The Quote AI Priced Off Rates It Invented

If you use AI to help write quotes, you have probably noticed how quickly it produces a number. Copper pipe per metre. Labour rate for a second-fix electrician. Concrete supply price per cubic metre. The figure appears in the output, formatted neatly, sitting inside a professional-looking document, and it looks exactly like a number you looked up yourself.

It is not. In most cases, it is a figure the model constructed from pattern-matching across text it was trained on, which may be months or years out of date, may reflect a different country entirely, and was never verified against a real supplier invoice. The model does not know this is a problem. It has no mechanism for knowing. It produces the figure because producing a figure is what it does.

This article is about that specific failure, why it happens, what it costs in practice, and how to keep using AI for the parts of quoting it is actually good at while keeping a human in charge of every number that goes into a price.

This is operational guidance only. Nothing here is financial, pricing, contractual or legal advice. Your own accountant, lawyer or industry body is the right place for advice that applies to your specific situation.

Why the Model Gives You a Number Instead of Saying It Does Not Know

Language models are trained to be helpful, and "I do not know" is a pattern they have learned to avoid unless they are specifically prompted to use it. When you ask a model to draft a quote for a bathroom renovation, it sees a task: produce a complete, useful document. A document with blank cells where rates should be is less complete, so the model fills them.

The number it fills in is not random. It is the output of pattern-matching across a large body of text that includes old forum posts, trade publications, supplier catalogues from previous years, and articles about construction costs. The model finds patterns in that text and produces a figure that fits those patterns. In some markets and at some points in time, that figure might be close to current rates. In others, it will be wrong by a wide margin.

The critical point is that the model cannot tell you which situation you are in. It does not flag the number as estimated, approximate, or unverified. It presents it the same way it presents everything else: as coherent, confident text.

Why You Cannot Spot the Invented Rate in the Finished Document

A quote produced with AI assistance looks like a quote. The line items are formatted consistently, the descriptions are clear, the totals add up. There is no visual difference between a rate that came from your current supplier's price list and a rate the model generated from training data.

This is not a criticism of AI tools specifically. It is a description of how text generation works. The output is text. Text does not carry metadata about where each figure came from.

If you are working fast, which is most of the time, you read the quote to check whether it sounds right, not whether each individual rate is current. A figure that is in the plausible range for a job will pass that check. The check that catches it, which is comparing each line to an actual supplier quote or your own materials database, takes time you may not spend if you trust the AI to have handled it.

When the Error Surfaces

The error surfaces after the client has accepted the quote. This is not coincidence. It is the nature of the problem. The quote looked correct at the time of sending. The client saw a professional document. They agreed to it.

The gap appears when you go to buy materials or subcontract the work. The copper you quoted at the model's figure costs thirty percent more at your actual supplier today. The subcontractor rate in your area is not what the model suggested. You are now either absorbing the difference or having a difficult conversation with a client who has a signed agreement in their inbox.

Neither option is good. The first one erodes margin on work you have already committed to. The second one damages the relationship and your reputation.

What AI Is Actually Good For in the Quoting Process

Structuring a quote is genuinely useful work, and AI does it well. Describing the scope of work in plain language, making sure the exclusions are clear, writing a payment schedule that is easy for the client to follow, standardising the format across all your quotes so nothing gets missed: these are areas where a language model helps and where the risk of error is low. The words it produces are easy to read and verify. You know when a sentence does not describe the job correctly.

The numbers are different. Rates, quantities, material costs, labour allowances: these need to come from a source that is current, specific to your location, and tied to your actual suppliers and subcontractors. That means your own price lists, your supplier quotes for the specific job, and your own records of what labour costs you in your market.

The practical split is this: use AI to write and structure, use your own data to price.

If you run your quoting and AI use through a documented process, a policy that makes this split explicit and covers which tools your staff are allowed to use and for what, that is worth having in writing. There is a template built for construction and trades at getplainstart.com/ai-policy-for-construction-and-trades.

A Worked Example: Ridgeway Electrical

Ridgeway Electrical is a fictional two-person electrical business used here to illustrate the problem.

The owner, running behind on admin, uses an AI tool to draft a quote for a commercial fit-out. The scope is straightforward: first and second fix for a small office suite, twelve circuits, data and lighting. He pastes the scope notes into the tool and asks it to produce a detailed quote.

The AI produces a complete document. It includes a line item for cable at a per-metre rate, a day rate for a second electrician, and a materials allowance for consumer units and accessories. The document looks exactly like the quotes he sends every week.

He reads it, adjusts the scope description in two places where the AI got the detail wrong, and nearly sends it.

Before he does, he runs the check he has recently added to his process: he opens the document, takes every rate line, and compares it against the last three supplier invoices in his accounting software. The cable rate is seventeen percent below what he paid on his last order, which was placed two months ago. The day rate for the second electrician is based on a figure he has not seen in his area for at least two years.

He updates both figures from his actual records. The revised total is higher than the AI-generated one. He sends the revised quote. The client accepts it.

The AI saved him time structuring and wording the document. His own records supplied the numbers. The combination worked. The AI alone would not have.

The Practical Check Before a Quote Goes Out

This does not need to be a long process. It needs to be a consistent one.

Before any quote goes to a client, go through every line that contains a number. For each rate, identify where it came from. If the answer is "the AI put it there", replace it with a figure from one of three sources: your current supplier's price list or a quote for this specific job, your own recent invoices for the same material, or a subcontractor quote received for this job.

This check takes longer than skipping it. It takes considerably less time than repricing a job after a client has accepted.

The model is not going to improve at knowing your local supplier's current price for 2.5mm twin and earth. That information is not in its training data, will not be in its training data, and changes regularly anyway. The check is not a temporary workaround. It is the permanent process.

The Short Version

AI produces numbers with the same confidence it produces words. It cannot verify those numbers against current, local, real prices. The error shows up after acceptance, at the point where it is hardest to fix. Use AI for structure and wording. Use your own data for every figure that goes into a price. Check every rate line before the quote leaves your hands.

That is the whole system. It is not complicated. It just has to happen every time.