AI in a construction business is not good or bad. It is both, depending on the task and who checks the output. Here is how it plays out on real work, so you know where to lean in and where to slow down.

The Good: Paperwork AI Does Well

AI is strong at turning a pile of text into something usable. That describes half your office.

Bids. Paste in an RFP and ask for a scope checklist, exclusions list, and questions to send back. You get a first pass in minutes. You still own the numbers.
Submittals. AI can compare a submittal package against the spec section and flag what looks missing. You review the flags.
Contracts. Ask for a plain-English summary of a subcontract, with payment terms, notice deadlines, and indemnity language pulled out. It reads fast. It is not your attorney.
Safety documents. A draft toolbox talk or job hazard analysis for tomorrow’s task takes two minutes instead of twenty. Your competent person still walks the site and edits it.

The pattern: AI drafts, a person decides. Used that way, AI in a construction business buys back hours you currently spend on typing.

The Bad: Confident and Wrong

AI does not know when it is guessing. It sounds the same either way.

It invents code sections, spec references, and quantities that look right.
It misses the one exclusion buried on page 47.
It gives you a clean summary of a contract you never gave it, if you forget to attach the contract.

This is not hypothetical. In 2023 a federal judge sanctioned two New York lawyers $5,000 in Mata v. Avianca after they filed a brief with court cases ChatGPT made up. Swap “brief” for “bid” and you see the risk. A made-up number in an estimate does not get a judge’s attention. It gets your margin.

The fix is boring. Verify anything that carries money, code, or safety. Check the source, not the summary.

The Ugly: Data and Liability

This is where contractors get hurt, and it is the part most people skip.

Client data. Plans, owner financials, and employee records do not belong in a public chatbot. In 2023, Samsung banned staff from using generative AI after employees pasted sensitive source code into ChatGPT. You do not need to be Samsung for the same thing to happen with your bid history or a client’s security drawings.

Liability. If your AI tells a customer something wrong, you own it. In Moffatt v. Air Canada (2024), a Canadian tribunal held the airline responsible for bad information its own chatbot gave a customer. “The bot said it” did not work as a defense. A chatbot on your website quoting a warranty you do not offer would land the same way.

Contracts you signed. Some owners and GCs now restrict AI use on their projects. If you paste their documents into a tool without checking, you may be in breach before you notice.

A Simple Rule Set

Strip it. Remove names, addresses, and pricing before you paste anything. (That is the first letter of S.A.F.E., the Iron Thread data rule.)
Approve it. Only use information you are allowed to share.
Check it. A person verifies every number, code reference, and contract term.
Write it down. Put a one-page AI policy in front of your crew. If you do not have one, you have a policy anyway. It is “everyone does whatever.”

Where to Start

Pick one low-risk task, like drafting toolbox talks or summarizing meeting notes. Run it for a month. Then decide what comes next.

That is the approach behind every Iron Thread workshop. You get the speed without the lawsuit.

FAQ

Is AI safe to use on bids?
Yes for drafting scope checklists and questions. No for final pricing without a person checking every number.

Can I paste a client’s contract into ChatGPT?
Check the contract and your tool’s data settings first. Strip identifying details either way.

Who is liable if AI gets something wrong?
You are. Courts and tribunals have treated AI output as the responsibility of the business or lawyer who used it.

Do I need an AI policy?
Yes. One page is enough to start.

Sources

Book a consultation to set your AI rules before your crew sets them for you

The Good, the Bad, and the Ugly of AI in Construction