The subcontract landed in Dana's inbox at 3:40 on a Friday afternoon. It ran 38 pages with four exhibits, plus a note from the general contractor's project manager asking for a signature by Monday so electrical rough-in could start on schedule. That weekend is a useful way to see what AI contract review for contractors actually does. It did two useful things and one dangerous thing, and most articles on the subject leave out the dangerous one.
Dana is a composite, not a client. I built the character from the kind of commercial subcontract that trade owners in the $1M to $5M range sign every quarter. Picture a 14-person electrical shop outside Austin with a strong residential service base, trying to grow into commercial tenant-finish work. The contract language below is typical of GC-drafted paper. The Texas statutes are real and linked, so you can check every one of them yourself.
Friday, 3:40 p.m.: thirty-eight pages and a Monday deadline
The job was a medical office build-out, and the electrical subcontract came to $412,000. It was the biggest commercial number Dana's shop had ever signed. The GC had a good reputation, paid reasonably on time on a smaller job the year before, and Dana had signed their paper then without reading much past the scope exhibit and the schedule.
Most owners I talk to admit the same habit. They read scope, price, and start date closely. They skim the general conditions, and they never open the exhibits that say "incorporated by reference." That habit isn't laziness. A 38-page subcontract takes two to three focused hours to read properly, and nobody running trucks has those hours on a Friday.
Dana's office manager had already been using a business-tier AI assistant to triage the inbox, the kind of setup described in our piece on AI email assistants for contractors, and suggested uploading the PDF. Before doing that, they checked two things:
- Data settings. Business and team plans of the major assistants generally don't train on uploaded files by default. Consumer plans often need a setting switched off. A subcontract contains the owner's name, the contract value, and the GC's terms, so it shouldn't go into a tool that learns from it.
- The contract's own confidentiality clause. Some GC agreements restrict sharing contract documents with third parties. This one limited disclosure to "advisors and professionals," which the office manager decided covered a private business AI workspace. If yours reads more tightly, ask before uploading.
The prompt that made the difference
They didn't type "is this contract okay?" That question gets a reassuring paragraph and nothing you can act on. Instead, they wrote a prompt that forced the tool to point at actual text:
- "You are reviewing a subcontract for an electrical subcontractor in Texas. Summarize it in plain English in under 400 words."
- "List every clause dealing with payment timing, retainage, indemnity, insurance, notice deadlines, delay, termination, warranty, and flow-down. For each one, quote the exact language and give the section number."
- "Mark anything that shifts risk to the subcontractor beyond its own work."
- "List every document the contract incorporates by reference that is not attached to this PDF."
The quote-the-language instruction does more work than the other three combined. A summary can be wrong without you knowing. A quoted clause with a section number can be checked in ten seconds. The tool answered in about four minutes. Dana then spent roughly forty minutes checking each flagged item against the PDF, which is a long way from three hours of cold reading.
Wondering which AI tool pays for itself first? Request a free AI audit and find out.
What did the AI contract review actually catch?
The tool returned eleven flagged clauses. Seven were routine. Four changed how Dana saw the job.
A pay-if-paid clause wearing pay-when-paid clothes
Section 7.3 read: "Subcontractor's receipt of payment is expressly conditioned upon Contractor's receipt of payment from Owner for Subcontractor's Work." The AI labeled it a condition precedent, which is correct. That makes it pay-if-paid, not pay-when-paid. A pay-when-paid clause controls timing: you'll get paid, maybe late. A pay-if-paid clause shifts the risk: if the owner never pays the GC, the GC may never owe you.
Texas allows these clauses but regulates them in Chapter 56 of the Texas Business & Commerce Code, which limits when they can be enforced and gives subcontractors a written-notice procedure. Other states treat them very differently. Courts in California and New York have refused to enforce pure pay-if-paid clauses. The AI mentioned Chapter 56 without being asked, which was helpful. Whether the clause would hold up on this particular job is a question for a lawyer.
The companion statute matters just as much. According to the Texas Legislature's Property Code Chapter 28, a contractor must pay a subcontractor within 7 days of receiving payment for that subcontractor's work, and overdue amounts bear interest at 1.5 percent per month. Dana hadn't known the interest figure. It's now a line in every past-due letter the shop sends.
Retainage tied to the wrong finish line
Section 8.1 held back 10 percent of each progress payment "until Final Completion of the Project and acceptance by Owner." That means the whole project, not Dana's scope. On this build, electrical rough-in and trim would be done around month seven. Final completion of a medical office with specialty equipment and a health-department sign-off could easily run to month fourteen. That left $41,200 of Dana's money sitting in the GC's account for half a year after the crew had moved on.
For comparison, federal jobs run under FAR 52.232-5 allow retainage of up to 10 percent only when progress is unsatisfactory. That isn't the rule on private work, but it shows that holding 10 percent flat until the end is a choice, not a law of nature. The AI's suggested ask was reasonable: release retainage on completion and acceptance of the subcontractor's scope, or cut it to 5 percent at 50 percent completion. Owners who track their numbers with AI dashboards tend to spot this cost fast, because outstanding retainage shows up as cash they've earned and can't spend.
Indemnity that reached past Dana's own work
Section 12.2 required Dana to indemnify the GC and owner for claims "arising out of or related to the Work, regardless of whether caused in part by a party indemnified hereunder." That phrasing is broad-form indemnity. Read literally, Dana's shop could be on the hook when the GC's own negligence helped cause the loss.
The AI correctly noted that Texas Insurance Code Chapter 151 generally voids provisions requiring a subcontractor to indemnify a GC for the GC's own negligence on most construction contracts. It then called the clause "likely unenforceable in Texas." That's where Dana had to slow down. Chapter 151 has exceptions, the biggest being claims for bodily injury or death of the subcontractor's own employees. And "likely unenforceable" doesn't protect you until a court says so. The practical question is whether the shop's general liability policy covers the contractual liability the clause tries to create, and only the insurance agent can answer that.
A seven-day notice trap
Section 15.4 required written notice of any claim for extra cost or time "within seven (7) calendar days of the event giving rise to the claim," and waived any claim not noticed in time. On a busy job, seven calendar days can pass before the office even hears about the problem. The AI flagged the deadline. It couldn't tell Dana that the foremen had no habit of reporting delays in writing. Fixing that meant a new field process, not a better contract.
Where does AI contract review go blind?
This is the dangerous part, and it's why the story is worth telling.
It can't read what isn't there. Exhibit C said the prime contract between the GC and the owner was "incorporated by reference," and the flow-down clause bound Dana to every obligation the GC owed the owner "to the extent applicable to the Work." The prime contract wasn't in the PDF. Because the prompt asked for missing documents, the AI listed it, but it couldn't review it. When Dana requested a copy, the GC sent a redacted version containing a daily liquidated damages figure that would flow down to any sub who delayed the schedule. None of that was visible in the 38 pages.
It blended two clauses into one wrong answer. The plain-English summary said "payment due within 30 days of invoice." That language came from the general conditions. A rider in Exhibit D replaced it with payment due 45 days after the GC receives payment from the owner, and an order-of-precedence clause in Section 2 said the rider controlled. The summary had merged the two. Dana caught it only because the itemized list quoted both sections, and the numbers didn't match the summary. A skim-the-summary owner would have planned cash flow around a date that didn't exist.
It can be out of date on state law. Texas overhauled its mechanic's lien statute effective January 1, 2022, and changed the notice deadlines subcontractors rely on. A general-purpose tool may describe the old rules with complete confidence. Treat any legal citation the AI gives you as a lead to verify on the statute's official page, never as the answer.
It doesn't know your position. It can't tell you that this GC always rejects retainage changes but will move on notice periods, or that losing this job would leave two crews idle in March. You bring that context. The AI doesn't have it.
Should contractors send every flagged AI contract review item to a lawyer?
No. Send the short list, prepared well. An AI first pass works best as intake for an attorney review, not as a substitute for one. The owner does the sorting, and the attorney spends their time only on the clauses that carry real legal or insurance risk.
Dana's attorney packet was one page: six clauses, each with the quoted language, the section number, and a specific question. "Is 7.3 enforceable under Chapter 56 if we send notice?" is a question an attorney can answer quickly. "Can you look over this contract?" is not. Here's roughly how the sorting worked:
| Clause type | AI is good at | A human must decide |
|---|---|---|
| Payment timing and retainage | Finding every related section and spotting conflicts | Whether to push back, and how hard |
| Indemnity and additional insured | Recognizing broad-form language | Enforceability and insurance coverage (attorney plus agent) |
| Pay-if-paid | Telling condition precedent from timing language | State-specific enforceability (attorney) |
| Notice deadlines | Pulling every deadline into one list | Whether your field team can actually meet them |
| Flow-down and incorporated documents | Listing what's missing | Reviewing the documents once you get them |
Once the attorney weighed in, the AI had one more job: drafting a polite, specific request letter to the GC's project manager, listing the proposed changes by section number. Dana edited the tone and sent it. The same idea runs through our guide to what ChatGPT actually does well for service business owners: it drafts, you decide.
Does AI contract review work on service agreements and supplier terms?
Yes, and for many trade owners these documents matter more than GC subcontracts, because they get signed without anyone thinking of them as "contracts." The same prompt works. Just swap in a different list of clause types.
Commercial service agreements
HVAC maintenance contracts with property managers, recurring pest control for retail centers, and janitorial agreements for office parks are usually written on the client's paper. Tell the AI to look for automatic renewal with a narrow non-renewal window, termination for convenience that only the client can use, net-60 or net-90 payment terms, response-time guarantees with service credits, and caps on price increases at renewal. If you already sell your own plans, our piece on AI for maintenance plans and service agreements covers the operations side. The contract review happens before any of that.
Supply house credit applications
This is the one owners miss most. A credit application with a distributor often includes a personal guaranty in small print, a monthly service charge on late balances, venue in the supplier's home county, and attorney's fees if they have to collect. Ask the AI specifically: "Does this document make me personally liable for the company's debt?" A personal guaranty turns a business account into a lien on your own finances, and it rarely comes up at the counter.
Software and equipment agreements
Field service software, fleet telematics, and equipment leases have their own traps. Look for auto-renewal terms, how and when price-change notices arrive, whether you can export your data at termination, and early termination fees. A quick pass before signing helps you avoid being stuck on a platform you've outgrown.
What Dana does now
The weekend turned into a habit, and the habit fits on an index card:
- Every contract above a set value gets the AI pass. For Dana's shop, anything over $25,000 or any recurring agreement longer than a year.
- The prompt is saved, not retyped. One version for GC subcontracts, one for service agreements, one for supplier and software terms.
- Every summary gets checked against the quotes. If the summary and a quoted clause disagree, the quote wins and the summary gets thrown out.
- Missing documents get requested before signing. No prime contract, no signature.
- A one-page packet goes to the attorney. Quoted clauses and specific questions only.
- A clause log tracks negotiation history. A simple spreadsheet noting which GC accepted which changes, so the next negotiation with the same GC starts from what worked before.
In this composite, the GC accepted two of Dana's five requested changes: retainage released on completion of the electrical scope, and a fourteen-day notice window instead of seven. Pay-if-paid stayed. Dana signed anyway, knowing exactly what the risk was and with the Chapter 56 notice procedure written into the project manager's checklist. For shops that also work as general contractors writing their own subcontracts, the same exercise run in reverse is just as useful: have the AI read your template the way your best sub's attorney would.
The AI didn't make Dana's decisions or replace the attorney. It turned three hours of dread into forty minutes of focused checking, and it put the right six questions in front of the right professional. That's a realistic return, and it's enough.
Questions contractors ask about AI contract review
Is it safe to upload a subcontract to ChatGPT or another AI tool?
Use a business or team plan that doesn't train on your uploads by default, or confirm that training is turned off in your settings. Read the contract's confidentiality clause first. Some GC agreements restrict sharing contract documents outside your company and its professional advisors.
Can AI tell me whether a clause is enforceable in my state?
It can point you to the right statute, but it shouldn't make the call. General-purpose tools can be out of date on state law. Texas, for example, rewrote its lien statute in 2022. Verify every citation on the official statute page, and take enforceability questions to an attorney licensed in your state.
What's the difference between pay-if-paid and pay-when-paid?
Pay-when-paid sets timing: the GC pays you within a set window after the owner pays the GC, and you still get paid if the owner is slow. Pay-if-paid makes the owner's payment a condition, so if the owner never pays, the GC may not owe you. Enforceability varies by state, and Texas regulates these clauses in Business & Commerce Code Chapter 56.
Which clauses should I never sign without an attorney reading them?
Broad-form indemnity, pay-if-paid, any personal guaranty, waivers of lien or payment-bond rights, custom lien waiver forms that replace your state's statutory forms, and liquidated damages that flow down from a prime contract you haven't seen. AI can find these clauses reliably. Deciding what to do about them is a legal call.
How long does an AI first pass take on a typical GC subcontract?
With a structured prompt, the tool returns results in a few minutes. Budget 30 to 60 minutes to check each flagged clause against the actual PDF. Skipping that check is how a confident but wrong summary makes it into your cash-flow plan.
Frequently asked questions
Is it safe to upload a subcontract to ChatGPT or another AI tool?
Use a business or team plan that doesn't train on your uploads by default, or confirm that training is turned off in your settings. Read the contract's confidentiality clause first. Some GC agreements restrict sharing contract documents outside your company and its professional advisors.
Can AI tell me whether a clause is enforceable in my state?
It can point you to the right statute, but it shouldn't make the call. General-purpose tools can be out of date on state law. Texas, for example, rewrote its lien statute in 2022. Verify every citation on the official statute page, and take enforceability questions to an attorney licensed in your state.
What's the difference between pay-if-paid and pay-when-paid?
Pay-when-paid sets timing: the GC pays you within a set window after the owner pays the GC, and you still get paid if the owner is slow. Pay-if-paid makes the owner's payment a condition, so if the owner never pays, the GC may not owe you. Enforceability varies by state, and Texas regulates these clauses in Business & Commerce Code Chapter 56.
Which clauses should I never sign without an attorney reading them?
Broad-form indemnity, pay-if-paid, any personal guaranty, waivers of lien or payment-bond rights, custom lien waiver forms that replace your state's statutory forms, and liquidated damages that flow down from a prime contract you haven't seen. AI can find these clauses reliably. Deciding what to do about them is a legal call.
How long does an AI first pass take on a typical GC subcontract?
With a structured prompt, the tool returns results in a few minutes. Budget 30 to 60 minutes to check each flagged clause against the actual PDF. Skipping that check is how a confident but wrong summary makes it into your cash-flow plan.