draft that shows where each figure came from, and nothing goes to the client until you approve it.

Start 7 days freeNexley vs generic AI

Common questions

Sources

1. OpenAI: Why language models hallucinate
2. OpenAI Help: Does ChatGPT tell the truth?
3. OpenAI Help: How your data is used to improve model performance

Every rule, rate and date above was checked against these sources on the date shown at the top.