How to Use AI as a General Contractor: Four Weeks, Four Jobs

Start with paperwork you already write: scope letters and RFIs in week one, site notes in week two, subcontractor bid comparisons in week three, and a review in week four. One ChatGPT or Claude plan covers it, about $20 a month for the owner or $40 to $50 for two seats, plus roughly ten owner hours across the month.

Leave estimating software, site cameras and scheduling tools for later. The first month is about finding out where an assistant is dependable on text you already have, and where you must check every line. Prices below are USD list prices and the hours are planning figures, not measured results, so replace them with your own timings as you go.

Follow me on Instagram@sagnikteaches

What the first month costs: plan, seats and owner time

The software is the small number. Owner time is the real cost, and it is worth planning before you buy anything.

Connect on LinkedInSagnik Bhattacharya

Those seat prices are $25 per user monthly or $20 per user on annual billing, for the standard seat on both products. The pricing pages move, so confirm them on the day you buy. ChatGPT offers Projects on every plan from Free upwards, with a per-project file limit that rises with the plan. Claude's help article says Projects are open to everyone, with free accounts capped at five, and that sharing a Project with colleagues is limited to Team and Enterprise. If your office already pays for a Microsoft 365 business plan, Copilot Chat comes with it at no extra cost, and Google Workspace plans now include Gemini, so a no-cost trial of either is possible.

Subscribe on YouTube@sagnikteaches

Now the time. These are planning figures for an owner who writes most of the paperwork personally:

At an illustrative $60 an hour for the owner's time, 10.5 hours is $630, against $20 to $50 for the plan. That ratio is the reason to spend week four on a proper review rather than carrying on by habit.

Before week one: one Project and one rule about client data

Create a Project in whichever assistant you chose and name it for the company, not the job. Put five short files in it: your standard scope exclusions, a one-page summary of your subcontract terms, a boilerplate paragraph about the company, one daily log you consider well written, and one RFI you consider well written. Then paste the standing instructions below into the Project's instruction box, so every chat starts from the same rules.

You help a small general contracting firm with paperwork.
Use only facts I give you or that appear in the files in this Project.
If a detail is missing, write NOT STATED. Never invent drawing numbers,
dates, prices, quantities or names. When you summarise a document,
quote the sentence you relied on for each point. Write plainly, in
short sentences, without sales language. Mark anything that commits
the company to money or time with [CHECK].

Then agree one rule with everyone who will use it: no client names, addresses, phone numbers, signed contract prices or bank details go into a personal or free account. Call people by role ("the client", "Sub A") and let the real names go in when you copy the finished text into your own documents. On a business or team plan you can relax that rule a little, but the habit costs nothing and survives a change of plan. For the policy wording, the AI policy tutorial for construction companies gives a longer version.

Week one: scope letters, RFIs and addendum summaries

Documents come first because text goes in and text comes out, so you can read every line before it leaves the building. Three jobs fill the week.

A messy request becomes a scope list. Clients ask for extra work in passing, and the request sits in your phone until it turns into a dispute. Paste the message and ask for a list of what was asked, what is missing and what to ask back. Take this message as the input:

hi, can you also do the downstairs loo while you're there, same tiles as the kitchen if possible, and probably sort the radiator thats making a noise. no rush on the loo. thanks

With the Project instructions in place, the reply (illustrative) arrives as two requests, neither yet an instruction. The downstairs WC has no size, no scope and no tile supply stated, and "if possible" is not a commitment. The radiator has no location and no fault described. The assistant then lists four questions: how big is the WC, is it a strip-out or a refresh, who buys the tiles, and which radiator makes the noise and when. A first draft of the reply email can easily promise "we will tile the WC to match the kitchen". The [CHECK] rule flags that sentence because it commits money, and you rewrite it as "we can price matching tiles once we know the size". Owner time: about 30 minutes including the questions to the client.

A drawing conflict becomes an RFI. Give the assistant the facts and the format your Project file shows, nothing more:

Draft an RFI from these facts only. Drawing A-201 shows door D4 at
900 wide. The door schedule on A-601 lists D4 at 826. We cannot place
the joinery order until this is answered; we want to order next Friday.

RFI 014 (draft)
Subject: Door D4 width, drawing and schedule conflict
Reference: Drawing A-201 (D4 shown 900 wide); door schedule A-601
  (D4 listed 826)
Question: Which width applies to door D4, 900 mm or 826 mm?
Why it matters: joinery order is on hold until answered
Response needed by: [CHECK: confirm the date before sending]
Impact on cost or programme: NOT STATED

The last line is the point. A draft that is allowed to guess can fill it with a plausible delay, which then reads as your claim. The instruction block turns the guess into NOT STATED, and you add the real impact only if you know it. Before sending, you still open both drawings and confirm the references yourself, because an assistant cannot see them. About 30 minutes, including the check.

An addendum becomes a change list. Paste the text of the addendum, not the whole tender set, and ask for one row per change: item, what changed, the exact sentence it relied on, and any cost or programme effect, or NOT STATED if the text is silent. The quoted sentence is what saves time. You skim the sentences against the original instead of re-reading 30 pages, and a row with no sentence is an invented row. In this illustration an addendum with six changes takes the owner about 25 minutes, against an hour of reading. One row cites the wrong clause number, which the quote makes obvious in seconds. Allow about an hour for this job the first time, while you learn what to paste, plus another hour at the start of the week for building the Project and gathering its five files.

Review every week-one output the same way, so the habit forms early. Read the NOT STATED lines first, because they show what the assistant refused to guess and what you may need to find out. Then read every [CHECK] flag, then open the source for each quoted sentence. If an output has no flags and no quotes, the instructions did not take, and the first fix is to restate them at the top of that chat. Keep the original message, the draft and your final version together in one folder for the month, because week four needs all three to count what changed.

Week two: site notes that turn into daily logs

Dictate a note at the end of each day into your phone's keyboard microphone, or type bullet fragments, and paste the result in. The aim is a record in a fixed shape, not a better story. The daily report tutorial goes deeper on voice notes, and this week uses the simplest version. First give the assistant the shape:

Turn this dictated note into a daily log with these fields: date,
weather, people on site, work done, problems, instructions or change
requests from anyone, safety, actions with owner and day. Use only
what the note says. Write NOT STATED for anything missing. Never turn
a request into an instruction.

Here is a dictated note, as spoken (illustrative):

Tuesday. Cold, bit of rain in the afternoon. Six on site, four of ours,
plumber's lad and the spark. Finished second fix on the first floor
bathroom apart from the extractor, ordered wrong spigot size. Screed in
the kitchen still tacky so no flooring. Windows were due, didn't turn
up, ring supplier tomorrow. Client popped in, wants the island moved
200 mm, I said I'd price it. Nothing on safety apart from the ladder
in the hall, moved it.

Two things commonly go wrong with a draft like this. One is the weather: a log with a blank weather field looks wrong, so an assistant may fill it with "dry and mild". The other is the island line, which can come out as "Variation instructed: island moved 200 mm". That one matters. A log is a record you may rely on later, and turning "wants it moved, I said I'd price it" into an instruction changes who asked for what. The last line of the prompt exists because of that, and the read-back is the second defence: before you file a log, read the instructions row aloud and ask whether the client has actually told you to do it.

The first day takes about an hour with the template. After that each log costs roughly 10 minutes: two to dictate, five to read, three to file. Five logs plus the template is about 1.8 hours.

Two habits keep the logs usable. Name each file with the date first, in year-month-day order, so a folder of logs sorts itself and the date is never left to the assistant to guess. And when a note is poor, say because it was dictated next to a saw, do not ask the assistant to tidy it into something fuller. A short, honest log with five NOT STATED fields is worth more in a dispute than a fluent one that quietly filled the gaps. If a day matters (an incident, a delay claim, an instruction you disagree with), write that entry by hand from your own memory and use the assistant only for the layout.

On Friday, paste the five logs and ask for a client update. The rules are the same: only facts that appear in the logs, nothing about next week unless a log says it, open decisions listed separately. A draft from the logs above reads (illustrative): "This week the first-floor bathroom second fix was completed apart from the extractor, which needs a corrected spigot. The kitchen screed is still curing, so flooring is on hold. The window delivery missed its date and we are chasing the supplier. You asked for the kitchen island to move 200 mm; we will send a price. Decision needed from you: whether to proceed with the island move once priced." That takes about 40 minutes with checking, and it replaces the Friday evening you used to spend remembering the week.

Week three: levelling three subcontractor bids

Bids are where an assistant earns its keep and where it can cost you most, so the rules tighten. It reads and organises. You price and decide. Three electrical bids arrive for the same ground-floor fit-out. Strip names and contact details and call them Sub A, B and C. Here are the headline figures and the sentences that matter (illustrative):

Give the assistant your own scope list and ask it to mark each item per bidder. It must not price anything:

Compare the three bids below against my scope list. For each item and
bidder write INCLUDED, PARTIAL, EXCLUDED, UNCLEAR or NOT MENTIONED and
quote the exact sentence you relied on. Do not price anything. List
every phrase that shifts risk (by others, subject to survey,
provisional, allowance) with the bidder's name. Finish with one
clarification question per UNCLEAR, PARTIAL or NOT MENTIONED item.
SCOPE: 38 light fittings supplied and fitted; 24 double sockets;
6 data points (cabling only); replace distribution board; fire alarm
alterations; making good after first fix.

A first reply can easily mark Sub A's fire alarm line INCLUDED, because the sentence contains "fire alarm" and "included". The quoted sentence gave it away in about twenty seconds: the interface is in, the panel alterations are out, so the item is PARTIAL. That is the reason for demanding the quote. A table cell without a sentence is a cell you cannot check quickly, so you should not trust it.

The gaps get priced from your own numbers, not the assistant's. Suppose your supplier's quote puts the 38 fittings at $5,400, the panel alterations at $2,900 and making good at $900:

• Sub A: $41,800 + $2,900 (panel alterations) + $900 (making good, not mentioned) = $45,600.
• Sub B: $38,500 + $5,400 (fittings) + $900 (making good) = $44,800, with the survey condition still unpriced.
• Sub C: $44,200, nothing to add.

The lowest headline bid becomes the second lowest once the gaps are priced, and the highest headline bid becomes the lowest. The spread between lowest and highest falls from $5,700 (about 15% of the lowest bid) to $1,400 (about 3%). The clarification emails then write themselves: ask Sub B what the survey condition covers and what the board costs if it fails, ask Sub A to confirm making good, and give both a date. Leave other bidders' prices and names out of every email. A draft to Sub B (illustrative) reads:

Hi [first name],

Thanks for the electrical bid. Before we compare, please confirm:
1. What does "subject to survey of existing distribution board"
   cover, and when will the survey be done?
2. If the board needs replacing, what is the cost, and is it
   inside or outside your price?
3. Your price assumes fittings supplied by others and making good
   by others. Please confirm there are no other exclusions.
Please reply by Thursday so we can issue a decision on Friday.

[your name]

Two practical notes on this week. First, if you are unsure which assistant to use, run the same prompt and the same three bids through ChatGPT and Claude, then compare only the cells where they disagree and settle each against its quoted sentence. A month of two personal plans is about $40, and the disagreements teach you faster than any published comparison. Second, keep the assistant's table as a working document. The bid you award still needs your own signed comparison, because the assistant's table is not a record anyone else has agreed to.

For a deeper test of what this catches and misses, see can AI do construction bid levelling. Owner time is about 1.5 hours for the levelling, 30 minutes for the emails and an hour for a one-page recommendation to the client.

Week four: keep, change or drop each job

Before deciding anything, fill in a scoreboard with your own numbers. These are illustrative figures for the jobs in this plan. Checking time is included in the "with assistant" column, which is timed on repeat runs once each prompt has settled, so it sits below the first-run times quoted in weeks one to three:

The last row is a job you tried and dropped, and a scoreboard with no dropped job is probably flattering itself. Use a simple rule for the rest. Keep a job when it saves at least a third of the time including checking, when the errors you caught were the kind you find in under five minutes by reading, and when no data rule was broken. For a week with one scope letter, one RFI, half an addendum, five logs, one client update and half a bid comparison, the kept jobs take about 330 minutes by hand and about 143 with the assistant, a saving of roughly three hours. If your real figures show a tenth of that, you have learned something cheaply.

Keep a one-page errors log beside the scoreboard. Each line has three parts: what the assistant got wrong, how you spotted it, and which rule you added or tightened because of it. After four weeks the log is a better training document for your staff than a generic list of tips, because every entry is about your paperwork. It also tells you where the standing instructions need another line, and which jobs deserve a second look every time rather than a skim.

Be honest about what four weeks cannot show. Rare errors, such as a misread clause that only matters on one job in twenty, will not appear in a month of small tests. So set a standing spot-check: once a month, take one finished document, find the source for every number and every commitment in it, and note anything that does not match. If the spot-check keeps finding problems, tighten the rules before you widen the use.

Then do three small things. Write the one-page rule on what never goes in (client personal data, signed prices, bank details) and put it where staff will see it. Decide the plan: stay on the owner's $20 plan, or move to two business seats at $40 to $50 a month if the office manager now shares the work. And choose one job for month two by hours lost and risk, not by novelty. Turning site messages into change orders is a natural next job because the daily log already holds the raw material, and the ten prompts for bids, subs and updates give you wording to start from.

Some things stay with people regardless of how the month goes. Contract clauses need an adviser. Prices come from your suppliers and your costs. Method statements and safety decisions belong to the competent person on site. Anything that commits the company to money or time is yours to approve, which is exactly why the [CHECK] flag sits in the standing instructions.

Questions contractors ask before starting

Do I need a paid plan from the first day?

No. Week one works on a free plan if you keep client names and prices out of it, but free limits change and personal plans can train on your chats unless you switch that off. A paid personal plan at about $20 a month raises those limits. Move to a business or team plan once staff share the work or client prices go in.

Can I paste subcontractor prices into an assistant?

Treat bid amounts as commercially sensitive. Strip names and contact details, call the bidders Sub A, B and C, and use a business or team plan, which does not train on your content by default. Check each plan's current data terms before pasting anything you would not email to a competitor.

ChatGPT or Claude: which suits a contractor?

Both have Projects for standing documents, and the individual plans cost the same. Run the week-three bid comparison in both on the same three bids and keep the one whose output you can check fastest. Two monthly plans for one month cost about $40, which is cheaper than choosing badly.

How many hours a week will this save after the first month?

It depends on how much paperwork you write. Track minutes before and after for each job type in week four, including checking time, and keep only the jobs that save at least a third. The figures in the scoreboard here are illustrative, so use your own timings rather than anyone's promise.

Further reads

• How Can AI Help General Contractors? Ten Jobs, Ranked by Effort — Ten more jobs to hand over once the first month works.
• Best AI for Builders: Pick the First Job by Hours Lost and Risk — Choose month two's job by hours lost and risk.
• AI Assistant for Contractors: ChatGPT, Claude and Gemini on Real Jobs — Compare ChatGPT, Claude and Gemini on real contractor tasks.
• Finding Construction Bids Worth Pricing: An AI Bid/No-Bid Screen — Screen incoming tenders before you spend a week pricing them.
• Subcontractor Management: Track Insurance Expiry and Onboarding With AI — Track insurance and onboarding for the subs you level.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI ChatGPT pricing and Business pages, Anthropic Claude pricing page and Projects help article, and the series fact sheet for seat prices; all figures are USD list prices and may change.