AI for contractors is usually pitched as estimating software, which is the least honest way to sell it. A
language model cannot measure a room, judge whether a subfloor is sound, or know what lumber costs in your
county this month. What it can do is write. And in most contracting businesses, the writing is the part that
happens at 9pm, unpaid, after the tools are packed away.
This guide is about that layer: bid narratives, change orders, client updates, job notes and the
paperwork that decides whether a job is profitable after the fact. If you are still shaping the quote
itself, start with writing quotes and estimates with AI and
come back here for the running-a-job side.
What AI is genuinely good at in a contracting business
Be specific about this, because vague enthusiasm is how owners waste a month. The tasks that respond well
share three traits: they are language, they repeat, and a wrong draft costs you a review rather than a
lawsuit.
- Turning rough site notes into scope text. You write “kitchen, 12 cabinets, remove old
counters, patch drywall where tile comes off, homeowner supplying sink” and get clean scope language back
that a client can read without calling you. - Drafting the bid narrative around your numbers. You supply the price. It writes the
paragraph explaining what is included, what is excluded, and what would change the price. - Change orders. The highest return of anything on this list, covered below.
- Client updates and delay messages. The “we found rot behind the tile” conversation is
easier to send when the wording is already written. - Summarising long documents. Supplier terms, warranty fine print, a subcontractor
agreement someone emailed you at 6am. - Hiring paperwork. Job ads, screening questions, onboarding checklists for new
labour.
What it cannot do, and should not be asked to do
Set this boundary before you start, not after a bad bid:
- It cannot price a job. If you paste a scope and ask what to charge, you will get a
confident number built from nothing. The model does not know your labour rate, your waste factor, your
travel time, or that this particular client changes their mind weekly. - It cannot measure. Takeoffs from drawings are a separate category of software, and even
there the output needs checking. Do not accept a quantity a model invented. - It cannot assess conditions. Rot, asbestos, an out-of-level floor, a panel with no
spare breaker: these are judgement calls made on site by someone with a licence and eyes. - It is not a compliance authority. Permit requirements, code questions and safety
obligations belong to your local authority and your insurer, not a chatbot.
Bidding: give it your numbers, ask for the words
The reliable pattern is that you do the arithmetic and the model does the explaining. Build the numbers
the way you always do, then hand over three things: the scope as you understand it, your inclusions and
exclusions, and the assumptions that would change the price.
That last one is where contractors leave money behind. An exclusion list written plainly (“does not
include permit fees, hauling, or replacing decking found to be soft”) is the difference between a
conversation and an argument in week three. A model writes that list without flinching, which matters,
because most owners soften it out of politeness and then eat the cost.
A reusable bid narrative prompt
Replace the bracketed parts. Keep the rule about not inventing figures; it is the whole reason this
works.
You are drafting the written narrative for a construction bid. Use only the facts I give you. JOB: [one-line description and address type] SCOPE AS AGREED: [bulleted list of work] INCLUDED: [what your price covers] NOT INCLUDED: [what it deliberately excludes] ASSUMPTIONS: [conditions under which the price holds] PRICE: [your figure, exactly as it should appear] TIMELINE: [start window and duration] Write four short sections: What this work includes. What it does not include. What would change this price. Next steps to confirm. Plain English, no marketing language, no adjectives about quality. Do not invent quantities, prices or durations. If a section has no information, write "To be confirmed" rather than guessing.
Read the output as if you were the client looking for a reason to distrust it. Anything that sounds like
a brochure gets cut. The section that earns its keep is “what would change this price”, because it converts
future surprises into pre-agreed variations.
Change orders: the highest-return use in the whole toolset
Unbilled change work is the quiet killer. It is rarely a client refusing to pay; it is usually that
nobody wrote the change down while the wall was open. A model fixes the writing problem, which is the actual
bottleneck.
Keep a standard shape and reuse it:
Draft a change order for a residential renovation. ORIGINAL SCOPE: [what was agreed] WHAT CHANGED: [condition found or client request, in plain words] WHY IT WAS NOT IN THE ORIGINAL: [concealed condition / client instruction / code requirement] ADDITIONAL COST: [your figure] SCHEDULE IMPACT: [days, or none] APPROVAL: signature line, date, and a sentence confirming this work is additional to the original agreement. Write it in neutral, factual language. Do not blame the client. Do not speculate about causes you cannot confirm. Keep it under 200 words.
Send it the same day, in writing, before the work proceeds. The language being neutral is not a
courtesy; it is what makes it easy for the client to sign without losing face.
Scheduling, dispatch and the messages nobody wants to write
A job list turned into a week plan is a language problem more than a scheduling one. Paste the jobs,
the crews and the constraints, and ask for the week laid out with the riskiest dependencies first. Treat
the output as a draft that a human with local knowledge corrects; you know which supplier always runs
late.
The other half is the messages. Running behind, waiting on materials, weather, a no-show, a client who
needs to move a date. These are two-line messages that take twenty minutes of dread to send. Write five
templates once, in your own voice, and keep them. If inbound calls are the bigger leak than scheduling,
the AI receptionist guide covers what answering
services can and cannot absorb.
Job documentation and the paper trail
Photograph as you work, then use the model to turn a voice note into a daily log entry: what was done,
what was found, what is next, any instruction received verbally. Two minutes of talking becomes a record
that settles a dispute six months later. The same approach handles punch lists (“turn these photos into a
numbered defect list, grouped by room”) and closeout documents.
Keep the raw photos and voice notes. The write-up is a summary, not evidence; if a claim goes further
than a conversation, the original is what supports it.
Hiring and managing subcontractors
Job ads are a good fit because the failure mode is cheap: a weak ad costs you a week of bad applicants,
not a job. Ask for the ad to be written around the actual work (“second fix carpentry on residential
renovations, own tools, 3 days a week”) rather than generic traits. Screening questions are where a model
earns its keep: turn your standard criteria into five questions that each force a specific answer instead
of “yes, I am reliable”.
For subcontractor agreements, ask for a plain-English summary of a document you already have, then have
your solicitor confirm anything that matters. A summary is not legal advice, and treating it as such is how
people get hurt.
What stays human, permanently
- Final pricing and margin calls. Yours, every time.
- Site assessment and safety. A model has never stood on a ladder.
- Licensing and code compliance. Your local authority, your insurer, your responsibility.
- Anything you would have to defend in court. If the wording is going into a contract
dispute, a person with a name on it reads it first.
Where to start this week
Do not start with the exciting task. Start with the one that already cost you money last month.
- Write the bid narrative template with your inclusions, exclusions and assumptions baked
in. One hour, reused on every quote. - Build the change order template. Then use it on the next variation instead of
“we will sort it out at the end”. - Save five client-message templates in your own voice: delay, material issue, running
late, change requested, payment reminder. - Try daily logs from voice notes for two weeks. If they survive two weeks, they become
a system. If they do not, you learned that cheaply.
Everything else can wait. The tools are listed in the directory if you want to
compare what each one is actually good at, and the about page explains how guides
here are written and checked.
Questions contractors ask
Can AI do takeoffs from plans?
Not reliably from a general-purpose chatbot, and not from a photo of a drawing. There is software built
specifically for quantity takeoff, and even that output gets checked by someone who knows the trade. Use a
language model for the written description of the work, not the quantities.
Will clients notice I used AI to write the quote?
Only if it reads like it. Generic AI output is long, adjective-heavy and says nothing specific. Edit
until the sentences sound like the way you talk on site, keep your own phrasing for anything about quality,
and the question stops coming up.
Is it safe to paste client details into a chatbot?
Strip what is not needed: full addresses, phone numbers, payment details, anything a client would be
uncomfortable seeing in a third-party tool. Most paid plans let you turn off training on your data; check
that setting rather than assuming it. For anything sensitive, describe the situation without identifying
the person.
What is the single best first use?
Change orders. They are short, they repeat, the language barely changes between jobs, and every one you
send the same day is money you would otherwise have argued about later.
Last reviewed and updated on November 4, 2026.
