Can AI Check a Construction Payment Application Before You Send It?

Yes, AI can check a payment application and its schedule of values for arithmetic and consistency before you send it: percent-complete maths on every line, totals, stored materials against invoices, the retention line and reconciliation to the contract sum. It cannot see the site, so it misses overclaimed progress, missing evidence and contract-specific terms; a person, usually your accountant, signs it off.

Think of it as a tireless second reader who reads every row. Payment applications fail in dull ways: a percentage typed as 85 instead of 75, a column that no longer adds up after a late edit, retention taken from the wrong base. A chatbot is good at those, because they are about rows of numbers that either agree or do not. It is bad at the question the owner's representative will ask, which is whether the work is really there.

Follow me on Instagram@sagnikteaches

What you hand it, and what it hands back

A payment application built on a schedule of values (the contract sum broken into priced lines, sometimes called a pay app or an application for payment) has a predictable shape: one row per line, then totals. The check works best when you give the chatbot that table as plain text rather than a photograph of it, with these columns:

Connect on LinkedInSagnik Bhattacharya

• Line number and description
• Scheduled value
• Work completed in previous applications
• Work completed this period
• Materials presently stored (not yet installed)
• Total completed and stored to date, and the percentage of the scheduled value
• Balance to finish

Underneath sit the summary figures: contract sum, any approved change orders, retention percentage, total earned less retention, previous certificates, and current payment due. If any of those summary cells is a typed number rather than a formula, say so, because typed numbers are where drift hides.

Subscribe on YouTube@sagnikteaches

Checking an electrical subcontractor's third application

Take an electrical subcontractor with a contract sum of $480,000, ten per cent retention and a monthly application. This is the third application. The sums below are illustrative.

The summary block as sent: contract sum $480,000; completed and stored to date $282,000; retention $24,240; earned less retention $257,760; previous certificates $147,420; current payment due $110,340.

The prompt:

You are checking a construction payment application before it is sent.
Below is the schedule of values as a table, then the summary block.
Contract sum: 480,000. Retention: 10 per cent on work completed and
stored materials. Previous certificates (net of retention): 147,420.

Do these checks and show your working for each:
1. Recompute every line: previous + this period + stored = total to
   date. List any row where my total is different.
2. Recompute the percentage of each line (total to date / scheduled
   value). List any row where my percentage differs by more than
   0.5 points.
3. Confirm the scheduled values add up to the contract sum.
4. Recompute totals, retention, earned less retention, and the
   current payment due. List differences.
5. Flag any line that moved more than 30 percentage points in one
   period, and any line with stored materials, as needing evidence.
6. Do not guess. If you are unsure about a figure, say so.

[paste table and summary block]

A realistic response, labelled illustrative, comes back like this. Line totals all agree. The scheduled values sum to $480,000. Line 4's percentage is wrong: $81,000 against $108,000 is 75%, not 85%. Total completed and stored is correct at $282,000. Retention differs: ten per cent of $282,000 is $28,200, but the application shows $24,240, which is ten per cent of work completed only ($242,400). Earned less retention should be $253,800, and the current payment due should be $106,380, not $110,340. The chatbot then flags line 3 and line 6 as carrying stored materials of $28,800 and $10,800 that need supplier invoices and proof of storage, and flags line 5, which moved from 10% to 30%, as an increase of 20 points, below the threshold but worth confirming.

What you would fix: the 85% typo, which does not change the money but would embarrass you in front of the owner's quantity surveyor; and the retention base, which here overclaimed $3,960. Whether the contract really applies retention to stored materials is a contract question the chatbot cannot answer, so you check the clause. In many contracts it does; in some, stored materials carry no retention or a different rate. You then recompute in the spreadsheet using the formula, not the chatbot's number, and send the corrected figure.

Where each check stops being reliable

Percent complete is a claim, not a number

The chatbot can confirm that 81,000 over 108,000 is 75%. It cannot tell you whether the first floor is 75% wired. Percent complete is a judgement, and the usual dispute is front-loading: claiming more than the work done on early lines to improve cash flow. If your contract allows a front-loaded schedule, fine, but owners and their quantity surveyors read these lines closely. A useful extra prompt asks the chatbot to compare this period's percentages with the previous two applications and list the lines whose rate of progress looks out of step with the others, then you ask whether you can explain each in one sentence.

Stored materials need paper

Stored materials are goods paid for in the application before they are installed, usually on site or at a bonded warehouse. Contracts typically want invoices, proof of ownership and insurance cover, and sometimes photos or an inspection. A chatbot can tell you which lines carry stored materials and whether their amounts match the invoice totals you paste in. It cannot know whether the switchgear invoice already includes the deposit you also billed under line 3's work. That overlap is a classic double count. So list each stored item beside its supplier invoice number and amount, then ask the chatbot to compare the two lists and flag any difference or any item that appears in both stored and completed columns.

Retention is a contract term

The retention percentage, the base it applies to, any reduction at practical completion and any cap are in your contract. Paste the retention clause as text and ask the chatbot to restate it in two sentences, then check its restatement against the clause with your own eyes. If it paraphrases wrongly, you catch the error before any sum is built on it. For the follow-on problem of releasing held money, see tracking retainage and chasing release.

Reconciling to the contract sum and change orders

The scheduled values must add up to the contract sum, plus any approved change orders added as their own lines. Pending or unapproved changes do not belong in the claim at full value unless your contract says otherwise. Ask the chatbot to list every line described as a variation or change and compare against your change order log. It will happily find a variation line with no matching approval number, which is exactly what an owner's representative would find next.

Wording, certified figures, cover notes and cash held back

Rewriting a line description that invites argument

Before: line 5, "Second floor", with no scope. After: "Second-floor branch wiring: rough-in, conductors and devices, excluding fixtures (line 6)". The after version matters because a vague description invites an argument about what is complete. Ask the chatbot to read your schedule and list any line whose description could overlap with another, then fix the wording. In the example above it flags "Feeders and switchgear" and "Light fixtures" as separate, but notes line 3 does not say whether panel boards are included, which is a scope question for you.

The previous-column check

The most common silent error is a "previous" column that does not match what the owner certified last time. Paste the last certified figures beside your current previous column and ask for differences. If you changed a number, perhaps because the owner reduced a line last month, the application must carry the certified figure, not the claimed one. In the sample, if line 2's previous figure had been typed as $47,000 instead of $45,000, the chatbot would flag a $2,000 mismatch against the paste of last month's certificate.

Plain-English summary for the project manager

After the numbers are fixed, ask for a short note to go with the application.

Write a four-sentence cover note for this payment application, for the
owner's project manager. Give: the application number, the period,
total completed and stored to date as a percentage of the contract
sum, the amount now due, and a list of the stored materials with
their supporting invoices attached. Use a neutral, factual tone and
do not promise anything about the next application.

The illustrative output reads "Application 3 covers work to the end of the month. Total completed and stored is $282,000, 58.75% of the contract sum of $480,000. The amount now due, after ten per cent retention, is $106,380. Supporting invoices for stored switchgear and light fixtures are attached." You adjust wording to your style, and leave the amounts as the spreadsheet gives them.

The contract sum that was typed, not calculated

Here is an illustrative slip. A contractor raises line 4's scheduled value by $12,000 after a change order, instead of adding the change as its own line, and leaves the contract sum in the summary block as a typed $480,000 rather than a formula. Every row looks right on its own. The contract-sum check, if the chatbot is asked to run it, shows the lines adding up to $492,000 against a summary of $480,000. Left alone, the error would understate the contract value on every later application and blur the record of what the change order added. Two lessons follow: ask for check 3 every time, even when nothing seems to have changed, and give each change order its own line.

Cash still held back when only testing remains

Say a job reaches 100% on every line except testing and closeout, which together are $30,000 of the $480,000 contract. Work earned is $450,000, ten per cent retention holds back $45,000, so the most that can have been certified is $405,000. When the last two lines are billed, retention rises to $48,000, and that sum waits on the release terms. A chatbot can lay out what is claimable now against what waits for practical completion, which helps you plan cash. Check the practical completion and release terms in your contract before telling anyone a retention release date.

Why the chatbot slips on arithmetic, and how to guard against it

Chatbots are language models, not calculators, and long columns of figures are where they slip. A model may add a column of eight numbers correctly nine times and misread a digit on the tenth. Three habits cut the risk. First, ask it to show every sum line by line so you can see where a figure came from; a wrong total with visible working is easy to spot, and a bare wrong total is not. Second, where your plan offers a code or data-analysis tool, ask it to run the sums in code and show the result, which makes the arithmetic mechanical. Third, treat any difference it reports as a prompt to look, not as a verdict: recompute that one cell in your spreadsheet, and you will quickly learn whether the chatbot or the sheet was wrong.

A good test of the whole setup is to run it once on a past application where you already know the answer, or to plant one deliberate error in a copy and see whether it is caught. If the chatbot misses a planted 85% typo, you know to tighten the prompt before trusting it on a live claim.

Spreadsheet habits that make the AI check shorter

The cleaner the schedule, the less the chatbot has to untangle. A few habits help:

• One formula per column. Total to date is previous plus this period plus stored, on every row, with no typed overrides. A single typed override is the usual cause of a number that does not reconcile.
• A percentage column calculated, not typed. Total to date divided by scheduled value. Typed percentages are how 85% replaces 75%.
• Change orders on their own lines, each with the approval reference in the description, so the contract sum check has something to match against.
• A locked "previous" column copied from last period's certified figures, never retyped.
• A check cell at the bottom that subtracts the sum of lines from the contract sum and shows zero when they agree.

With those in place, the AI check becomes a confirmation rather than a hunt. Many firms find the spreadsheet fixes catch most errors on their own, and the chatbot earns its place by catching what the structure cannot: odd wording, an unexplained jump, a stored item with no invoice behind it.

Anticipating the owner's objections

Once the numbers are right, run a second pass from the other side of the table. Ask the chatbot to act as the owner's quantity surveyor reviewing the application and to list the five questions most likely to be raised.

Act as the owner's cost consultant reviewing this payment
application. Using only the table and notes I have pasted, list the
five questions you would ask the contractor before certifying.
For each, say which line it concerns and what evidence would answer it.

An illustrative reply for the sample application: why line 5 moved 20 points in a month when line 4 moved 25 and is further ahead; what proof of storage and insurance supports $28,800 of switchgear held off site; whether the light fixtures stored under line 6 are the specified product; whether line 2 at 95% is awaiting an inspection that would bring it to 100%; and whether any change order is pending. You do not need to answer all five in the cover note, but having a one-line answer ready for each saves a round of emails and often shortens the time to certification. It also exposes weak spots, such as a stored line with no photo, while you can still fix them.

Unit-rate lines for a paving subcontractor

For a landscaping or paving subcontractor billing by measured quantities, the schedule of values is often unit-priced: square metres laid at a rate, linear metres of kerb. The same check applies with one addition. Ask the chatbot to recompute quantity times rate on every line, then compare the quantities against the measured-up sheet. A line showing 1,200 square metres paved against a total design area of 1,150 square metres is exactly the sort of impossible figure the chatbot spots in seconds. Whether the extra 50 square metres is a genuine variation or a typo is again for you to determine from the site measurements.

A routine you can repeat every month

1. Build the application in your spreadsheet with formulas for every total, so the arithmetic is mechanical.
2. Export the table as text and strip the client's name, bank details and any private terms.
3. Run the six-point prompt and keep the answer with the job file.
4. Fix each difference at its source, in the spreadsheet, never in the chatbot's output.
5. Attach evidence: invoices for stored materials, change order approvals, lien or payment waiver forms if your contract uses them.
6. Send the draft to your accountant or bookkeeper for a final read of the retention and totals, then submit.

The whole routine takes about twenty minutes after the first month. Keep the prompt in one saved place so the checks do not drift. A shared Project in ChatGPT or Claude lets you store the six-point instructions and the retention clause once, so each month you only paste the new table.

What AI will not catch, so you must

• Work that is claimed but not on site, or not yet inspected.
• Supplier invoices that were never paid, which can create payment-chain problems.
• Contract deadlines for submitting the application and the owner's right to reject or reduce it.
• Whether a payment notice must accompany the claim. See drafting notices and checking the deadlines.
• Tax treatment, which belongs to your accountant.

Treat the chatbot's clean bill of health as meaning "the arithmetic and internal consistency pass", nothing broader. It is worth saying so in your own routine notes, so nobody on the team reads a clean check as approval.

What to strip out before pasting a pay application

Paste only what the check needs. Owner names, project addresses and bank details add risk and no value. Before uploading client documents, look at your plan's privacy settings: consumer plans have a switch that stops your chats being used to train models, and business plans generally leave business content out of training by default. When in doubt, replace names with "Owner" and "Project" and paste figures only. The build-up of a payment application is also a decent place to practise the pattern of asking a chatbot to audit rather than to produce, a habit that carries over to cash planning, covered in building a cash flow forecast with AI. For the broader picture, see the AI guide.

Payment application checks: your follow-up questions

Can I paste the whole application into ChatGPT or Claude?

You can, but strip the owner's name, bank details and any contract terms you do not need. For the maths check, a table of line numbers, scheduled values, previous work, this period, stored materials and totals is enough. Upload a spreadsheet or PDF only if your plan's data settings suit client documents.

Will AI spot an overclaim against the real progress on site?

No. It only sees what you give it. It can flag a line that jumps from 20% to 90% in a month or a trade that looks complete before its predecessor, but only a walk of the site, or the project manager's progress report, tells you what is truly installed.

Should AI calculate the amount due, or should the spreadsheet?

The spreadsheet should calculate it and AI should audit it. Ask the chatbot to recompute each line independently and list differences, which turns it into a second pair of eyes rather than the source of the number.

Who signs off the final figure?

You do, and where your firm uses one, your accountant or bookkeeper checks it before it goes out. The application is a legal claim for payment under your contract. A chatbot's check is a working aid, not an approval.

Further reads

• Can AI Check an Interim Valuation? A Quantity Surveyor's Test — The same claim seen from the quantity surveyor's desk.
• Construction Final Account: Variations, Deductions and Sign-Off With AI — The end-of-job version of this check, with variations.
• Electrical Change Orders Written With AI: Ask for Extra Work Clearly — Write the change orders your application relies on.
• Fergus for Builders: Job Costing and Progress Claims, With Fergus AI — Progress claims raised inside a job management app.
• Tradify for Builders: Set Up Quotes, Variations and Stage Payments — Stage payments if you work in a job app instead.
• Bill of Quantities Template: A Worked Example Checked With AI — Check the quantities your schedule of values came from.
• AI Automation for Construction Companies: Daily Logs to Invoices — Six paperwork steps from dictated daily log to progress invoice, one tool each, with a ten-person firm's costs, sums and who signs what.
• AI in Construction Accounting: Job Costs, WIP and Retainage Checks — Four month-end checks for a small contractor: code invoices, catch job-cost errors, recalculate WIP and track retainage, then hand your accountant a clean pack.
• Lien Waivers for Small Contractors: Track Conditional and Final Releases — A three-tab sheet that matches each conditional and final lien waiver to the payment it covers, with red-flag formulas and a worked HVAC example.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: general construction progress-payment practice (schedule of values, percent complete, stored materials, retention); ChatGPT and Claude help pages on file upload.