Six paperwork steps in a construction company suit AI: turning dictated notes into a daily log, filing and summarising it, flagging logged extra work as a variation, capturing bills and hours, checking progress percentages and drafting the invoice's covering note, and wording payment reminders. The site manager signs the log and the owner approves the invoice; AI never does either.
The chain only works if each output feeds the next. A log that records who instructed what becomes a priced variation, the variation becomes a line on the valuation, and the valuation becomes the invoice. Where the chain breaks is usually a missing hand-off, such as an instruction mentioned on Tuesday that nobody priced, not a lack of software.
One tool per step: what goes in, what comes out, who signs
The worked example is a ten-person building firm: the owner, a site manager, a part-time office administrator, five tradespeople and two labourers, running four jobs at once. It uses ChatGPT or Claude for the writing and checking, and JobTread as the job software because its pricing and AI tools are published. Each step has one tool.
| Step | Tool | In | Out | Signed off by |
|---|---|---|---|---|
| 1. Log | ChatGPT or Claude, in a Project holding your log template | Dictated notes | A structured daily log with instructions flagged | Site manager, same day |
| 2. File and summarise | JobTread daily logs and its AI summaries | The checked log and photos | A weekly summary for the team and customer | Site manager reads before it is shared |
| 3. Variations | ChatGPT or Claude, same Project | The week's logs | A table of possible variations and a draft notice | Owner prices it; the client approves |
| 4. Costs | JobTread receipt scanning and time tracking | Photos of bills; timesheets | Bills and hours on the right job | Office checks totals; site manager approves hours |
| 5. Invoice | JobTread customer invoice, or QuickBooks Online progress invoicing | Contract sum, percentage complete, approved variations, payments to date | A draft invoice and covering note | Owner |
| 6. Chasing | Your accounting package's reminders | Overdue invoices | Reminder drafts | Office sends |
Switch the names if you already use other software. Raken's site describes mobile daily reports with notes, photos and videos that sync to a web app and export as PDFs, plus timecards, and CompanyCam turns job photos into reports. Keep what you have; steps 1 and 3 work on the exported text just the same.
Daily log: dictate the day, let ChatGPT or Claude shape it
A daily log fails when it is written late, from memory, in a different style by each foreman. Dictation fixes the lateness, and a template fixes the style. The site manager dictates for two minutes on the way out, using the phone's built-in dictation, and pastes the transcript into a Project that holds your log template. Claude's Projects and ChatGPT's both do this; Claude Free accounts can create up to five projects, and ChatGPT Projects work on every plan. For more on the basics of voice-note logs, see the tutorial on whether AI can write a construction daily report from voice notes.
Turn the dictated notes below into a daily site log in this exact order: Date and job; Weather (as stated); People on site (headcount by trade); Work done, by area; Deliveries; Visitors and inspections; Instructions received (who gave each one, and whether it is confirmed in writing); Delays and their stated cause; Safety observations; Possible variations; Tomorrow.
Rules: use only what the notes say. Where the notes are silent, write NOT RECORDED. Never estimate a number, name a cause or add a time I did not give. Keep quantities and times exactly as dictated. Under "Possible variations" list any instruction or request that might change the contract scope, but do not price it or say whether it is in scope.
Notes: [paste the transcript, with names and addresses removed]Here is a messy dictation and what comes back, both illustrative.
Dictated: "Tuesday, extension job. Dry, cold start. Six on site, three bricklayers, a labourer, me and the scaffolder in the morning. Blocks to first floor level on the back wall done, side wall about half. Lintels came at ten, two short, driver says Thursday. Engineer visited at two and said the steel over the new opening needs to be bigger, I've got that in an email, so that's extra. Client's partner asked if we could move the kitchen window along by 300, I said I'd come back to them. Lost an hour at lunchtime, the pump didn't turn up. Tomorrow roofers measure and we carry on with blockwork."
Date and job: Tuesday, extension. Weather: dry, cold start. People on site: 6 (3 bricklayers, 1 labourer, site manager, scaffolder in the morning). Work done: blockwork to first-floor level on the rear wall complete; side wall about half complete. Deliveries: lintels at 10:00, two short; driver says Thursday. Visitors: structural engineer at 14:00. Instructions: (1) engineer, steel over the new opening to be larger, confirmed by email; (2) client's partner, asked to move the kitchen window 300 along, not yet agreed. Delays: about one hour lost at lunchtime, pump did not arrive, no reason stated. Safety observations: NOT RECORDED. Possible variations: larger steel (written instruction); window move (verbal request). Tomorrow: roofers to measure; blockwork continues.
The site manager reads it in under two minutes, which is the sign-off. Three things to check every time: the headcount, because it feeds labour costs and any later delay claim; every number and time against what was dictated; and the NOT RECORDED lines, which are the model being honest about gaps you may want to fill. If the log says "dry" and it rained, or invents a safety talk, the template rule has failed and needs a stricter sentence. Keep the raw dictation as an attachment to the final log, so that if anyone later asks what was said and what was written, the first is on file.
Filing: keep the log and photos where the job lives
Paste the checked log into JobTread's daily log for that job. Its page describes logs with notes and customisable fields, plus uploaded files, photos and videos with markup. The page does not describe a sign-off step, so add one yourself: a customisable "Checked by" field, filled in by the site manager. Photos belong here, not in the chat; the AI needs only the text.
JobTread also lists AI Daily Log Summaries, which condense a run of daily logs into a single progress update you can share with the team, the customer and suppliers. Use it for the weekly customer update, with one rule: read the summary against the logs before it is shared. A summary is a compression, and compressions drop things. The commonest loss is the delay's cause, which turns "pump did not arrive" into "minor delay", a change that matters if you later need to show who caused it. For how the photo side works, the tutorial on construction photo documentation apps and their AI plans compares the options.
Variations: catch them while they are still one line in a log
Unpriced extra work is the most expensive leak in a small building firm's paperwork. It starts as a sentence in a log, "client asked to move the window", and ends as a dispute over the final account. The fix is a weekly read of the logs by a model that is told only to look for instructions that might change scope.
Below are the daily logs for one job since my last review. List every instruction, request or change that might alter the contract scope, in a table with the columns: Date; What was asked or changed; Who asked; The log wording (quoted); Written or verbal; Priced yet (yes, no, not stated). Do not decide whether anything is in scope: write "check the contract" in that column. Do not price anything. If you find nothing, say so.
Logs: [paste]Run over the extension's last two weeks of logs, the result might be three rows: the larger steel (engineer, written, not priced); the window move (client's partner, verbal, not priced); and an extra wall chase mentioned the following week with no instruction recorded. The third is the useful one, because it is a cost without a trail. The owner decides which rows are real variations and builds the price. Here are two, with the firm's own figures.
- V1, larger steel over the new opening. Materials $1,900.00; labour 14 hours at $62.50 = $875.00; plant $375.00; total $3,150.00.
- V2, kitchen window moved along. Labour 12 hours at $62.50 = $750.00; materials $530.00; total $1,280.00.
Only now does ChatGPT or Claude write the notice, from facts and numbers you supply, ending with a plain approval step: "Reply 'approved' to this email to confirm." Store the reply against the change order in JobTread, whose features list change orders. The tutorial on turning site messages into change orders goes deeper on the drafting, and the model should never decide whether work is extra: that depends on your contract, and you decide.
JobTread says you can connect ChatGPT or Claude to your JobTread data through its MCP server, which could let this step read the logs directly. Start with copy and paste. It is slower by two minutes and keeps you in the loop on what leaves the system.
Costs: bills and hours land on the right job the same week
Cost capture is where small firms lose the link between a site and its money. Two inputs matter: supplier bills and hours. JobTread lists AI receipt scanning, where you photograph a receipt and the AI captures the details, and time tracking that the office reviews and approves. The page does not say whether scanned details are saved without a check, so treat every scan as a draft: confirm the supplier, the date, the total and, above all, which job the bill belongs to.
Take a firm with 60 bills a month. Keyed by hand at four minutes each, that is four hours; photographed and checked at two minutes each, it is two. The saving is small per bill and large in effect: bills that reach the job the same week let you see a cost overrun while you can still act on it. The realistic mistake is a misallocation. A $312.40 timber bill photographed from the van is saved against the shop fit-out instead of the extension, and nothing looks wrong until the extension's cost report comes up short. Checking the job name against the delivery note takes ten seconds at the time and an afternoon later. If you keep your books in Xero, its pricing page lists Smart Document Capture on all three plans, so office bills can be captured there instead, though the cheapest plan, Early, allows only five bills, far short of 60 a month. Bills captured in the accounts still need tagging to the right job. The tutorial on job costing with AI shows how the same data answers which jobs actually made money.
Invoice: build the progress claim from approved numbers
An invoice is arithmetic and judgement. The judgement, how complete each section is, belongs to the site manager. The arithmetic belongs to software. AI's role is to check both and to write the covering note. For the extension, with a contract sum of $212,000.00, a valuation after six weeks looks like this.
| Section | Contract value | Complete | Value to date |
|---|---|---|---|
| Preliminaries and groundworks | $38,000.00 | 100% | $38,000.00 |
| Foundations and slab | $27,000.00 | 100% | $27,000.00 |
| Superstructure | $62,000.00 | 60% | $37,200.00 |
| Roof | $41,000.00 | 0% | $0.00 |
| First fix services | $24,000.00 | 0% | $0.00 |
| Finishes | $20,000.00 | 0% | $0.00 |
| Approved variations V1 and V2 | $4,430.00 | Done | $4,430.00 |
| Total to date | $106,630.00 | ||
| Less previous invoices | $60,000.00 | ||
| This invoice | $46,630.00 |
The sections add up to the contract sum ($38,000 plus $27,000 plus $62,000 plus $41,000 plus $24,000 plus $20,000 is $212,000), the work done is $102,200.00, and the two variations add $4,430.00 to make $106,630.00. Your software or a spreadsheet does that sum. JobTread turns orders into invoices and covers payment applications for progress billing, and QuickBooks Online lists progress invoicing on its Plus and Advanced plans. If your contract holds back retention, add that line too; the tutorial on tracking retainage and chasing its release covers it.
Two prompts then do their checking. The first recalculates, with the rule "do not correct, list differences". The second is the one with real value: it compares the logs with the percentages.
Below are the daily logs for the invoice period and the valuation table. For each section, say whether the logs support the percentage complete, quoting the log lines. List any section where the logs suggest more or less progress than claimed. Do not change any figure. If the logs are silent on a section, say so.
Logs: [paste]
Valuation: [paste]A typical return confirms the superstructure at 60% (rear wall complete, side wall about half), confirms the roof at 0% (roofers only measured), and queries the slab: an earlier log mentions "slab edge to be made good", which may make 100% premature. That query costs nothing to resolve and removes an argument a client could have started. For more tests on the claim before it goes out, the tutorial on whether AI can check a construction payment application adds them.
The covering note then restates the same facts in plain words, from the logs and the table only. An illustrative draft: "Progress to date: groundworks and foundations complete; superstructure 60% (rear wall to first-floor level, side wall about half); roof and first fix services not started; variations V1 and V2 completed. Total to date $106,630.00, previous invoices $60,000.00, this invoice $46,630.00." The owner reads the whole pack, note included, and approves it.
Chasing: collect what is unpaid without souring the job
Use the reminder tools in your accounting package first, since they already know what is overdue. Ask ChatGPT or Claude only for the wording of the two or three messages: a friendly check on day three, a firm request for a payment date on day ten, and a phone call before anything formal. Give the model the invoice number, amount, due date and the section it covers, and forbid it from mentioning stopping work, interest or any contract remedy. Those are the owner's decisions.
A day-ten draft for the extension's invoice reads: "Hi [first name], a reminder that invoice 0057 for $46,630.00, for progress to date on the rear extension, fell due on [date], ten days ago. Please let me know the date you plan to pay it. If a line on the valuation needs explaining, I can go through it with you on site." Notice that it cites only the number, amount, due date and scope. It does not mention work stopping, because that decision, and any right to take it, sit in your contract and with the owner.
Who signs what: the 90-second and five-minute checks
- Site manager, 90 seconds a day. Read the log; check headcount, times, quantities and the NOT RECORDED lines; fill the "Checked by" field. Weekly, five minutes: read the AI summary against the logs before it is shared.
- Owner, five minutes a variation. Check that the instruction is real, the price is built from your own rates, and the notice states what happens if the client declines.
- Office, ten minutes a week. Check that every scanned bill sits on the right job, that no bill was scanned twice, and that timesheets match the log's headcount.
- Owner, ten minutes an invoice. Check the percentages against the log comparison, the variations against approvals on file, and the arithmetic against the software.
A ten-person firm: what it costs and what it saves
JobTread's pricing page lists $199 a month plus $20 a month for each internal user beyond the first, or 20% less on an annual subscription, with no free trial and a 30-day money-back guarantee on monthly plans. Vendor, subcontractor and customer portal users are free. With four internal users (owner, site manager, administrator and one lead tradesperson), that is $199 plus three at $20, so $259 a month. Watch the seat count for the crew. JobTread lets field crew who only upload photos and log daily notes join free, but its page counts people who track time among internal users, so if the other six workers clock their own hours, ask JobTread whether each needs a paid seat: at $20 each that would add $120 a month. The alternative is for the site manager to enter the day's hours from the log, which also keeps the headcount and the timesheets in step. ChatGPT Business or Claude Team needs two seats at about $20 to $25 each, so $40 to $50 a month for the site manager and the office. The total is roughly $300 a month at list prices, so check the current pages before you buy.
The savings, with illustrative timings: 80 daily logs a month (four jobs, 20 working days) at 6 minutes saved each is 480 minutes; 60 bills at 2 minutes saved is 120 minutes; 6 progress invoices at 25 minutes saved is 150 minutes. That is 750 minutes, or 12.5 hours, worth about $690 at a blended $55 an hour. Add one thing the time sums cannot show: if the weekly variation read catches a single unpriced $1,280 job a month, it has paid for the tools on its own. Measure the real figures over one month: logs finished the same day, days from a section finishing to its invoice, and variations priced before the work.
A four-week order for switching it on
Do not start all six steps on the same Monday. A staged start shows you which step the crew will actually use.
- Week 1: logs on one job. One site manager, one job, the Project and the template. Compare each AI log with what you would have written, and tighten the rules where it strayed.
- Week 2: the weekly variation read. Run the variation prompt on that job's logs on Friday. Price whatever it finds, and note how many items were real.
- Week 3: bills and hours. Add receipt scanning and timesheet approvals for the same job, and check the allocations every day.
- Week 4: the invoice cross-check. Build the valuation by hand as usual, then run the log comparison and see whether it finds anything. Only then roll out to the other three jobs.
How the paperwork fails quietly
- Style drift. Different foremen dictate differently. The template and a rule of "keep quantities as dictated" absorb most of it.
- Quiet invention. A model fills a gap with a plausible detail. The NOT RECORDED rule and the site manager's read are the defence.
- A variation logged but never approved. The weekly read finds it; the owner must then price it and chase the approval.
- A bill on the wrong job. A classic scanning error. Check the allocation, not just the total.
- Personal data. Strip names and addresses before pasting, keep photos out of the chat, and use a business plan so the content is not used for model training by default.
- Lock-in. Before committing to any job software, find out how to export logs, photos and invoices, because the records are yours and must outlive the subscription.
Questions construction firms ask about automating paperwork
Can AI write a daily log that is safe to rely on in a dispute?
Only if a person checks and signs it. Keep the raw dictation as an attachment, tell the AI to write NOT RECORDED wherever your notes are silent, and have the site manager approve the final version the same day. A log that nobody checked is a record of what the AI wrote, not of what happened on site.
What if site managers will not dictate a log every day?
Make the log shorter, not optional. A two-minute voice note before leaving site, prompted by a fixed phone reminder, is quicker than typing, and the template does the formatting. Show the payoff: the weekly customer update and the variation list are built from those logs, so fewer people chase the site manager for answers. If a day is missed, the office rings for the notes that afternoon.
Do I need JobTread to follow this chain?
No. It is the example because its pricing and AI tools are published. The same steps work with any job software that holds logs, change orders, bills and invoices, or with a mix such as Raken for daily reports and QuickBooks Online for invoicing. What matters is one tool per step and a named person signing each one.
Should ChatGPT or Claude see site photos?
Usually not. A log needs text, not images. Keep photos in your job software, where access is controlled, and send the AI only the dictated notes with names and addresses removed. If you do share photos for a specific job, check who and what is visible in them first, including neighbours and workers.
Further reads
- Five AI Apps for Construction Site Work: Photos, Logs, Checks — Five AI apps for site photos, logs and checks.
- Best AI for a Small Construction Company's Financials and QuickBooks — Where QuickBooks fits in a small construction firm's finances.
- 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.
- Best AI App for a Construction Company: Logging Near Misses — Five ways to log near misses from site, compared on photo capture, voice notes and corrective-action follow-up, with a two-week trial plan.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: JobTread's pricing, features, daily log and AI pages; Intuit's QuickBooks Online pricing and invoicing pages; Xero's pricing page; Raken's home page; CompanyCam's pricing page; OpenAI's ChatGPT pricing page; Anthropic's help article on Claude Projects; read in late October 2026.