Can ChatGPT check a pay application?

Published 17 October 2026. Updated 17 October 2026.

A controller with a long pay application and a pile of field reports will sooner or later wonder whether a general AI chat assistant such as ChatGPT could just do the check. It is a fair question. The honest answer is that it can help with parts of the job, and that how much you can rely on it depends on how you set it up and what you do with the answer.

This guide covers what the job requires, where a chat assistant is useful, where it tends to go wrong, and how to get a trustworthy result whichever tool you use.

What you are asking it to do

Checking a pay app against the field has two halves:

1. Read a table correctly. Every line needs the scheduled value, the previous work, this period's work, stored materials and the resulting percent, taken from the right row and column.
2. Match prose to lines. The superintendent writes "units 1 to 7 set, 8 to 10 waiting on hookups". Someone has to decide which schedule line that belongs to and what percent it supports.

The first half is exact. The second is judgment. A tool can be good at one and weak at the other. The longer method is set out in how to check a pay app against field reports.

What a chat assistant does well

• Summarising notes. Turn weeks of free text into a list by area or trade, which you then check against the original.
• Explaining the form. What a column on the continuation sheet means, or what retainage is, in plain words.
• Drafting. A polite note asking a sub for specifics on a line with a vague claim.
• Brainstorming the questions. What would I want to know about a mechanical piping line at 55 percent?

These are useful, low-risk jobs because you can see at once whether the output makes sense.

Where it tends to go wrong

These are problems of the approach, not of any one product, and they vary with the tool and the prompt, so test your own setup.

• Long tables. Reading dozens of rows from a PDF, or from a scan or photo, can drop or misplace a row, and the answer still reads confidently.
• Arithmetic. Adding a column of large figures is exactly the kind of work where fluent text and exact sums can diverge.
• Quotes. If you ask for the evidence behind a percent, it may paraphrase rather than quote, and a paraphrase can drift from what the note says. Unless the output is checked against your notes, you cannot tell.
• Inventing a number. Asked how complete a line is, an assistant may give you a percent even when the notes say only "good progress". The honest answer there is that there is no evidence.
• Different answer next month. The same prompt can give a different structure each time, which makes month-to-month comparison hard.
• Nobody checks the second time. If the same model reads, calculates and judges, the result has no independent look.
• Privacy and records. What you paste goes to a service. Check the terms, and keep your own record of what you checked.

How to get a better result from any assistant

If you want to use a chat assistant anyway, these habits help:

1. Give it the table as text or a spreadsheet, not a photo, whenever you can.
2. Do the sums in a spreadsheet and ask the assistant only to read and match.
3. Ask for a quote from the notes for each line, then search your notes for that quote.
4. Tell it to answer "no evidence" when the notes give no number, and check that it does.
5. Work from the biggest dollar lines first and verify those by hand.
6. Keep the prompt and the output with the pay app file.

Side by side

When a chat assistant is the better choice

Use it when the task is open ended: drafting a letter, explaining a clause, summarising a long daily log, or thinking through how to approach a difficult sub. It is also the sensible tool if you check one or two pay apps a quarter, you enjoy tuning prompts and you will verify every figure yourself. Its flexibility is a real advantage that a single-task tool does not have.

When a purpose-built tool is the better choice

It makes more sense when you check every pay app on every job, the result has to look the same each month, and you want the arithmetic and quotes checked by something other than the same model that wrote them. Then the value is in the repeatable process and the independent check, not in the conversation.

How GroundClaim does this step

GroundClaim is built for the one task. You upload a job's monthly pay application (a spreadsheet, a Word or PDF table, or a scan or photo, up to 400 lines) and the field notes as free text. Code does the sums, including the G702 and G703 tie-out, retainage and schedule totals, and works out the dollar gap on each line. The notes are searched for progress on each line, and every excerpt it quotes is checked word for word against your notes. Where the notes give no number, the line is marked as insufficient evidence and no percent is made up.

You get an exceptions memo ranked by dollar exposure and a line-by-line reconciliation that also shows the clean lines, plus a one-line recommendation that is only a suggestion. A second, stronger AI model reviews every result, fixes what it can, and lists what is left for you to decide. Results are model output checked by code and by a second model; they are not verified by a person. GroundClaim does not accept or reject a pay application and works only from the documents you upload.

One pay application costs 25 credits, with no subscription. Create an account and try it on a closed-out project where you already know the answer.

Questions

Is it safe to paste a pay application into a chat assistant?

That depends on the service and your agreement with it. Check how it stores and uses what you upload, and what your contracts and clients allow, before you paste anything. Remove names you do not need to share.

Can I just ask the assistant to check the math?

You can, but verify the result. Sums over long tables are where fluent text and exact arithmetic can part ways. Recompute the totals yourself or in a spreadsheet, whatever the assistant says.

What is a good use of a chat assistant here?

Drafting the question to send a sub, summarising a long set of field notes into a list you then verify, or explaining what a column on the G703 means. Use it for support work, not for the final number.