How to Implement AI in Your Business: First Job for a Trades Firm

Score each candidate job on hours lost, money at stake, safety if wrong, ready inputs, ease of measuring and a named owner, then run the winner for two weeks beside your current process: one shadow week where a person reads every draft, one live week with limits, and a success check agreed before day one.

That order is what protects live work. A trades firm cannot pause a boiler install while an experiment sorts itself out, so the first job should be one where a mistake costs a minute of someone's time, not a booking. Big-company frameworks assume a project team and a quarter. A five-person firm has a fortnight of spare attention at most, and one sheet of paper.

Follow me on Instagram@sagnikteaches

What protecting live jobs means in a trades firm

Live work is the stuff that is booked, promised or paid for this week: customers waiting for a visit, an engineer relying on the right parts, a quote someone is about to accept, a call that could be a gas leak. Eight rules keep the pilot away from it.

Connect on LinkedInSagnik Bhattacharya

1. Run beside the old process, never instead of it. For the whole fortnight, the way you do the job today keeps working.
2. Pick a quiet fortnight and a Tuesday start. Not the week before a holiday or the first cold snap, and not a Monday when the diary is already a mess.
3. Change one thing at a time. No new job software, new phone number or new quote template in the same fortnight.
4. Nothing goes out unread in week one. A person reads every draft before it reaches a customer, and in week two they still read it.
5. Keep contact details out of the AI tool. Paste first names and job details, not phone numbers, email addresses or full addresses.
6. Cap the blast radius. Pilot on a slice: quotes under a set value, calls after 6 pm only, one engineer's notes.
7. Write down the off switch. Who turns it off, how, and how long it takes. Two minutes is the target.
8. Name the person on duty each day. If the tool misbehaves at 4 pm on a Thursday, someone must know it is theirs.

The generic version of the shadow-week idea is in piloting AI in shadow mode before customers see it. The rest of this page is about what changes when the firm has vans, a diary and no spare staff.

Subscribe on YouTube@sagnikteaches

The pick-one-job sheet: six questions, six candidate jobs

List every job you could hand to AI, then score each from 0 to 3 on six questions. Do the scoring with whoever does the work today, not alone.

• Hours. How many hours a week does it eat? 0 is under half an hour, 1 is half an hour to an hour, 2 is one to three hours, 3 is more than three.
• Money at stake. What does it cost when it slips? 0 is nothing noticeable, 3 is about a job a week.
• Safe to be wrong. 3 means a person reads the output before any customer sees it. 0 means a mistake reaches a customer or moves money automatically.
• Inputs ready. 3 means the text is already in your job software or email. 0 means it lives on paper or in someone's head.
• Easy to measure. 3 means you already count it. 0 means you would have to invent a measure.
• Owner. 3 means one named person does it today and wants it fixed. 0 means nobody owns it.

Here is a sample sheet for a five-person heating and plumbing firm: the owner, two engineers, an apprentice and a part-time office person. Your own scores will differ.

Quote follow-ups win on 16. The office person spends about two and a half hours a week on them in the evenings, a quote is the nearest thing to money in the firm, the text is already in the job software, and a person can read each draft first. Invoice reminders come second and become the second job. After-hours calls score 9 because the output reaches a customer immediately and nobody clearly owns the problem; that is not a reason to avoid them for ever, only for now. The wider question of customer-facing against back-office first is covered in where AI should go first, customer-facing or back office.

If you want to see what the routes to each job cost, here are the ones I would try first. Prices are list figures from the vendors' pages in October 2026 and change, so confirm them.

The sheet gives different winners in different trades, which is the point of scoring instead of copying a list:

• A roofing firm that surveys on Monday and quotes on Friday may find the winner is chasing quotes after a survey, because the money at stake is high and the notes already sit in the CRM.
• An electrical contractor buried in certificates and paperwork may find that turning site notes into a draft certificate summary scores highest on hours, but it fails the owner and safe-to-be-wrong questions until one person is named to check every draft.
• A landscaper with a spring rush of enquiries may score reply-to-enquiry drafting highest, because the volume is high and a person can still read each reply.
• A painter and decorator who quotes from photos may find the winner is the "send us two more photos" request, which is repetitive, low risk and easy to count.

Vetoes that remove a job before you score it

Some jobs should not be a first job however well they score. I apply four vetoes before the sheet.

• Emergency calls. A call about a gas smell, a burst pipe or smoke needs a person, and the safe pattern for an AI receptionist is "contact the emergency services" first, then a transfer. That is a later job with its own rules, which AI receptionists for construction companies, with real costs and call routing goes through.
• Anything that moves money or makes a legal promise. Refunds, price changes, deposits, guarantees and contract wording stay with a person.
• Anything with no owner. If nobody wants the task fixed, nobody will review the drafts.
• Anything that needs field staff to change habits. Asking engineers to dictate notes differently is a change-management job, and the first fortnight has no room for one.

Two weeks, day by day: shadow first, then live with limits

The pilot assumes the winner is quote follow-ups. Swap in your own job and keep the rhythm.

1. Before day one: take the baseline. From the last four weeks of records, count quotes sent, how many were followed up within three days, how many were accepted, and how many hours the follow-ups took. For the example firm: 25 quotes sent, 10 followed up within three days (40%), 6 accepted (24%), about 2.5 hours a week of office time.
2. Days 1 and 2: set up. Write the prompt, export the list of open quotes (first name, job type, days since sent, value, notes), and save a manual fallback template.
3. Days 3 to 7: shadow week. Each morning the office person pastes the rows into the assistant, reads the drafts, writes the follow-up they would have written anyway, and sends their own version. The drafts are never sent.
4. Day 8: review. Count the drafts accepted as written, edited and rejected, and read every rejection. Change the prompt rules, not the person.
5. Days 9 to 14: live week with limits. The AI drafts, the office person reads each one for thirty seconds and sends it from the normal system. Quotes over $4,000 stay with the owner.
6. Day 14: success check. Compare the fortnight with the baseline using the numbers you agreed on day one.

If your records are thin, the baseline takes one week of tally marks instead: a sheet by the phone or on the office screen, a mark each time a quote goes out, a mark each time one is followed up, and the minutes spent on follow-ups written at the bottom of each day. A rough week of real counts beats a precise guess. Whatever you record, keep the definition fixed: "followed up" means a message or call that is logged, not a thought about doing one.

You draft follow-up messages for a five-person heating and plumbing firm. A person reads every message before it is sent.

Rules:
- Use only the details in the row. Never mention a discount, a deadline, a price change, a guarantee or availability unless it is in the row.
- Under 70 words, plain and friendly, no exclamation marks. Sign off as [office name].
- Message type depends on days since the quote: 2 to 4 days, check they received it and offer to answer questions; 5 to 9 days, add one useful detail about the job that appears in the notes; 10 or more days, a short closing message that leaves the door open.
- If the value is over $4,000, or the notes say declined, write the single word SKIP instead of a message.

Rows (first name | job | days since quote | value | notes):
[paste rows]

A draft, and what the shadow week caught

For a row reading "[first name] | boiler replacement | 6 days | $3,450 | wants to decide after payday", an illustrative draft came back like this:

Hi [first name], following up on the boiler replacement quote we sent last week. Most replacements take a day and your heating is usually back on by the evening, so I can send over the install-day timings if that helps. Happy to answer anything before you decide. [office name]

Three things to fix. "Most replacements take a day" and "back on by the evening" are not in the row, so the assistant has invented them from general knowledge, and the rules said not to. Replace them with your firm's real timings or cut the sentence. The note about payday was ignored, which is right, but the draft should not push a customer who has said they are waiting. And "last week" is a guess at the quote date; the row gave only days, so write "on [date]" from the quote record.

The shadow week also caught a mistake of a different kind. On day 5 the export included quotes with the status "declined", and the assistant happily drafted follow-ups for customers who had already said no. Nobody was embarrassed, because nothing was sent. The fix was two lines: export only open quotes, and add the SKIP rule for rows whose notes say declined. That is the whole point of a shadow week: the errors show up with the customers safely out of reach.

The live week has its own kind of slip. On day 11, say, a follow-up went to a customer who had phoned that morning to book the work, because the office person marks accepted quotes in one batch each evening and the export still showed the quote as open. The customer laughed it off, but the firm treated it as a real error and logged it. Two changes came out of it: a "last updated" column in the export, and a ten-second check of the job diary for the customer's name before any message is sent. A pilot that records slips like this, and fixes the routine instead of blaming the tool or the person, is the one that survives to the handover.

The success check: numbers agreed before day one

Write the thresholds on the sheet before the pilot starts, because a fortnight's result always looks like a pass to a hopeful owner. Pick two process measures, one time measure and one harm measure. Here is the example firm's check.

Twelve quotes cannot show whether follow-ups change the acceptance rate, so treat that line as "no sign of harm" and look again after six to eight weeks. The rule I would use: keep the job if both process measures are met and nothing serious reached a customer; extend the pilot by a week if only one is met; stop if a serious error got through unread or the time measure did not move. Setting success criteria that hold up goes into the harder cases.

Now the money, because it decides whether the second job is worth the effort. The time saved is 1.3 hours a week, about 5.6 hours a month, which at $30 an hour for the office person is roughly $168 against a $20 chat plan. That is a fine result but not the reason to do the job. The reason is the 15 quotes in the baseline month that were not followed up within three days. If one extra quote in a month is accepted, at an average value of $2,400 and a 30% margin, that is $720 of gross profit. I would not promise that, and your numbers will differ, but it shows why "money at stake" is on the scoring sheet.

Handing the job to staff without handing them a mystery

A handover is finished when someone other than you has run the job for a full week while you watched, and has fixed one small problem without calling you. Give them a single sheet. The firm's version looks like this.

• What it does: every open quote gets a follow-up within three days.
• Owner and backup: the office person; the owner covers holidays.
• The daily routine: at 9.30, export open quotes that are two or more days old, paste the rows, read every draft, edit, send from the quote system, and mark the quote as followed up. Ten minutes.
• The rules the AI follows: the prompt above, saved in one place with a date on it.
• Staff may change: the sign-off, the tone words and the day thresholds. Only the owner changes: anything about prices, discounts, the SKIP value and the fact sheet.
• How to pause: stop pasting rows and use the saved manual template.
• Error log: a shared sheet with date, what went wrong and whether it reached a customer.
• Weekly review: fifteen minutes each Friday for four weeks, then monthly.

Run one thirty-minute session. Show the old way, then the new one, then let the person break it with three awkward rows: a declined quote, a customer who has already booked, and a quote with a blank note. For a fuller layout, a one-page runbook for staff has a template, and the same page covers what to do when the person who ran the job leaves.

Say plainly what the job removes from the person and what they keep. They keep the judgement and the final read; the chore of staring at a blank screen at 6 pm goes. Staff who are told that, and who wrote half the rules, are more likely to look after the routine. Staff who find a tool switched on in their inbox are less likely to.

After the pilot: keep, tune, drop, or add the second job

The fortnight ends in one of four places. Keep it if the check passed. Tune it if one measure missed and the error log shows a fixable pattern. Drop it if the time measure did not move or a serious error got through. Or extend by a week if the data was thin, as with the acceptance rate above.

Tuning looks like this in practice. Take the landscaper from the scoring section, piloting replies to spring enquiries. The time measure passed, but only 9 of 20 drafts in the live week were usable as written, because the assistant kept offering a free design visit the firm does not do. The error log showed the same fault eleven times, which is a fixable pattern and not a bad tool. One added rule ("we charge for design visits; never offer one free") and one extra week later, 16 of 20 drafts went out with light edits. That is a tune and a one-week extension, not a restart.

Dropping looks different. Picture a kitchen-fitting firm that skipped the field-habits veto and piloted turning surveyors' voice notes into job sheets. The notes were recorded in vans with the radio on, measurements came through garbled, and the office person ended up replaying every recording to check the drafts, so office time on job sheets rose from three hours a week to four. Nothing wrong reached a customer, which is the shadow week doing its job, but the time measure moved the wrong way. The fix is not a better prompt. It is a fixed notes template the surveyors use for a month first, after which the job can be scored again with inputs ready at 2 instead of 1.

Add the second job only after four weeks of normal running with the owner out of the loop and a quiet error log. Take it from the scoring sheet, not from the latest demo. In the example firm that is invoice reminders at 14, which share the same owner, the same kind of data and the same read-before-send habit, so the handover is a day's work and not a project. The after-hours call job, at 9, needs its own pilot with the emergency rule written first.

If you would rather pick and pilot the first job with someone who has seen the pattern, that is what my AI implementation consultation covers: scoring your own list, choosing the job, and handing over something your team can run.

First AI job for a trades firm: what owners ask

How long should a first AI pilot last in a trades firm?

Two weeks of pilot, after a baseline taken from the last four weeks of records. Run one week in shadow mode, where a person reads every draft and nothing goes out unchecked, and one week live with limits. Then give the job four weeks of normal running before you judge outcomes such as accepted quotes, which need more data than a fortnight gives.

Should the first AI job talk to customers directly?

Usually not. Put a person between the AI and the customer on the first job, so a wrong draft costs you a minute and not a booking. Tools that answer calls or chat with customers make better second or third jobs, once you have seen how the AI behaves. Emergency calls need a human route from the start.

Do I need to buy new software for the first job?

Often not. A chat assistant on a free tier or a plan around $20 a month, plus data already in your job software, covers jobs like quote follow-ups. Check your current software first: ServiceM8 lists AI quote and invoice drafting on every plan, and Tradify puts its AI tools on its Plus plan only.

What if my engineers or office staff do not want to use it?

Start with the job that removes a chore from the person who dislikes it most, and let them write the rules. Do not ask field staff to change how they log work during the first fortnight. Resistance usually drops when the person who does the task reads every draft and can say no.

When should I add a second AI job?

After the first has run for four weeks without you in the loop and the error log has gone quiet. Take the next job from the scoring sheet, not from a vendor's demo. Adding it mid-pilot muddies the results, because you will not know which change caused which effect.

Further reads

• The First 30 Days of AI in a Trades Business — A week-by-week view of the first month, after the pilot.
• Best AI for Small Businesses: What a Three-Van Trades Firm Buys First — What a three-van firm buys first, with prices.
• What Tradespeople Actually Use AI For: 12 Real Jobs It Handles — Twelve real jobs that tradespeople hand to AI.
• Seven AI Mistakes That Cost Tradespeople Jobs — Seven mistakes that cost tradespeople work.
• Chasing Unpaid Invoices With AI: Polite Reminders for Trades — A strong second job: polite invoice reminders.
• When Should a Trades Business Bring In an AI Consultant? — When a trades business should bring in help.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Smith.ai's AI receptionist pricing page, Jobber's pricing page and its help centre on Receptionist allowances, and Tradify's and ServiceM8's pricing pages (all read October 2026). The scoring sheet, pilot plan and example figures are the author's own illustrations.