What an AI quoting engine actually does

Most trade quotes get written after seven at night. The measuring happened on site at ten in the morning; the pricing happens at the kitchen table, supplier price list in one tab, a half-finished template in another, and the memory of what you charged for the last similar job somewhere in between. A March 2026 survey of 167 UK small business owners, commissioned by automation consultancy HeyBRB and reported in Electrical Times, found more than a third spend over 20 minutes writing a single quote from scratch, inside roughly eight hours a week of repeatable admin. It is a small, vendor-commissioned sample, so read it as directional. But nobody who quotes for a living will read it and think it exaggerates.

When we designed a scheduling system for a trades business, we deliberately left quoting out of version one. Not because quoting does not matter; because it is harder than scheduling, and a tool that does it badly is worse than no tool. This post is the longer answer: what an AI quoting engine actually is, where the human stays in it, and what the evidence supports today.

What is an AI quoting engine?

An AI quoting engine is a system that takes a job’s specifics, your cost data, and your quoting history, and produces an itemised draft quote for a person to review and price. Three things go in, three things happen in the middle, one thing comes out.

In: the job spec (the enquiry as the customer wrote it, site photos, measurements, notes), your cost data (materials pricing, labour rates, supplier price files, call-out structure), and your win-loss history (what you quoted on comparable jobs, and which of those quotes turned into work).

In the middle: scope matching, which finds the past jobs most like this one; pricing, which applies current costs and your margin rules to the matched scope; and consistency checking, which flags where the draft sits out of line with what you have charged before.

Out: an itemised draft. Line items, quantities, labour, materials, exclusions, and the comparable jobs it drew on, shown so you can check the reasoning rather than take its word.

This is the shape of what we call a Quoting & pricing engine, and the name has two halves for a reason. Quoting, the assembly of an itemised document from scope and cost data, is work a machine can draft. Pricing, the decision about what number goes on it, is not.

Why does the final price stay a human call?

Because a quote is not a calculation; it is a bet, and the person who signs it carries the loss if the bet is wrong. The engine prices what it can see: the scope as described, the materials as listed, the hours a comparable job took. It cannot see what you saw on site. The switchboard that has clearly been “fixed” twice before. The access that turns a two-hour job into a four-hour one. The client who will call daily. How much you want this job given what else is in the pipeline this month.

Two identical bathroom strip-outs can deserve different prices: one is for a builder who has sent you work for a decade, the other for a stranger renovating to sell. No line item captures that, and it should not. That margin call is your business judgement, which is exactly why it stays with you.

We hit the same boundary when we built an AI invoice verifier: a system can only assess what is in front of it, and the most expensive risks are usually the ones that are not. The honest division of labour is the engine drafts, the human prices. Every design decision follows from that.

What does the evidence actually say?

The evidence says AI is genuinely good at estimating and drafting, and not good enough to price unsupervised. Three sources, each with its class declared.

Peer-reviewed: a 2025 systematic review in the journal Modelling synthesised 39 studies of AI cost estimation published between 2016 and 2024. The best deep learning models estimated project costs with 85-90% accuracy; standard machine learning models averaged 75-80%, and regression models 70-80%. Read those numbers the way an estimator would: an estimate that is 85-90% accurate is missing by 10-15%, and on a job priced at a typical trade margin, that miss can be most of the profit. The studies also lean on large project datasets from construction, healthcare, and real estate. A five-person trades business generates a smaller, messier history than any of them.

Published study: a 2011 Harvard Business Review audit of 2,241 companies found that firms contacting a lead within an hour were nearly seven times as likely to qualify it as firms that waited even one hour more, and over sixty times as likely as firms that waited a day. That is lead response rather than quoting, but the mechanism transfers directly: when services are comparable, the first credible, itemised number in the customer’s inbox anchors the decision. Speed is where quoting engines earn their keep, not precision.

Our own evidence review: our State of AI review reached the same split in mid-2026. Quote drafting inside trade job-management platforms sits in the deployable-today column: the drafting layer works in production now. Autonomous pricing does not appear in that column, and on current evidence it should not.

Put together: the win is not that AI prices better than you. It is that a draft you would have written at nine tonight can exist by lunchtime, checked and sent while the competitor is still promising to “get something over by Friday”.

Where does an AI quoting engine not land yet?

It does not land anywhere the data is thin or the risk is invisible. Four failure modes come up repeatedly, and we would rather name them before a client discovers them.

Stale cost data. An engine drafting off last quarter’s supplier price file quietly erodes margin on every quote it touches. Materials pricing moves; the engine is only as current as the price file behind it, and keeping that file current is unglamorous plumbing that no model replaces.

Thin win-loss history. A business writing five quotes a week produces a couple of hundred data points a year, spread across job types. That is enough to match scope against past work. It is nowhere near enough to learn what your market will bear, which is the thing autonomous pricing would need.

The cost model that lives in one head. In many trades businesses the real pricing logic has never been written down; it is the owner’s accumulated judgement. An engine cannot draft from a model that does not exist. When we find this, and we find it often, the first step is not AI at all. It is getting the cost model out of the head and into a structure, which pays off whether or not an engine ever sits on top of it.

Autonomous sending. No system we would put our name to sends a binding price to a customer without human sign-off. A wrong email draft is embarrassing. A wrong quote is an offer you are expected to honour.

What is worth doing now?

Draft-and-approve, wired to your enquiry channel. The engine assembles the itemised draft, ideally straight from a structured enquiry captured on your website or phone line; you review the scope, set the number, and send the same day. That converts the speed advantage the response-time evidence points to, without surrendering the one decision that is genuinely yours.

That is also why quoting is not really a standalone problem. The engine’s best input is a well-captured enquiry: job type, address, urgency, photos, asked for consistently at the moment the customer reaches out. Enquiry handling and quoting are one workflow seen from two ends, which is how we build them.

Common questions

What data does an AI quoting engine need?

Three things: the job spec (the enquiry, site notes, photos, measurements), your cost data (materials, labour rates, supplier price files), and your win-loss history (what you quoted on similar jobs and which ones landed). If your cost model currently lives in someone’s head, writing it down is step one, before any AI is involved.

Can AI set the price on a quote without a human?

Not in any system we would put our name to. An AI quoting engine drafts the itemised quote and shows its reasoning; a person reviews the scope, adjusts the margin, and approves the price. A wrong email draft is embarrassing, but a quote is an offer you are expected to honour, so the number stays a human decision.

How accurate is AI at estimating job costs?

A 2025 peer-reviewed systematic review of 39 studies found the best deep learning models estimate project costs with 85-90% accuracy. That sounds strong until you notice the residual: a 10-15% miss on cost can be most of the margin on a trade job. Treat AI estimates as fast, consistent drafts to be checked, not final numbers.

Is an AI quoting engine worth it for a small trades business?

It depends on volume and response time. If you write more than a handful of quotes a week, quote after hours, or lose work to faster competitors, a draft-and-approve engine usually pays for itself in recovered evenings and faster responses. If you quote rarely and every job is bespoke, start with structured enquiry capture instead.

Perth AI Consulting builds enquiry and quoting systems for trades businesses, on one document-engine architecture, from enquiry handling through to full quoting engines, and tells you honestly which half the machine should own. Start with a conversation.