Choose the assistant that matches where your documents already live, then prove it on your own jobs. At about $20 a month each, ChatGPT, Claude and Gemini can all draft a quote, a scope of works and a tricky client email. Gemini suits firms already on Google's office tools; otherwise run ChatGPT and Claude through the same three jobs.
The brand matters less than three habits. Give the assistant your own rate sheet so it never guesses a price. Make it flag every gap instead of filling it. Recalculate every total yourself. These tools write fluent paperwork, which is exactly why a wrong quantity or an invented promise slips past a tired reader at nine in the evening.
The subscription is the cheap part: $20 is well under an hour of a foreman's time. The expensive part is a quote that leaves out the tile waste, or an email that promises a date nobody agreed. The test below is built to catch those two failures before you commit to anything beyond a trial month, and it uses one imaginary job so you can copy the method onto a real one.
The job notes every assistant gets: a bathroom refit
Use one real job from last month with the names stripped out. For the examples here, take an imaginary bathroom refit: a 2.4 m by 1.8 m room, a walk-in shower replacing a bath, the toilet moved 400 mm, a new vanity, an extractor fan and about 14 m² (roughly 150 sq ft) of tiling. The notes are deliberately messy, the way they sound when you dictate them in the van.
JOB NOTES (dictated, untidied)
Bathroom refit, first floor. Room 2.4 by 1.8. Strip out bath, basin, WC and old floor tiles. Skip needed.
New walk-in shower, tray 1200 x 800 plus glass screen. Move WC across 400 mm, plumber to do it.
New vanity unit and basin. Extractor fan, electrician, plus bonding check. Tile walls and floor, about 14 m2 total.
Client wants large-format tiles, has not picked them yet. Client works from home, so no loud work before 9 or after 5.
Stud wall behind the bath looks damp at the bottom. Floor might be chipboard, not checked.
Client away 12th to 19th.
My hours: strip-out 10 (labourer), shower fit 4, vanity fit 3, making good and painting 12 (carpenter rate).
Plumber's written price $1,150. Electrician's written price $640.
RATE SHEET (illustrative, use yours)
Carpenter or tiler labour: $52 an hour
Labourer: $34 an hour
Tiling labour: $62 per m2
Tile supply allowance until the client chooses: $45 per m2, plus 10% waste
Shower tray and screen, supplier price: $520
Vanity unit and basin, supplier price: $410
Skip hire (small): $360
Subcontractors: pass through at their written price
Margin: 20% on the subtotal
Quote valid for 30 days
For each assistant, open a new chat and, if the app has a memory feature, switch it off for the test so earlier chats cannot colour the result. Paste the rate sheet, then the notes, then the prompt for the job. Do not nudge or correct the first answer. Write down two numbers per run: the minutes until you would be happy to send the document, and how many fixes you made. If you can, repeat on a second job of a different type, because one job proves very little.
Job one: turn the notes into quote lines you can check
Quotes are where a general assistant earns or loses its place, because every figure can be checked against a calculator. The prompt below does three things on purpose. It bans guessing, so a missing rate becomes a visible gap. It demands the basis of each line, so you can audit quantity times rate in seconds. And it forbids additions, which stops the polite padding that turns into a promise.
You are helping a small building firm write a quote. Use ONLY the rates in the rate sheet.
If a line has no rate, write RATE NEEDED and do not estimate one.
Show each line as: description, basis (quantity x rate), amount.
Then show subtotal, margin and total.
List assumptions and exclusions at the end.
Do not add anything the job notes do not mention.
RATE SHEET:
[paste]
JOB NOTES:
[paste]
A usable draft, built from the notes above, looks like the table below. Every figure in it can be re-derived from the rate sheet, which is the point.
| Line | Basis | Amount |
|---|---|---|
| Strip-out and waste removal | 10 h labourer at $34, plus skip $360 | $700 |
| Tiling labour | 14 m² at $62 | $868 |
| Tile supply (provisional sum) | 14 m² at $45, plus 10% waste | $693 |
| Shower tray and screen, supply and fit | $520 supply, 4 h at $52 | $728 |
| Vanity unit and basin, supply and fit | $410 supply, 3 h at $52 | $566 |
| Plumbing subcontract | Written price | $1,150 |
| Electrical subcontract | Written price | $640 |
| Making good and painting | 12 h at $52 | $624 |
| Subtotal | $5,969 | |
| Margin | 20% of subtotal | $1,193.80 |
| Total | $7,162.80 |
Now the fixes. The commonest slip is the tile line: a draft that shows 14 m² at $45 as $630 has dropped the 10% waste allowance. That single omission turns the total into $7,087.20, which is $75.60 short once the margin is added. If the same slip repeats on 30 quotes a year, that is $2,268 of work given away without anyone noticing.
Check three more things. Every line the notes could not price should say RATE NEEDED rather than carry a number. The damp stud wall and the possible chipboard floor belong under assumptions or exclusions, not buried inside a price. And the subcontract lines should carry your margin once, not twice. If you asked for a shorter version and a line disappeared, put it back: a shorter quote that has quietly lost an exclusion is a worse quote.
The tutorial on checking a quote rather than writing it with ChatGPT goes deeper on the habit that keeps the numbers yours.
Job two: a scope of works the client can sign
The scope is the document clients argue about months later, so it rewards an assistant that asks questions rather than one that fills silence with confident prose. The prompt tells it to park anything the notes do not settle under a heading for you, which is where the real value sits.
Using the quote and job notes, write a scope of works a homeowner can understand.
Sections: what we will do (numbered); what is not included; what we are assuming;
what we need from the client and by when; working hours and access; how changes are
handled; payment stages.
Plain English, no legal wording, no guarantees, no dates unless I give them.
If the notes are silent on something a client might assume, list it under
CHECK WITH THE CLIENT instead of deciding for me.
On this job a good draft includes the tiles being a provisional sum until the client chooses, the WC move as a separate numbered item, and the quiet-hours rule under working hours. It should list, under exclusions, anything found behind the stud wall, redecoration outside the bathroom and moving the client's belongings. Under Check with the client it should raise the questions you did not think to write down: who isolates the water, where the skip can stand, whether the client wants the old bath kept for salvage.
Then run the cross-check, which is the most useful minute in the whole test. Paste the quote and the scope into the same chat and ask:
Compare the quote and the scope of works. List every item that appears in one
but not the other, and every quantity, material or allowance that differs.
A mismatch here is how a firm ends up building something it never priced. In the sample job, a scope line saying "waterproof the shower walls" with no matching quote line is the classic find. Strike or price it before the client sees either document.
Finally, read the scope for sentences that behave like guarantees. "Work will be completed within 10 working days" is a promise unless it is conditional on deliveries and access. "Fully compliant with all regulations" is a claim an assistant has no way to know. Replace both with wording your insurer or solicitor has already approved.
Job three: the awkward email about a delay and a variation
This is the best test of judgement, because tone and omission both matter and no calculator can mark it. Here is the client message, followed by the facts you would feed in.
"Nobody came Tuesday and the tiles still aren't here. You told me two weeks. Now you're saying there's rot? I want a firm date and I'm not paying extra for your surprises."
Draft a reply to the client message below. Facts: the tile supplier has delayed
delivery; the supplier's written new date is [date]; we found soft chipboard under
the old bath. Repair option A patches it ($380); option B replaces the 2 m2 section
($640). Shower work can continue; tiling cannot start until the repair is done.
Tone: calm, specific, no blame, no jargon. Do not apologise for anything I have not
listed. Do not promise a finish date; give the earliest the work could finish if the
repair is approved by [date]. Ask for written approval.
CLIENT MESSAGE:
[paste]
A first draft from a decent assistant should open with the facts, not feelings, something like this: "Hi [first name], the tiles are now due on [date], confirmed in writing by the supplier, and the shower work carries on in the meantime. When we lifted the old bath we found soft chipboard underneath, which is not something we could see before opening up. There are two ways to repair it..." It then lays out options A and B with their prices, states the earliest finish if you approve by the date given, and asks for a written reply. Sign-off: [your name].
What you would fix in that draft: remove any unprompted apology for the supplier's delay if your contract allocates that risk differently; check every date against the supplier's email; shorten the opening paragraph; and make sure the approval deadline is stated once, plainly. An assistant tends to soften the cost message with phrases like "unfortunately we will need to charge", so delete those and let the options speak. If a variation is likely to be disputed, ask your solicitor how delay notices should be worded under your contract before you send anything. For a standard set of replies, five customer messages to draft and check goes further, and turning site messages into change orders covers the paperwork that follows an approval.
Where the draft lands differs by product. Gemini on a Workspace plan drafts inside Gmail, and Claude's Gmail connector can create drafts and replies in your mailbox as well as send them, so keep it to drafts and do the sending yourself. Whatever the assistant writes, you press send. The email carries your name and your legal position, and the draft is only a first pass.
Plans and prices side by side, as read from the vendors' pages
These are USD list prices as read in October 2026. They move often, so confirm on the vendor pages before you buy: ChatGPT Business, Claude and Google AI plans.
| What you need | ChatGPT | Claude | Gemini |
|---|---|---|---|
| Individual paid plan | Plus $20 a month (Go $8; Pro from $100) | Pro $20 monthly, or $17 a month billed annually | Google AI Pro $19.99 a month (AI Plus $4.99) |
| Team plan | Business: $20 a user a month billed annually, $25 monthly, from 2 seats | Team Standard: $20 a seat annually, $25 monthly, 2 to 150 seats | Included in Google Workspace: Starter about $7 and Standard about $14 a user a month on annual billing |
| Where it works inside your email and files | Chat and uploads; apps (formerly connectors) link other tools | Chat and uploads; Gmail and Calendar connectors can write drafts, replies and events | Starter: Gmail and the Gemini app. Standard and above: Docs, Sheets, Slides, Meet, Drive and Chat |
| Saved standard wording | Projects; file limits Free 5, Go and Plus 25, Business 40 | Projects; Free allows up to five projects | Gems (becoming a skill from November 2026) |
| Sharing that saved wording | Project sharing on every plan; Business adds group sharing | Project sharing on Team and Enterprise only | Gem sharing lets recipients see its instructions and files |
| Training on your content | Business: none by default. Individual plans: a switch in privacy settings | Team: none by default. Free, Pro and Max: a switch in privacy settings | Workspace: none by default. Consumer app: privacy controls |
Two practical notes. First, the two-seat minimum on ChatGPT Business and Claude Team means a one-person firm either pays for two seats (about $40 to $50 a month) or stays on an individual plan. Second, upload limits differ and matter for plans and photos: Claude accepts up to 20 files per chat at 500 MB each, ChatGPT takes images of up to 20 MB each, and the Gemini app takes around 10 files per prompt. Large drawing sets may need splitting whichever you pick.
Scoring your own trial: a one-page rubric
Resist choosing on feel. Score each assistant out of 100 on the same sheet, using the same three jobs.
| What you check | Points | How to score it |
|---|---|---|
| Totals and quantities correct | 30 | Lose 10 for each wrong figure you find |
| Stuck to your rate sheet | 20 | Lose 10 for each invented price or rate |
| Gaps flagged, not filled | 15 | Full marks if RATE NEEDED and CHECK WITH THE CLIENT appear where they should |
| Tone of the email | 15 | Paste the three emails into one document marked A, B and C, and ask your foreman or office manager to rate each from 0 to 15 without knowing which is which |
| Minutes to something you would send | 10 | Under 10 minutes: 10. Ten to 20: 5. Over 20: 0 |
| Use from the van | 10 | Dictate the job notes by voice on your phone; 10 if nothing needed retyping |
Here is how a tally might look for three unnamed assistants. Assistant A scores 30 on figures, 20 on the rate sheet, 15 on gaps, 9 on tone, 10 on time and 5 on phone use: 89. Assistant B scores 20, 20, 15, 14, 5 and 10: 84. Assistant C scores 30, 10, 5, 12, 10 and 10: 77.
The reading matters more than the ranking. A beat B by five points, but a single wrong figure costs 10, so that gap is inside the noise and counts as a tie. Break a tie by price and by where your files live. C fell down on the rate sheet and on gaps, the two things that cost money, so it is out whatever its tone score says. As a working rule, do not pay for anything scoring under 70, and treat a gap smaller than one wrong figure as a tie.
Four firms, four different answers
A solo contractor on Google Workspace. Business Starter already includes Gemini in Gmail and the Gemini app, so the cost of testing Gemini is nothing. Standard, at about $7 more per user a month, adds Gemini inside Docs, Sheets and Meet. If Gemini scores 80 or more on the rubric, stop there: running a second paid assistant for the same jobs is spending without a reason. Test one of the other two for a month on a free or monthly plan only if the drafts frustrate you.
Two partners on personal email and a shared van phone. Each needs a login of their own, because sharing credentials breaks the terms of OpenAI and Anthropic. Two individual plans cost about $40 a month. A Business or Team plan costs the same or slightly more ($40 a month on annual billing, $50 monthly) and adds the promise that your content is not used for training by default plus shared project files, so the standard exclusions and rate sheet live in one place. For two people who write quotes daily, I would run ChatGPT and Claude through the rubric and take the Business or Team plan of whichever wins.
An eight-person firm with an office manager, an estimator and three crews. Only the three people who write documents need a seat: owner, office manager and estimator. On annual Business or Team billing that is $60 a month. If the firm is about to move its email to Google anyway, Workspace Standard for all eight comes to about $112 a month and covers email, calendar and storage as well, so the assistant is close to free.
A firm that lives in Microsoft 365. Add Copilot to the trial. Copilot Chat is included with the business plans at no extra cost, and the paid Copilot licence adds reasoning across your own emails and files, so test the free tier first. For a wider look at how the three compare outside trades work, see the general ChatGPT, Claude and Gemini comparison.
Mistakes all three share on construction paperwork
These came up in the sample job, and each one reads as perfectly sensible on screen.
- Invented quantities. A draft says 15 m² of tile because it added wall area to floor area twice. Always ask for the working: length times height for each wall, listed.
- Supply and fit blurred. "Vanity unit, $566" can mean supply only or supply and fit. State which in the prompt and in the line description.
- Quiet promises. "All permits and inspections included" appears where the notes were silent. Anything the notes do not mention should be excluded, not assumed.
- Confident rule-quoting. An assistant will cite a standard or clause number with total assurance and sometimes be wrong. Ask for the reference, then look it up yourself, or leave rule references out.
- Mixed units. If your notes switch between metric and imperial, a total can mix them silently. Put one unit system in the prompt.
- Calendar against working days. "Ten working days" and "two weeks" are not the same promise. Choose one and state it.
- A concession in a polite sentence. "We understand this is unacceptable" admits more than you meant. Read emails for what they concede, not for how they sound.
A short red-pen routine before anything leaves the office: recompute the total on a calculator, scan every date, scan every adjective like "fully" and "all", read the exclusions aloud, then check the quote and scope agree. Five minutes, and it catches nearly everything above.
What to keep out of a consumer chat window
Consumer plans of ChatGPT and Claude let you switch off the use of your chats for model training in privacy settings, and the setting names change, so look for the model-training switch rather than a label you read once. The business plans (ChatGPT Business, Claude Team, Gemini in Workspace) do not train on your business content by default. Either way, replace client names with [first name], drop full addresses, and leave out alarm codes, key-safe numbers, bank details and insurance policy numbers. ChatGPT's Temporary Chat is not used for training, but a copy may be kept for up to 30 days for safety. Deleting a chat does not erase the provider's copy straight away either.
For contracts and anything a client would not want a stranger to read, use a business plan or leave it out. If your quotes mention a client's circumstances, swap them for a code until the final document.
A yearly cost sum for a five-person firm
| Option | Working | Per year |
|---|---|---|
| Five ChatGPT Business seats, annual billing | 5 x $20 x 12 | $1,200 |
| Five Claude Team Standard seats, annual billing | 5 x $20 x 12 | $1,200 |
| Three seats for the three who write documents | 3 x $20 x 12 | $720 |
| Google Workspace Standard, five users | 5 x $14 x 12 (about) | $840 |
| Upgrade from Workspace Starter to Standard | 5 x $7 x 12 (about) | $420 |
Set that against time saved. Say the assistant takes 25 minutes off each of six quotes a week. That is 2.5 hours a week; at an assumed $35 an hour for the person doing the quoting, it is about $87.50 a week, or $4,550 across 52 weeks. Replace my assumptions with your own, but if you cannot show 25 minutes saved during the trial, do not pay yet. For the wider office-software picture, see an office stack for a five-person firm and free against paid AI tools for a small trades business.
Picking after the trial
Take the highest score. If two are within a single wrong figure of each other, take the cheaper one, and if cost is also level, pick the one that sits where your files already live: Gemini for Google, Copilot for Microsoft, otherwise whichever team plan fits your seats. Start on monthly billing and switch to annual only after a month of real use, because the annual saving (about $3 a month on Claude Pro, $5 a seat on the Business and Team plans) is smaller than the cost of paying for the wrong tool for a year.
Save the three prompts, the rate sheet and your standard exclusions in a Project, or in a Gem if you chose Google, so every new quote starts from the same wording. Then rerun the rubric in six months. These products change every few weeks, and a ranking that held in October will not necessarily hold in April.
Questions contractors ask before paying for an assistant
Can ChatGPT, Claude or Gemini price a job for me?
Not reliably. They do not know your labour rates, your suppliers or what materials cost this month. Give the assistant your own rate sheet, tell it to write RATE NEEDED where a figure is missing, and recalculate every total yourself before the quote leaves your hands.
Can two partners share one paid login?
No. OpenAI's terms and Anthropic's consumer terms both bar sharing account credentials. Use a Business or Team plan (two seats minimum), two individual plans, or Google Workspace, where each person has their own account and Gemini comes with the plan.
Is a free plan enough to run the trial?
Usually yes for a one-week trial, because the three jobs are short. Free tiers cap usage and change often, and consumer plans have a privacy switch for model training, so keep client names, addresses and access codes out of the test.
Can these assistants read site photos and drawings?
They can look at photos and PDFs within each app's file limits. Treat any dimension an assistant reads off a drawing as a guess until you have checked it against the figure written on the plan, and never let a photo stand in for a site measure.
Should Microsoft Copilot be in the trial too?
If your email and documents already live in Microsoft 365, yes. Copilot Chat comes with business plans at no extra cost, so adding it costs nothing but time. Run it through the same three jobs and score it on the same one-page rubric.
Further reads
- Which AI Tools Should a General Contractor Buy? What to Skip — Which paid AI tools a general contractor should skip.
- AI Voice Agent for Contractors: What It Handles and What It Hands Off — What a phone agent handles and what it passes to you.
- AI Apps for Contractors: A Phone-Only Setup for One-Truck Firms — A phone-only setup if you run one truck.
- ChatGPT Productivity Tool for Plumbers: Free, Plus or Business? — Free, Plus or Business plans, priced for a plumbing firm.
- AI Marketing Agency for Contractors: What to Pay and What to Ask — Published agency prices, the ChatGPT do-it-yourself bill, a worked lead audit and the contract and reporting terms a contractor should insist on.
- Best AI App for General Contractors: Site and Office Tools Compared — Site and office AI tools priced per month for a six-person firm, each with a worked example, plus four stacks from $189 a month.
- AI Marketing for Contractors: Reviews, Case Studies and Local Posts — Three AI marketing jobs for a small contractor, built on work you already do: review replies, a case study from each job's photos, and weekly Google posts.
- Best Free AI Tools for Construction: Where Each Free Tier Stops — Free AI tools a small construction firm can use today, with the exact limit on each: job cards, calls, scans, seats and chats, plus a three-month cost example.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: ChatGPT pricing page (chatgpt.com), ChatGPT Business pricing page (openai.com), Claude pricing page (claude.com), Google AI subscriptions page (gemini.google), Google Workspace pricing page, OpenAI help articles on project and file limits, Anthropic support article on file uploads. Prices read in October 2026.