Yes, AI can review a construction contract as a first reader. It is good at summarising payment terms, notice periods, retention and liquidated damages, and at pointing to one-sided wording. It is poor at what the contract leaves out, at conflicts between documents, and at telling you whether a clause will hold up. It is not legal advice.
The useful way to think about it: the assistant is a fast, tireless junior who has read the text once and never visited a site. It will give you a confident summary in two minutes. Your job is to decide which parts of that summary you trust, which you check against the original, and which go to a lawyer before you sign.
What a first pass can realistically give you
Upload the contract to ChatGPT or Claude and ask for a structured read, and you get three things reliably. First, a plain-English summary of who pays what and when. Second, a list of obligations that fall on you rather than on the other side. Third, a set of clauses it considers unusual, one-sided or unclear. That is worth having, because most small contractors read the price, the programme and the start date, and skim the rest.
What you do not get is assurance. A model reads the words it has been given. It does not know which edition of a standard form you are on unless the file says so, cannot see the drawings that the contract incorporates by reference, and has no idea how your local courts treat a given clause. Treat every output as a reading list, not a verdict. For a comparison with purpose-built tools, see what builders should test in contract review software.
Set the file up so the AI reads all of it
Most review mistakes start before the first prompt. Do these four things first.
- Combine the pieces. A construction contract is rarely one document. There is an agreement, conditions, a scope or specification, a price schedule and sometimes a programme. Upload every piece, and name each file clearly ("1 Agreement", "2 Conditions", "3 Scope", "4 Price schedule"). Tell the assistant which document wins if two conflict, if the contract says so.
- Check the PDF has real text. Click and drag over a paragraph. If you cannot highlight it, the file is a scan. Run it through text recognition first, or ask the assistant at the start to state how many pages it can read. The limits differ by tool: Claude looks at images and layout only in PDFs of 100 pages or fewer and reads the text layer alone from 101 to 1,000 pages, so a scanned price schedule at page 140 of a tender bundle can come through blank. It reads only the text of a Word file, never its embedded images. OpenAI's help page on file uploads does not say how ChatGPT treats image-only scans, so test one page before trusting the rest. Splitting a long bundle into its parts (agreement, conditions, scope, schedules) avoids most of this.
- Ask for the page count back. Begin with "How many pages and which schedules did you receive?" If it says 38 and your file has 52, find out why before you continue. When you later ask about a specific page, use the page number your PDF viewer shows, not the number printed in the footer; Claude's help centre gives the same advice, and contracts with roman-numbered front matter trip this up constantly.
- Decide what you may share. Many contracts carry a confidentiality clause. Business plans from the main vendors do not train on your content by default, and consumer plans have a switch for it, but your contract may still bar disclosure to third parties. When in doubt, redact the client's name, the site address and the price.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowThe prompt that works for a first read
Asking "is this contract OK?" gets a bland reassurance. Ask for a structured extraction with the clause number and a quote, so you can verify each line in seconds. Here is a prompt you can paste, then adapt.
You are helping a small building contractor read a subcontract before signing. I am the subcontractor. This is a first read, not legal advice.
For each item below, give: the clause number, a word-for-word quote (max 40 words), a one-line plain-English meaning, and a flag of GREEN (normal), AMBER (unusual or unclear) or RED (heavily one-sided). If the contract is silent on an item, write "NOT ADDRESSED" and say what that could mean for me.
Items: payment due dates and certification; pay-when-paid or pay-if-paid wording; retention percentage and release; liquidated damages; variations and how they are priced; notice periods and time bars; extension of time; indemnities; insurance requirements; termination by either side; defects period; dispute process; warranties and design responsibility; set-off and deduction rights.
Finish with the five clauses I should ask a lawyer about first, ranked, and any place where two documents seem to conflict.The "NOT ADDRESSED" instruction does most of the work. Without it, models tend to skip items the contract never mentions, which is exactly where the risk hides. The same structure carries over to other paperwork; the ChatGPT prompt for flagging risky clauses in a commercial lease uses it for a yard or unit lease.
Clauses AI flags well
These clauses are self-contained, use fairly standard language and sit in one place, so a model finds and explains them well.
Payment terms and retention
Due dates, interim certificates and retention percentages are easy to extract. Say a subcontract states payment 45 days after invoice, retention of 5 percent held until a year after completion, and a right to deduct "any sums the contractor considers due". The AI will list the 45 days, the 5 percent and the year, and will usually flag the deduction wording as AMBER or RED. That is a correct and useful catch. A 45-day term is not itself wrong, but it changes your cash flow, so run the numbers: on a $60,000 subcontract billed monthly, a 5 percent retention ties up $3,000 for a year or more.
Liquidated damages and delay
A clause that charges a daily sum for late completion is easy to spot. Models also tend to notice when the daily rate looks large against the contract price, though "large" is a judgement call. A $500 daily charge on a $40,000 job deserves a question. A $500 daily charge on a $4 million job probably does not.
Pay-when-paid wording
Phrases such as "payment to the subcontractor is conditional on receipt of payment from the employer" are a classic. Models recognise them reliably. What they cannot tell you is whether that wording is permitted where you work, because some places restrict or ban it. That is a lawyer question, and the AI's flag is simply your reminder to ask it. The subcontractor's pay-when-paid check before signing works through what that wording does to your cash flow.
Insurance and indemnity requirements
A model will list the cover limits, the types required and whether you must name the other party on your policy. Compare that list with your own certificate. This is a task where the assistant saves real time, because the requirements are often scattered across three clauses and a schedule.
Termination and notice periods
Short notice periods, one-sided termination rights and cure periods are typically found and summarised well. A realistic catch: the contractor may terminate on 7 days' notice, while you need 28 days and a formal warning. The model will line the two up in a table if you ask.
Clauses AI misses or gets wrong
These are the failures that matter, because they feel safe. The summary looks complete, and the gap is invisible.
| Where it goes wrong | Why it happens | How to catch it |
|---|---|---|
| What the contract does not say | The model reads text, not absences. Silence on weather delays, access or design responsibility produces no flag. | Make "NOT ADDRESSED" a required answer for a fixed list of topics. |
| Conflicts between documents | Long files exceed what the model holds in mind. A specification can contradict the conditions on a time bar. | Ask a second, separate question: "List every clause in the scope that contradicts the conditions." |
| Cross-references and defined terms | A clause says "as defined in Schedule 4" and the model does not follow it. "Completion" may mean three things. | Ask it to list every defined term and the clause that defines it. |
| Amendments and special conditions | Edits to a standard form often sit in a separate schedule that overrides the printed clause. | Ask: "Which standard clauses are amended, and how?" Check the schedule yourself. |
| Time bars buried in other clauses | A notice deadline may live in the variations clause, not under "notices". | Search the contract for "within" and "days" and compare with the model's list. |
| Numbers in tables and schedules | Tables lose structure in some extractions, so rates and percentages can shift columns. | Read every figure against the source page. |
| Local law | The model may apply general principles that do not match the rules where you work. | Ask a lawyer who works under the governing law. |
Running the prompt on an $84,000 house-extension subcontract
Take an illustrative case. A three-person groundworks and brickwork firm is offered a subcontract by a larger builder for a house extension. Price: $84,000. The firm uploads the agreement, the conditions and the scope, and runs the prompt above.
The assistant returns a table. Payment: monthly, 30 days from the valuation date (GREEN). Retention: 5 percent, half released at practical completion, half after 12 months (GREEN). Liquidated damages: $350 a day (AMBER, about 0.4 percent of the price per day). Variations: "only if instructed in writing before the work starts" (AMBER). Set-off: the builder may deduct "any losses it considers reasonable" (RED). Pay-when-paid: present (RED). Design responsibility: NOT ADDRESSED.
Here is where the human part begins. The owner checks the three highest-risk items against the original text. The variations clause does say written instruction first, but a later clause in the scope says the contractor "shall not be entitled to additional payment for work reasonably inferable from the drawings". The assistant had flagged the first and missed how the second narrows it. The owner adds a note: "Ask for a change so that instructions given verbally are confirmed by the builder within 48 hours."
The design-responsibility gap is the most valuable finding. The scope shows a foundation detail but does not say who is responsible if the ground conditions differ from the drawing. The firm adds it to a question list for the lawyer, along with pay-when-paid and set-off. That is a one-hour conversation rather than a four-hour review, and the lawyer's time goes where it matters. The whole exercise took 40 minutes on an ordinary chat subscription, and it caught the gap a skim would have missed.
Check the flags against the original
For every AMBER and RED flag, do three things. Find the clause number the assistant quoted. Read the full clause, not only the quote. Then read the clauses before and after it, because a limiting phrase often sits in the next paragraph.
Also spot-check a few GREEN items, say three out of fifteen. If one of them turns out to be wrong, assume others may be, and slow down. A good habit is to ask the model to critique its own answer: "Which of your GREEN ratings are you least sure about, and why?" A second look often surfaces a qualification the first pass smoothed over, though it can also invent doubt, so the original text stays the referee.
A risk checklist to run before you sign
Print this and run it with the AI's table beside you. Each line is a question for the contract, not for the model.
- Money in: when is each payment due, what triggers it, and can the other side delay or reduce it?
- Money held back: retention percentage, release dates, and whether it is held in a separate account.
- Money out: deductions, set-off rights, back-charges and who pays for the other side's delay.
- Time: start date, completion date, who can extend it, and the notice you must give to claim more time.
- Changes: how variations are instructed, priced and approved, and what happens when instructions are verbal.
- Risk you carry: ground conditions, existing structures, design errors, third-party delays, materials price rises.
- Insurance: required cover, who is named, and whether your current policy meets it.
- Ending the deal: termination rights, notice periods and what you are paid if it ends early.
- After handover: defects period, warranties, and any long-stop liability.
- Disputes: process, deadlines, who pays costs.
- Missing items: anything on the list the contract does not mention.
More examples of what AI does with specific clauses
A before-and-after on a notice clause
The original text: "The Subcontractor shall give written notice of any event likely to cause delay or additional cost within 7 days of becoming aware of it, failing which the Subcontractor shall have no entitlement to an extension of time or additional payment."
An illustrative plain-English rewrite from the assistant: "You have 7 days to tell them in writing about anything that delays you or costs more. Miss that, and you lose the right to claim." That is accurate and, importantly, it tells you what to do on site: if rain stops work on Tuesday, a written notice goes out by the following Tuesday. Ask the assistant to turn every time bar into a calendar entry or a one-line site rule and you have something the whole team can use.
A realistic mistake: the quoted number that was not in the file
A contractor asked a general question about retention and the assistant replied "Retention is 3 percent, released at completion". The contract in fact said 5 percent, with half released at completion. The model had blended the wording with a typical figure. The fix is the quote requirement in the prompt: if the model cannot give a word-for-word quote with a clause number, treat the answer as unreliable. Never rely on a number the model cannot point to.
A quick sum on liquidated damages
Ask the assistant to turn the clause into exposure. Contract price $120,000, programme 16 weeks, damages $400 a day. Four weeks late is 28 days, so $11,200, which is 9.3 percent of the price. Ask it to show the arithmetic and then check it yourself with a calculator. Models make mistakes in multi-step maths, as this look at AI and arithmetic explains. The point is not the sum itself. It is seeing that a month's delay could wipe out a typical margin.
Comparing two versions of a contract
When the other side sends a revised draft, upload both and ask: "List every difference between version 1 and version 2, grouped by clause, and mark which side each change favours." This works well for wording changes and badly for formatting noise, so tell it to ignore layout. It is a strong use of AI, because a person comparing two 40-page documents by eye will miss changes.
Questions for the other side
Ask the assistant to draft a short email of clarifying questions from your RED and AMBER flags, in a neutral tone, without conceding anything. Review it before sending. A good output reads like: "Clause 14.3 allows deductions for losses you consider reasonable. Could you confirm that deductions will be itemised in writing and notified before payment is due?" That is a better opener than a flat refusal.
Where the lawyer still comes in
Use the assistant to shorten the lawyer's job, not to replace it. Send your lawyer the AI's table, your verified notes and your five questions. You are paying for judgement about enforceability, the governing law, how the wording has been read by courts, and what to push back on. Those are the things a language model cannot see.
A rough rule: if the contract is small and on terms you have signed before, an AI read plus your own checks may be enough. For anything large, for any new type of work, or for any contract that moves design risk or ground risk to you, get advice. If you want to set up a repeatable routine for this with your own contract types, that is what an AI implementation consultation can cover, but the routine above works without it.
Keeping your review notes useful
After each review, save three things in one folder: the contract version, the AI's table with your corrections, and the questions you sent. Next time you see the same builder or client, you can run the new contract against your last notes and ask what changed. Over a year this becomes a small library of how each counterparty writes its terms.
Also keep a short list of clauses you will never accept without discussion, such as unlimited liability, pay-if-paid, or open-ended set-off. Put it into the prompt ("flag any of these as RED automatically"). That turns your experience into a standing instruction, and it is the single best way to make the assistant behave more like someone who knows your business. For how contract terms feed through to profit by job, see construction job costing with AI.
Questions contractors ask before uploading a contract
Is it safe to upload a client's contract to ChatGPT or Claude?
Check the contract's confidentiality clause first, then the plan you are on. Business plans from both vendors do not train on your content by default, while individual plans, free or paid, need the model-training switch in privacy settings turned off. If the contract bars sharing with third parties, redact names and prices before uploading.
Can AI tell me whether a clause is enforceable?
No. Enforceability depends on the law that governs the contract, the facts and how a court reads the wording. AI can say a clause looks one-sided or unusual compared with common practice. Whether it holds up is a question for a qualified construction lawyer.
How long should an AI first pass take?
For a 40-page subcontract, plan on 30 to 45 minutes: 10 to run the structured read, 15 to check the flagged clauses against the original text, and the rest to list questions. If it takes five minutes, you probably skipped the verification step.
Should I paste the contract or upload the file?
Upload the file when the tool accepts it, because tables and schedules survive better than pasted text. Scanned pages are the exception: Claude reads page images only in PDFs of 100 pages or fewer, and a Word file's embedded images not at all, so run text recognition first on any scanned schedule.
Further reads
- AI Tools for Contracting Businesses: Subcontracts and Change Orders — Subcontract and change-order paperwork once the contract is signed.
- AI Contract Review for Small Firms: What It Catches and Misses — The same first-pass method for non-construction contracts.
- AI Contract Management Software Demo: 10 Questions for Contractors — Ten questions to ask when a vendor demos contract tools.
- How Small Builders Use AI to Price Variations and Change Orders — Price the variations your contract should already control.
- AI in Construction Accounting: Job Costs, WIP and Retainage Checks — See how payment terms flow into job costs and WIP.
- How to Evaluate an AI Implementation Proposal or Quote — Read supplier proposals with the same sceptical checklist.
- Can AI Answer Building Code Questions? A Builder's Safe-Use Routine — A builder's routine for using ChatGPT or Claude on building code questions: find the section, check the edition, confirm with the building department.
- Can AI Draft a Remodelling Contract Template in Plain English? — Prompts, sample clauses and a checklist for drafting a remodelling contract with AI, and the terms a solicitor must still review.
- Can AI Draft a Scaffolding Hire Agreement? Terms a Solicitor Checks — Hire period, extension, damage, access and alteration clauses drafted with AI, plus the liability terms a solicitor must review.
- Construction WIP Report: What It Shows and How AI Builds One — What a work-in-progress schedule shows, a three-job worked example of over and under billing, and where AI helps and where it must not.
- Material Price Escalation Clause: What AI Can and Cannot Check — Use AI to summarise an escalation clause and list its triggers, then take a short confirm-list to your solicitor.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: ChatGPT and Claude plan and privacy pages; Anthropic support page on uploading files to Claude (PDF page limits); OpenAI help page on uploading files to ChatGPT; general construction-contract practice. This tutorial is not legal advice.