AI Draft Quote From Site Notes: What to Let It Write

An AI draft quote from site notes works if you keep the AI on the words and keep yourself on the numbers. Paste your 40 words of scribble into the assistant and ask it for one thing only: a customer-facing description of the work, a clean scope list, and an exclusions list. Then you type the prices, the hours and the material quantities yourself, from your own price list. The AI is a copywriter with no access to your rates, your supplier's stock or your van. It cannot know that the consumer unit is a 1990s board on a first floor with no loft hatch. If you let it guess at any figure, it will produce a confident, plausible, wrong one, and once a customer signs, that wrong line is your contract.

Why this matters in 2026

AI assist is now bundled into most trade systems, so the drafting button sits right next to the send button. That is the risk. Two years ago you had to deliberately open a separate chatbot. Today the draft appears inside the quote you are about to email, already formatted, already looking finished. Finished-looking text does not get read. It gets skimmed and sent.

The failure is not that the AI is stupid. It is that a language model completes patterns. Your note says "check the shower pump". Thousands of quotes in its training data say "supply and fit new shower pump, test and commission". So it writes that. You never said "supply". You never said "new". You said "check". That is a two-word gap that costs you a pump.

The 40-word test: a worked example

Here is a realistic site note. Messy, abbreviated, written one-handed on a doorstep:

Mrs H, Ashgrove. Old CU under stairs no RCD. Wants EV charger on drive, ~8m from board. Meter tails short. Garage light not working, prob just lamp. She asked about a shower pump too, didn't look. Away 3rd wk Aug.

That is 41 words. It contains four jobs at three different levels of certainty, one thing you never inspected, one diagnosis that is a guess, and a date constraint. A good AI draft turns that into readable customer text. A bad one flattens the certainty levels into one confident list, and that flattening is the whole problem.

What a safe draft looks like

• Scope you are sure of: "Replace the existing consumer unit under the stairs with a new board including RCD protection. Install one EV charge point on the driveway, approximately 8 m from the board. Renew the meter tails as required for the new board."
• Investigate, do not promise: "Attend to the non-working garage light. Our first check is the lamp. If the fault is in the fitting or the circuit, we will tell you the cost before going further."
• Not seen, not quoted: "You mentioned a shower pump. We have not inspected it and it is not included in this price. We are happy to look while we are on site."
• Constraint restated: "You are away the third week of August. We will book around that."

Notice what the AI did. It wrote full sentences, gave the customer certainty where you had it and honesty where you did not, and it named the thing you never looked at instead of quietly dropping it. That is the job. Notice also what is absent: not one figure, not one hour, not one date it made up.

What AI can safely draft, and what it must never touch

The dividing line is simple. The left column is language, which the model has seen millions of examples of. The right column is facts about your business and this specific property, which it has seen zero examples of. It does not know your rate. It does not know that 8 m across a drive becomes 14 m of cable once you go round the door frame. It will still produce a number if you ask, because producing plausible text is the only thing it does.

The material-quantity trap

This is the one that catches people out. Ask an assistant for a materials list from a note and it returns a beautifully formatted one: cable, glands, clips, board, breakers, sundries. It looks like a takeoff. It is not a takeoff. It is a list of things that usually appear near those words, with quantities rounded to whatever looks tidy. If your system holds a real supplier price list, the quantities come from there and from your measurement. Never from a sentence.

A prompt that behaves

Most bad drafts come from a lazy prompt: "write a quote for this". That instruction invites invention. This one does not:

Turn these site notes into customer-facing quote text for a domestic client. Rules: do not invent any price, hour, quantity, material spec or date. If the note does not say it, do not write it. Split the work into Included, Not included, and Assumptions. Where the note is uncertain or something was not inspected, say so explicitly in plain language. End with a bullet list of every question you need answered before this quote can be sent. Plain English, short sentences, no sales language.

The last rule is the useful one. A good model comes back with "the note does not say whether the garage light fault has been diagnosed" and "no cable route length is given". That question list is the AI doing the thing it is genuinely good at: finding the holes in your own note while you can still ring the customer. Ten seconds of reading it saves the callback.

The two-minute review, in order

1. Numbers first. Read only the digits. Every price, hour, quantity, distance and date. Each one either came from you or it gets deleted. No exceptions, no "that looks about right".
2. Verbs next. Hunt for supply, fit, replace, install, renew, upgrade, certify, guarantee. Each one is a promise. Did you make it? "Check" is not "replace".
3. Absolutes. Delete "all", "any", "full", "complete", "throughout". "Rewire all sockets" and "rewire the four sockets in the kitchen" are different contracts.
4. Exclusions. Is everything you did not look at named? Silence in a quote reads as included.
5. Read it as the customer. Not as the person who wrote the note. You know what the note meant. They only have the text.

If you cannot do those five things, do not send the draft. A quote is an offer. In most markets, once it is accepted you are bound to it, and "the AI wrote that line" is not a defence anyone will entertain.

Where the legal edge actually is

A quote is a contract document, not marketing. Consumer rules across Sweden, the UK, Australia and New Zealand point the same way: a fixed-price offer to a consumer binds you at that price, and vague scope is read against whoever drafted it. In Sweden the reference points are konsumenttjanstlagen and Hantverkarformularet, the standard contract template developed with Konsumentverket. In the UK it is the Consumer Rights Act plus your own terms. Different paperwork, same principle: ambiguity loses.

Two AI-specific traps follow from that:

• Invented guarantees. Models like writing "all work guaranteed for 12 months" because the phrase is everywhere online. If your terms say six months, or say nothing, you have just given away half a year on a job you never priced for it.
• Invented compliance claims. Certification, notification and regulators are not for the AI. In Sweden that means Elsakerhetsverket for electrical work and Boverket for building rules. In the UK it is Part P and your scheme provider. Write those lines once, by hand, and reuse them as a saved block that no model ever rewrites.

The Swedish deduction case: keep AI out of the maths

Swedish readers know ROT: the tax deduction that lets a homeowner take 30% off the labour portion of renovation work, up to an annual per-person ceiling shared with RUT. You quote the net price and reclaim the balance from Skatteverket. Green-tech installations such as EV chargers and solar run on a separate scheme with its own rules. For readers outside Sweden, treat it as the structural cousin of a reduced-rate or grant-funded job: the customer pays less, you claim the difference from the state, and the split between labour and materials becomes legally significant.

This is the worst possible place for a language model. The deduction depends on the labour and material split, the customer's remaining annual allowance, property ownership, and which scheme applies. An AI will happily write "including ROT deduction" into a draft for a job that does not qualify, and you will not notice, because the sentence looks like every other quote you have sent. Deduction maths belongs in your system, calculated from your line items, or with your accountant. Check current amounts and eligibility at skatteverket.se, never with a chatbot.

How this fits a system that already holds your numbers

The clean version of this workflow is a quote tool where numbers cannot be hallucinated because the AI never generates them. In OdinTask the price lines come from your imported supplier price lists and your own rates, ROT and green-tech deductions are calculated by the system rather than typed, and the AI assist is pointed at the wording: turning your notes into scope and exclusions text. You review one screen, e-signing goes out, and the figure the customer signs is the figure your system produced.

The principle holds whatever you use. The AI drafts the sentences. Your price list owns the numbers. A human reads the digits before anything leaves the building. Keep that boundary and AI saves you fifteen minutes a quote. Blur it and it will cost you a job's margin, once, in a way you cannot argue your way out of.