Workflow / AI for subcontractors

How to use ChatGPT, Codex, and other AI tools in construction

Three ways to use AI for project office work, with a source and a human check for every important output.

AI tools in construction are a practical fit for work your team already knows how to check. A project engineer can review a draft action list against the meeting notes. An estimator can trace an unusual spreadsheet total back to its rows. Those are better starting points than asking a model to decide whether a drawing complies with code or whether a bid is ready to send.

Pick the tool based on the output. A chat assistant is useful for a draft written in ordinary language. A coding agent can build a repeatable utility. A spreadsheet assistant can help explore a table. None of them should become the project record or the person approving a change.

Use ChatGPT to turn meeting notes into draft actions

Start with approved, sanitized notes from a coordination meeting. Give ChatGPT a narrow output: an action table with an issue ID, owner, due date, and a reference to the note that supports each row. Number the notes before pasting them so a reviewer can trace the result. The ChatGPT prompting guide covers giving the assistant source material, context, and a defined output.

Use only the numbered notes below. Separate decisions, proposed changes, and action items. For each action, return the issue ID, owner, due date, and note number. Write "not stated" for missing details. Do not infer that a proposal was approved. List any conflicting notes at the end.

If the note says "mechanical to confirm access after the next model issue," the draft should show a follow-up and no calendar deadline unless one was stated. The meeting lead checks owners, dates, and commitments against the source before copying the agreed actions into the official log.

Use Codex to build a check you can run again

Codex can work with code and run local development tools. A contractor with a technical reviewer can use it to create a small utility for comparing two bid-log exports. The task is to write and test the utility. Once reviewed, the utility itself can run the same fixed rules each week without asking an AI model to interpret every file.

Using these two synthetic CSV fixtures, write a read-only comparison tool. Match rows by bid ID, report due-date changes, and flag duplicates or missing columns. Preserve the input files and write a separate report. Add tests for duplicate IDs, blank dates, and a bid present in only one export. Do not connect to live project systems.

Have the reviewer check what counts as a match, how dates are parsed, and whether malformed rows stop the run or get flagged. Test on copies before using approved project data. A missing row may reflect a changed export filter, so do not treat it as a withdrawn bid without checking the source.

Use a spreadsheet assistant to flag rows worth checking

In an approved Microsoft 365 environment, Copilot in Excel can help answer questions about structured data. For example, ask it to identify takeoff rows that have a quantity but no unit, group them by bid package, and return row IDs. Check the result against a normal filter. Use the source drawings and the estimator's review to correct quantities or units; do not let the assistant fill them in.

Illustrative construction takeoff with scope, quantities, and units to verify against the source
An illustrative takeoff, not an AI-generated estimate. Compare quantities and assumptions against the underlying documents.

Set a boundary before anyone uploads project information

Check client permissions and your company's rules for the specific account and tool. Pilot on synthetic or approved sanitized data, including attachments and metadata. Keep confidential drawings, bid pricing, payroll information, credentials, and personal details out of unapproved tools. Give a coding agent only the files and permissions it needs for the assignment.

Keep safety, design, code compliance, contractual commitments, bid sign-off, and field release with the responsible people. Start with one task for two cycles. Compare the draft against the manual result, count material errors, and measure time including review and correction. Drop the pilot if you cannot trace important outputs to a source.

For the process that comes before tool selection, read how to standardize and automate a construction workflow. To define who maintains the source records and who reviews the output, explore project coordination support or discuss a scoped office-support workflow.