AI for Contractors: Four Jobs to Trial Before You Pay for Anything

Start with four jobs where a mistake is cheap and a free tool already works: replying to enquiries with a quote, catching calls you miss, turning site notes into customer updates, and chasing unpaid invoices. Run each for two weeks against a written success test, and pay for software only where your own numbers show it earns its price.

A free trial answers a narrower question than a vendor demo. It will not tell you which product is best. It tells you whether a job is worth doing with AI in your firm, at your volume, with your customers, and that is the question that decides whether any subscription pays back. A fortnight of your own counts answers it better than any feature list.

Follow me on Instagram@sagnikteaches

Count seven things before you switch anything on

A trial with no starting figures proves nothing, because a good fortnight and a quiet fortnight look the same in memory. Take the baseline from records you already have: the phone log, the sent folder, the diary and last month's invoices. The firm below is an illustrative four-person kitchen fitting business, and every figure is invented to show the arithmetic.

Connect on LinkedInSagnik Bhattacharya

If you cannot find a number, that is your first finding. Start writing it down today, even before the trial. Whatever you count, count it the same way afterwards, or the comparison means nothing.

Subscribe on YouTube@sagnikteaches

Trial 1: reply to every enquiry within the hour, inside your rules

The slow part of quoting is rarely the price. It is the reply that waits until the evening because nobody had ten minutes to write it. A free chatbot account can draft that first reply in seconds if you give it your rules, and you keep control by reading it before it goes. Pick one chatbot (ChatGPT, Claude or Gemini all have free plans) and stay with it for the fortnight so your log is comparable.

The prompt below holds the firm's rules. The rates are illustrative, so replace them with your own, and keep the rule that unknowns go in square brackets, because a chatbot will otherwise invent a plausible figure.

You draft replies to customer enquiries for a four-person kitchen fitting firm. Use only the rules below. If something is not covered, put it in [square brackets] for me to fill in.
Rules: never give a final price without a site visit. We can give a guide for fitting only (not units or worktops): $95 a base unit, $65 a wall unit, $110 a metre of worktop. Plumbing and electrical moves are priced after the visit. Visits are free and take about 45 minutes. Ask for: number of units, metres of worktop, whether units are supplied, whether plumbing or electrics move, preferred dates. Never promise a start date. Sign off as [your name]. Keep it under 120 words.
Enquiry: "Hi, we've bought a kitchen (12 units and worktops) arriving in March and need it fitted. The sink is moving to the other wall. What would you charge?"

A typical first answer (illustrative):

Hi [first name], thanks for getting in touch, and congratulations on the new kitchen. For fitting only, my guide is $95 a base unit, $65 a wall unit and $110 a metre of worktop. Moving the sink means plumbing, which I price after a visit because it depends on where the pipes run. Could you tell me how many base and wall units there are, how many metres of worktop, and whether any electrics move too? I can visit for about 45 minutes, free of charge, on [day] or [day]. We'd be happy to start the week your kitchen arrives. [your name]

The reply is useful and it contains one fault: "we'd be happy to start the week your kitchen arrives" is a promise about timing that the rules forbade. Delete it, and add to the prompt "Do not say or imply when we can start." Most first trials of a quote prompt turn up one of these. The model wants to sound helpful, and helpful sentences are the ones that commit you.

• Free tool: any free chatbot plan, plus your own email or messaging app to send the reply.
• Success test: at least 8 of 10 replies sent within 60 minutes during working hours; nothing in any sent reply that your rules did not contain; and visits booked at least as often as your baseline, here 4 of 12.
• Stop sign: a reply that went out with a price you did not set, a promise about dates, or the wrong type of job. One of those is enough to rewrite the prompt. Two in a week means this job needs a person.

The trial usually shows a second thing: your own rules were never written down. A rate card, an exclusions list and a standard set of questions are worth having even if you never use AI. For a full quote layout, the kitchen installation quote template with scope and exclusions goes through the wording line by line.

Trial 2: answer the calls you cannot take, free for the first 25

Missed calls are the one job where free software does something a chatbot cannot, which is pick up the phone. Smith.ai's pricing page lists its AI receptionist plan at $0 for 25 calls a month and $3.00 a call after that. The free plan includes 24/7 call answering, recording, transcripts and summaries, lead qualification, scheduling (Calendly and others) and a test mode it calls Quality Studio. The page does not say how a summary reaches you or whether calls past 25 bill without your say-so, so find both answers in the sign-up screens before you connect your number.

1. Count first. Last month's missed calls from your phone log, and how many left a voicemail. That is your baseline.
2. Write the script. The questions in your order (name, mobile read back to the caller, what they want done, where, how soon), what it must never do (quote a price, promise a date, say a fault is fixable over the phone), and one emergency rule, such as putting a burst-pipe call straight through to your mobile.
3. Test before real callers do. Use the test mode, then ring it yourself from another phone with the five calls in the table below.
4. Forward only unanswered calls. Set your mobile to forward calls you do not pick up, so you stay in charge of the ones you can take.
5. Act on every summary. The trial only counts if you ring people back within two hours.

Run each scenario twice, once calmly and once as a mumbling caller on a bad line, because real callers are the second kind. Fix the script after each failure and test the same call again.

• Free tool: the free receptionist plan; no software to buy, but a monthly cap to respect.
• Success test: in the fortnight, at least 80% of real call summaries have the right name, number and job type; every caller is called back within two hours; and no call ends with the AI having promised a price, a date or a repair.
• Stop sign: a caller misled about anything, or a wrong number that cost you a job. If you hit 25 calls inside a week, that is not a failure: it is a volume signal, and the next step is to price a paid plan against what those calls were worth.

The longer comparison of paid options, with 100 calls priced across five vendors, is in AI answering service for contractors. For now the trial only needs to answer one question: do the calls you were missing turn into visits?

Trial 3: turn a day's scribbled notes into a customer update

Many calls a contractor takes are customers asking what is going on. A short written update at the end of each key day stops most of those calls, and it does something more valuable: it puts extras in writing the day they happen, while everyone remembers. Updates usually go unwritten because composing them costs twenty minutes at the end of a long day. Typing or dictating rough notes into a chatbot and letting it tidy them costs a few.

The prompt rule that matters is the same as before: use only what is in the notes, put gaps in brackets, and leave out anything marked skip. Here is a day's raw notes, with the draft that came back (both illustrative):

Notes: tues - units 1 to 8 in. waste pipe under sink had to be re-run, approx 2 hrs extra. wed - units 9 to 12 + island. worktop template thurs 10am. customer asked about moving two sockets above worktop, needs sparky, not quoted. skip: dishwasher not delivered, customer chasing.

Hi [first name], a quick update on the kitchen. All twelve units and the island are now in. Moving the waste pipe under the sink took about two hours longer than planned, and I'll add that to the final bill at our hourly rate unless you would like to talk it through. The worktop template is booked for Thursday at 10 am. You asked about moving two sockets above the worktop. That needs an electrician and I haven't priced it yet, so I'll send you a figure by [day]. Your dishwasher still hasn't been delivered, so please chase the supplier. [your name]

Two fixes and one check before sending. The dishwasher sentence goes, because the notes said skip, and models often include what they were told to leave out. "At our hourly rate" is a commercial term the AI invented: replace it with how you actually charge extras, as set out in your quote. The check is the count: "all twelve units and the island" matches the notes here, but counts are what drift, so compare every number with what you wrote. The value of the draft is the extra: it recorded two hours of unplanned work in writing, on the day.

Here is the same job as a before and after, to show what changes in practice:

• Free tool: the same free chatbot as Trial 1, with notes typed or dictated on your phone's keyboard.
• Success test: six updates in the fortnight, each sent within 10 minutes of the notes; no invented fact in any of them; no customer phone call asking what a message meant; and at least one extra recorded in writing that would otherwise have been agreed by phone.
• Stop sign: more than half your old time spent correcting the drafts, or any update that stated something false about price, dates or the state of the work.

If you write updates after a longer site day, can AI write a construction daily report from voice notes covers a fuller version for larger jobs.

Trial 4: invoice nudges that go out on the day, in your wording

Late payment on a small contractor's books is often not a customer refusing to pay. It is a reminder that goes out a week late, or never, because the day got away. Invoicing and job software often has reminders built in (Jobber's pricing page, for one, lists automatic invoice reminders by email and text with a one-click payment link), so check what yours already does before paying for anything extra. The chatbot's part is the wording of a ladder that sounds like you.

The hold list is the safeguard that matters. Before any message goes, glance at a column in your sheet marked "query open". A firm reminder on an invoice the customer has already disputed turns a small disagreement into a bad one. The seven-day message, drafted by the chatbot and edited by you (illustrative):

Hi [first name], invoice 2217 for $4,350 (kitchen fitting, completed 3 March) was due on 17 March and still shows as unpaid on our side. Could you let me know when to expect payment, or tell me if anything on it needs correcting? The payment link is below. Thanks, [your name]

Never let a chatbot add interest charges, collection threats or legal language. Those are decisions for you and, where it matters, an adviser, not wording to be generated. The longer treatment of tone, with samples, is in chasing unpaid invoices with AI: polite reminders for trades.

This is the one trial two weeks may be too short for. If your terms are 14 or 30 days, an invoice reminded today may not be paid within the fortnight. Judge it first on proxies you control, and extend it to six weeks if the other three pass.

• Free tool: the free chatbot for wording, plus the reminder feature in whatever invoicing tool you already use.
• Success test: every reminder on the ladder sent on its day; none sent to an invoice on the hold list; and, over six weeks, the average days from invoice to payment below your baseline, here 17.
• Stop sign: any customer complaint about tone, or a reminder sent to a disputed invoice. Fix the hold list before you restart.

The two-week log: one line per use, kept the same way every time

Keep the log in a notes app or spreadsheet, one line per use: the day, which trial, minutes spent (including checking), what happened and what you changed. Minutes spent are the column people forget, and it is the one that decides whether a job saves time. Stagger the starts, so you are never learning four things on one day. The excerpt below is the kitchen fitter's log (illustrative).

Notice how many lines say "change". That is the log working: each change is a rule the prompt or the script did not have, and by day 14 the early failures should have stopped repeating. A failure that keeps coming back after a fix means the job is the wrong one for AI. For a general method for any tool, not only these four, see how to run a two-week AI tool trial before you commit.

Turning two weeks of numbers into a yes or no

Paying is justified only when a free tool ran into a limit that cost you something countable. A good result on its own is a reason to keep using the free version. Three tests, all of which must pass before you pay for anything:

1. The trial passed its success test in both weeks, not only the first. A first week of novelty and close attention tends to flatter any tool.
2. The free version hit a named limit. A cap you reached three or more times, a feature it lacked, or a data rule that stopped you pasting what you needed. "It would be nice" does not count.
3. The cautious monthly gain from removing that limit is at least three times the monthly price. Cautious means the gain measured in the fortnight counted once for the whole month, which is the same as doubling it for a month and then halving it. The factor of three is my rule of thumb, not a standard. It leaves room for your own set-up and checking time, for a fortnight that was luckier than usual, and for the cost of cancelling.

Here is the kitchen fitter's fortnight scored against those tests. The margins and visit rates come from the baseline table, and every figure is illustrative.

On those figures the answer is to pay for nothing, even though the trials found about $2,700 a month of cautious gain. The gain came from the rules, the scripts and the habit of replying fast, and none of it needed a subscription. Three booked visits is also a small sample. One fewer would halve the call gain, which is why the rule asks for a margin of three, and why any firm with a thin month should extend the count to four weeks.

The calls row has a sum that is worth doing with your own volume, using Smith.ai's published figures: the free plan is $0 for 25 calls and $3.00 a call after that, and the Pro plan is $150 for 75 calls with overage at $2.50 on its smallest tier. At 28 calls a month the free route costs $9 if overage is billed. At 60 calls it costs $105, which still beats $150. At 75 calls both come to $150, and above that Pro is cheaper per call. So at this volume the reason to move to Pro would be a feature the free plan lacks, not the price per call. The page lists custom SMS branding as a Pro feature, for instance, but check what each plan includes before assuming it solves your problem. Goodcall, another receptionist, charges $79 a month for 100 unique customers, which its page equates to about 150 calls.

When a step-up does pass the three tests, it usually looks like one of these:

• Job software, as a fifth trial. Jobber's pricing page offers 14 days of full access to the Grow plan's features with no credit card needed, and says nothing is deleted if the trial ends before you add billing. It does not say whether Jobber AI or the Receptionist is included, so ask before you rely on them. Run it against the same baseline and the same three tests, and compare what it replaces (the chatbot, the reminders, the sheet) with its monthly price.
• A business chatbot plan, when the named limit is a data rule or a second user. ChatGPT Business lists $20 a user a month billed annually ($25 monthly) for teams of 2 to 200, with no training on business data by default. Claude Team lists the same $20 and $25 for 2 to 150 seats, with no model training by default. A two-seat minimum is $40 to $50 a month, so it has to earn more than $120 to $150 of cautious monthly gain. For one person, Claude Pro is $20 a month ($17 billed annually). In the kitchen fitter's case, scrubbing customer details from six updates saves nothing like that, so the firm keeps scrubbing.
• A paid phone line, when volume is far past the cap and your call-back rate stays high, which the log will show.

What free costs you: caps, privacy and the day you outgrow it

Free plans cost nothing in money and a little in everything else. Three things to watch from the first day, all of which belong in your log.

• Caps. ChatGPT's free plan lists text chats as unlimited, subject to its abuse guardrails, but file uploads, image creation and memory as limited, with no counts on that page (OpenAI's help centre puts free file uploads at three a day). Other free plans have limits too, and vendors change them. Write down each day you hit one and what you were doing, because three such days in a week is the named limit that the second pay test asks for.
• Privacy. Claude's pricing page lists model training as opt-out on its individual plans, Free included, and none by default on Team and Enterprise. Whichever chatbot you use, look at its data setting before the first prompt, and keep customer names, phone numbers, full addresses and contract prices out of a personal account. Swap them: "Mrs [surname], 14 [street]" becomes "Client A, a two-storey house", and a quote total becomes "about $9,000". The model needs the shape of the job, not the person.
• Reliance. If the free tool changes its limits or a plan disappears, the job should still work by hand. That is one more reason to write your rules down in a document and not leave them only in a chat history.

The data question is bigger than a paragraph, and the further reads at the end of this tutorial cover what free and business plans do with what you paste in.

Four things to keep out of the trial, free or paid

• Prices from photos or plans. A chatbot will give a confident total from a picture of a kitchen, and neither it nor a free app knows your costs, your suppliers or what is behind the wall. If you want to test estimating, do it later, against jobs you have already priced; an AI estimator against a human estimator sets out how to compare them.
• Safety paperwork. Method statements, risk assessments and certificates carry a signature. The person who signs is answerable for every line, so do not hand the first draft to a tool in a trial where you are watching something else.
• Anything sent without being read. Nothing in these trials goes to a customer unread. Auto-sending is a decision for later, after a log that shows a clean month.
• Real customer details in a personal account. Use the swaps above until you have moved to a plan whose data terms you have read.

Four jobs in two weeks, each with a pass mark, is a small experiment, and it should end with a decision you can defend with figures: keep, fix, extend or pay. Often the answer is that the habit was the gain and the subscription was never needed. If you would rather set up the trial and the log with someone, that is what my AI implementation consultation covers.

Questions contractors ask before a free AI trial

Is a free chatbot enough for a small contractor?

For writing jobs at low volume, usually yes. ChatGPT's free plan lists text chats as unlimited, subject to its guardrails, but uploads, image creation and memory as limited, so note the day you hit a cap. Keep customer names, numbers and addresses out of personal accounts, and move to a business plan only when the log shows you need one.

How many calls does a free AI phone plan cover?

Smith.ai's free AI receptionist plan lists 25 calls a month, then $3.00 a call, on its pricing page as checked in October 2026. Before connecting your number, find out in the sign-up screens whether calls past 25 bill automatically, and count your real missed calls first so you know whether 25 is enough.

Can I run all four trials at once?

Yes, if you stagger them. Start the quote replies and site-note updates on day one, the phone line once your test calls pass, and the invoice nudges around day five. Keep one log, and stop any trial the moment it hits its own stop sign rather than waiting for the fortnight to end.

What if I get too few enquiries to judge in two weeks?

You need roughly ten real enquiries and ten real calls to learn anything. With fewer, extend the trial to four weeks or run every old enquiry from your sent folder through the quote prompt and compare the draft with what you actually sent. Invoice trials usually need longer, because payment lags behind the reminder.

Further reads

• Free vs Paid AI Tools for a Small Trades Business — Which free tools hold up for a small trades firm, and which do not.
• AI Assistant for Contractors: ChatGPT, Claude and Gemini on Real Jobs — Three chatbots tried on real contractor jobs.
• AI Apps for Contractors: A Phone-Only Setup for One-Truck Firms — A phone-only AI setup for a one-truck firm.
• AI Automation for Contractors: Qualify, Book and Confirm Site Visits — Qualify, book and confirm site visits without ringing round.
• How to Measure an AI Receptionist's Return in the First 90 Days — How to measure an AI receptionist over a longer trial.
• Does ChatGPT Train on Client Data? Business vs Free Plans — What free and business plans do with the data you paste in.
• AI Software for Contractors: An Office Stack for a Five-Person Firm — What a five-person contracting firm's office stack looks like.
• How Can AI Help General Contractors? Ten Jobs, Ranked by Effort — Ten jobs for a small contractor in three effort tiers, each with a worked example and the check that keeps a person in charge of the result.
• AI for Concrete Contractors: Estimating, Pour Scheduling and Weather Calls — Where AI helps a concrete firm on estimates, pour scheduling and weather days, and the calls that stay with the foreman, with templates.
• AI for Garage Door Companies: What It Can Take Off a Small Team — Five jobs AI can take off a small garage door team, from missed calls to reviews, with the safety limits and what each one costs a month.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Smith.ai AI Receptionist pricing page, Goodcall pricing page, Jobber pricing page, ChatGPT pricing page, ChatGPT Business pricing page, Claude pricing page. Checked October 2026; prices and plan limits change, so confirm them before you buy.