Best AI for Builders: Pick the First Job by Hours Lost and Risk

Hand AI the job that costs you the most hours a week and does the least harm when it gets something wrong. For most small building firms that is missed calls, then quote follow-up. Chasing payments and site notes come next, and estimating comes last, because a wrong number costs real money.

Two clarifications before the scoring. "Builders" here means a building firm with vans, subcontractors and a phone that rings while you are up a ladder, not the AI app builders that crowd some search results (my tutorial on a Glide snag-list app covers building your own tool, which is a later step). And "hours lost" means the hours a job eats today plus the hours it would take if you did it properly and currently do not.

Follow me on Instagram@sagnikteaches

Score each job: hours lost times safety

In its simplest form the rule is hours lost against risk of error. I turn risk into a safety rating from 1 to 5 so that a bigger score is always better: 5 means a mistake is caught before anyone outside the firm sees it, and 1 means a mistake sets a price, a promise or a safety control. Multiply hours lost a week by the safety rating. The highest score is your first job.

Connect on LinkedInSagnik Bhattacharya

The rating belongs to the version of the job you start with, not the job in general. A missed-call text that you wrote and approved is a 4. A live AI voice answering questions is a 3. Same job, different risk, so begin with the safer version and move up once it has run cleanly for a month.

Subscribe on YouTube@sagnikteaches

Scores ignore urgency, so one override applies. If a large sum has been overdue for more than 30 days and cash is tight, chasing payments goes first whatever the score says. It is a low-risk job anyway, because the amounts come from your accounts and the AI only changes the wording.

Measure the hours first, because owners guess badly

Keep a two-week tally in a notes app: one line per task with the date, the job type, the minutes and what went wrong. Include the jobs nobody has time for, as "not done" lines with an honest guess at the minutes they would take, for example "quote chase, not done, 12 min". Those lines are where the hidden hours sit.

If the lines are messy, an assistant can classify them, provided you do the adding up yourself. A prompt that works:

Below is a two-week tally from a six-person building firm. Each line is: date, job type, minutes, note.
Classify every line as one of: missed calls, quote follow-up, estimating, chasing payments, site notes, supplier and sub chasing, other.
Then give the total minutes per type and the average hours per week (divide by 2).
List the lines you counted under each type so I can check the sum.
If a line fits two types, put it under the first and tell me.
Do not estimate any minutes that are missing.

[paste the tally here]

The firm I will use through the rest of this tutorial is an invented six-person building firm doing extensions and loft conversions: the owner, who also prices every job, a site manager, three carpenters and an office assistant who works two days a week. Here is its tally after the classification:

The raw output needed two corrections. The model filed three "rang back after the text" lines under "other", and on a second run it totalled the calls at 640 minutes where the lines add up to 600. Chat assistants add long lists unreliably. Use the AI to classify, paste the classified lines into a spreadsheet and let a SUMIF do the totals. The table above is the corrected version.

The six jobs, scored for one firm

Now apply the rule. The third column is the safest version of the AI job, because that is the version you would start with.

Missed calls win with 20. The phone log shows about nine missed calls a week, and the tally understates the damage, because the callers who never left a message are not in it. Quote follow-up is second at 15 and is the safest job on the list. Estimating has the most hours of any job on the list, yet it ties for last, which surprises most owners. Payments are low on hours, but the cash override above would bump them to the front if the firm had a large overdue balance.

Before you trust a ranking, stress-test it. Move each safety rating up or down by one and see whether the top job changes. Here, if the firm starts with a live AI voice instead of a text-back, calls drop from 4 to 3 and the score from 20 to 15, a tie with quote follow-up. When two jobs tie, take the one where money is leaking fastest, which is calls. If the top job survives every one-point change, the answer is solid. If it flips, you are choosing between close options, so decide on cash and on which job you would be least embarrassed to get wrong.

A quick sum shows why the first job matters even at this size. The firm loses 8 hours a week to calls and quote chasing. At a loaded internal cost of $40 an hour (an assumption, so use your own), that is $320 a week, or about $16,600 a year. If a calmer routine halves it, the saving is bigger than the $240 a year a chat assistant costs at a $20 list price, before counting any job won back.

Missed calls: start with the text-back, not the robot

Version one costs almost nothing and rates a 4. When a call goes unanswered, the caller gets a text within a minute, in wording you wrote:

Hi, it's [firm name]. We're on site and can't pick up. Reply with your name, what you'd like done and the best time for a call back, and [owner first name] will ring you by 6pm today. If it's an emergency such as gas, fire or flooding, please ring the emergency services first.

Replies and voicemails then go to an assistant that turns them into a morning list. Here is a sample output from three voicemails, with the problems I would catch before ringing anyone back:

1. [first name], new enquiry: single-storey rear extension, wants a site visit, best after 4pm. Number transcribed as 555 0142. Check the audio, transcripts mishear digits.
2. [first name], existing client on the kitchen job, asks whether the plasterer is still coming Thursday. Not a new enquiry; forward to the site manager.
3. [first name], new enquiry about "knocking through", budget "fifteen". Unclear whether that is $15,000 or $50,000; listen before assuming.

Version two is a live AI receptionist, which rates a 3 because it speaks for you. Do not buy one until the text-back has run for a month. When you do test one, the free tier is the right place to start. As of October 2026 the Smith.ai AI receptionist lists a free plan with 25 calls a month and $3 a call beyond that, with a Pro plan from $150 a month for 75 calls. Twenty-five calls would cover about three weeks of this firm's nine missed calls a week. My tutorial on AI receptionists for builders on site compares the options properly.

Whichever version you pick, write the guardrails into the script: no prices, no start dates, no site-safety advice, and anything mentioning a collapse, gas or fire gets "ring the emergency services" before a transfer to you. A realistic failure to design against: the assistant tells a caller the firm "can usually start within two weeks" when the order book is full until March. So the script says only that the office will confirm availability when it rings back. Listen to the first month's recordings, not a sample of five.

Quote follow-up: three messages, written once, sent on a timer

Before: six quotes out, last chased "when I remember", and customers going quiet for three weeks. After: every quote gets a message on day 2, day 7 and day 14, each drafted by the assistant from facts you supply and sent by you. Put the standing instructions in a saved project so you do not retype them:

You write short follow-up messages for a small building firm.
Facts I give you per quote: first name, quote reference, one-line scope, price, how long the quote is valid.
Never invent availability, start dates, discounts or guarantees.
Tone: plain and friendly, no sales language, under 90 words.
Write three messages: day 2 (check the quote arrived), day 7 (offer a ten-minute call and name the one thing customers usually ask about), day 14 (ask whether to hold or release the slot).

A trimmed sample of the day 7 message it returns for a rear extension:

Hi [first name], I wanted to check the quote for the rear extension (ref 1182) answered your questions. The two things people usually ask about are the steel beam supply and what the electrics allowance covers. I'm happy to talk either through on a ten-minute call, so reply with a good time, or tell me if you'd rather wait. [your name]

Trim what the first draft invented: it added "we pride ourselves on quality" and "we have a slot in March", neither of which appeared in the facts it was given. The prompt's "never invent availability" line exists because of that. Send the messages yourself, from your own number or email, so they sound like you and replies reach you.

For timing, use the reminders in your job software if it has them, and a calendar reminder at 2, 7 and 14 days if it does not. My tutorial on ten ChatGPT prompts for quotes, chasers and client emails has more wording to borrow.

The break-even sum is short. Take an average job of $24,000 and a margin of about 15%, which is $3,600 (assumptions, not benchmarks). An assistant at $20 a month costs $240 a year, so one extra job every 15 years pays for it. The real cost is the ten minutes a week it takes to send the messages.

The first week for the top two jobs, in minutes

Nothing here needs a project plan. For the example firm, calls and quote follow-up together fit into a week of short sessions:

• Monday, 45 minutes: write the text-back, switch on the automatic reply or call forwarding, and send yourself a test call from another phone.
• Tuesday, 30 minutes: list every open quote with the date sent, the value and the last contact. That list is the follow-up queue, and it usually shocks the owner.
• Wednesday, 30 minutes: save the follow-up instructions in a project and generate messages for the five oldest quotes. Edit each one, then send it yourself.
• Friday, 20 minutes: read the week's voicemail summaries against the audio, count the missed callers who got a reply, and write one line in the error log for anything the AI got wrong.

That is under two hours, and it touches the two jobs that scored highest. If the Friday review shows the summaries are unreliable or the text-back is annoying people, you have lost two hours rather than a subscription and a month.

Chasing payments: the wording is the risk, not the maths

Payment reminders reach a customer at a sensitive moment, so a badly judged tone can start a dispute that a plain note would have avoided. The figures must come from your accounting package. Let the AI draft only the words, in a three-step ladder: a friendly note two days before the due date, a clear one the day after, and a firmer one at fourteen days that states the amount, the invoice reference and what you need from them.

Here is a filled-in day-after reminder in the tone I would use:

Hi [first name], invoice 2041 for the second stage payment ($14,500) was due yesterday and I can't see it on our side yet. If it has already gone, thank you, and please ignore this. If anything about the invoice needs explaining, ring me and we will sort it today. [your name]

A realistic mistake to expect in a first trial: the draft adds "a late fee of 5% now applies" when the contract has no such clause. So the instruction reads: do not mention fees, interest or consequences unless I paste the contract clause. For the full reminder ladder and the wording for trades, see chasing unpaid invoices with AI.

Site notes: useful, but a diary can end up as evidence

A site diary matters most when something goes wrong: a disagreement over a variation, a delay, a defect. That is why it rates a 3 and why the instruction has to forbid embellishment. The workflow is a voice note when the crew packs up, a draft entry from the assistant, and the site manager signing off the draft the same day. Keep the original audio and photos alongside it.

Turn this voice note into a site diary entry with these headings: Date, Weather, Who was on site, Work done, Deliveries, Requests from the client, Problems, Photos.
Use only what I say. Where I do not say something, write "not stated".
Never add quantities, times, weather or names. A request from the client is not an agreed change.

[paste the transcript here]

A sample note, "Tuesday. Three of us plus the electrician from ten. Roof timbers on, felt half done, rain stopped us at two. Steel arrived at nine, left by the gate. Client came by, wants an extra socket by the sink, I said I'd price it. Photos of the steel and the roof", came back structured correctly except for one line: the draft put "Variation agreed: extra socket" under client requests. The speaker had only said they would price it. Left in the diary, that line could be quoted back at the firm later, so the last instruction in the prompt is there for a reason. My tutorial on client updates and snag lists with AI covers the client-facing side.

Estimating comes last, and a control can move it up

Six hours a week is a lot, and it is tempting to start here. The trouble is the rating: a wrong quantity or a missed item turns directly into lost margin, and nobody outside the firm will catch it. Controls change the rating. If the AI only does the takeoff, you hand-count three items on every job, and a second person checks the totals against the drawings, the rating moves from 1 to 2 and the score from 6 to 12, level with site notes and supplier chasing, and still behind calls and follow-up.

I would keep the rating at 1 until the control has run on at least three live quotes and the AI totals have matched your own. For what accuracy looks like, read whether AI can produce accurate estimates for a small building firm. The AI never sets your margin, your labour rates or the final price.

When the order changes: a firm with an office manager

The method gives different answers for different firms. Take an eight-person firm with a full-time office manager who answers every call, three site managers and two people sharing the estimating. Its weekly tally could look like this:

Here the first job is site notes, then supplier chasing, and calls drop to the bottom because a person already answers the phone. Same rule, opposite answer, which is why "usually calls and quote follow-up" is a starting guess and the tally is the evidence.

Four weeks in: how to tell the first job is working

Pick three checks before you start, and write them down:

• Hours: repeat the tally for that one job. Decide your target in advance, for example down by a third.
• Coverage: for calls, the share of missed callers who got a reply within five minutes; for quotes, the share that received all three follow-ups.
• Mistakes: keep an error log with one line per AI mistake, whether you caught it or a customer did.

Set a stop rule as well. If two mistakes reach a customer in a single week, pause the job, fix the instruction and restart. And keep to one job at a time: add the second only after four clean weeks, so you always know which change caused which result. If you would rather work through the scoring with someone who has seen how it plays out in other firms, that is what my AI implementation consultation covers.

Questions builders ask before starting

Should I use ChatGPT, Claude, Gemini or Copilot for the drafts?

Use whichever you already have open all day. Short messages built from facts you paste in work in all four, so the deciding factor is convenience: Gemini sits in Gmail, Copilot sits in Outlook, and ChatGPT or Claude work best with a saved project holding your tone and standard wording. Pick one for the first job and stay with it for four weeks.

Is it safe to paste customer details into an AI tool?

Strip what the job does not need. A quote chaser needs a first name, a reference and a one-line scope, not an address or phone number. Business plans of the main assistants do not train on your content by default, while consumer plans have a model-training switch in their privacy settings, so check which plan you are on before pasting anything personal.

Do I have to tell callers they are speaking to an AI?

Some places now require it for chatbots and voice agents, and the rules are still settling, so check yours. The safe habit anywhere is to say so in the greeting, for example: you have reached the office assistant for the firm, I can take a message and the team will call back today. It costs you nothing and avoids a complaint later.

Further reads

• Best AI for Building Companies: Quotes, Programmes and Site Admin — Match tool types to quotes, programmes and site paperwork once the first job runs.
• Builder Invoice Template: Stage Payment Requests Written With AI — Stage payment requests, if payments end up as your second job.
• How Small Builders Use AI to Price Variations and Change Orders — Pricing variations once site notes are being captured properly.
• AI Receptionist for Construction Companies: Real Costs and Call Routing — Real costs and call routing before you pay for a live AI voice.
• AI for Building Contractors: Chasing Subcontractors and Progress Updates — The sub and supplier chasing job that scored well in the larger firm.
• AI Consulting for a Construction Company: What a First Project Costs — Three gated phases priced in consultant days for a fourteen-person building firm, plus five candidate first projects and the payback test for each.
• 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 (checked October 2026).