Here is a truth that the AI vendors will never tell you: the most dangerous thing an AI can do on a construction project is write a confident, well-formatted document that cites the wrong contract clause. Not because the writing is bad, but because contract clauses control your rights, your deadlines, and your money, and a notice that names the wrong clause or asserts a deadline that does not apply can quietly forfeit the very entitlement it was meant to preserve. The AIA A201 General Conditions are the rulebook most of these documents live inside, and AI gets its clauses wrong constantly, conflating provisions that an experienced PM would never confuse. This lesson teaches you the contract clauses that govern AI-drafted documents, the specific ones AI blends together, and how to make sure an AI-assisted notice actually protects you instead of betraying you.

Why the Contract Is the Rulebook AI Plays Inside

Every RFI, ASI response, change order request, notice, and pay-app narrative is not free-floating text; it is a move inside a game whose rules are the contract. The AIA A201 General Conditions of the Contract for Construction is the most widely used set of those rules in the United States, and it defines what a Request for Information is, what authority binds the parties, how disputes resolve, and, critically, the notice and timing requirements that determine whether a claim survives. When AI drafts one of these documents, it is making a move in that game, and a move that violates the rules, misses a deadline, cites a clause that does not apply, asserts an authority the drafter does not have, can lose you the point regardless of how well-written it is.

This is why contract literacy is not optional for using AI safely on the business side of a project. The model can produce fluent contractual prose all day, but it does not understand which clause governs which situation, what the actual deadline is, or who has authority to do what, and it will assert all of those confidently and often wrongly. The verification here is the contract-authority gate from the cardinal-rule lesson, and the only person who can perform it is someone who knows the contract. So we walk the clauses that matter most for AI-drafted documents, and the specific failures the model produces against each.

The Three Clauses AI Conflates Constantly

There is a specific, recurring AI failure in A201 that is worth memorizing because it shows up in real drafted notices and it can be expensive. The model blends together three different clauses that govern three different things with three different timing rules, presenting them as if they were one. Knowing the real distinctions is both your defense against the AI error and a mark of genuine contract literacy.

The first is A201 Section 3.2.4, which concerns the Contractor's review of the contract documents and the obligation to report errors, inconsistencies, or omissions discovered, the provision under which the RFI obligation broadly sits. The key fact AI gets wrong: this clause has no fixed-day window. It is about the duty to review and report, not a countdown clock, yet the model routinely attaches a specific deadline to it that does not exist. The second is A201 Section 15.1.3.1, the actual Notice of Claims provision, which in A201-2017 establishes the twenty-one-day window for claims, the real deadline the model often misattributes to other clauses. The third is A201 Section 3.7.4, governing concealed or unknown physical conditions, which in the 2017 edition requires notice within fourteen days, shortened from the twenty-one days in the prior 2007 edition.

Watch how the AI failure works: asked about a notice deadline, the model confidently produces an answer that mashes these together, attaching the twenty-one-day claims window to the RFI clause, or citing the wrong section for concealed conditions, or applying the old fourteen-versus-twenty-one timing wrong. Each of these is a real-sounding statement that is contractually wrong, and a notice built on the wrong clause and the wrong deadline is a notice that can be challenged on exactly that basis. The discipline is that any clause citation or deadline in an AI-drafted contract document gets verified against the actual executed contract, because A201 is the default rulebook but your project's specific agreement may modify these provisions, and the model knows neither the default reliably nor your modifications at all.

The model blends A201 section 3.2.4 (contractor review and RFI duty, no fixed window), 15.1.3.1 (the 21-day Notice of Claims), and 3.7.4 (concealed conditions, 14-day notice in the 2017 edition) into one confident, wrong answer. Knowing they are three different clauses with three different rules is both the defense and the literacy.

Authority and Definitions: Who Can Do What, and What Words Mean

Beyond the timing clauses, A201 governs two things AI gets wrong in subtler ways: authority and definitions. Authority is about who can bind the parties, who can direct changes, who can issue an instruction that carries contractual weight. An AI that drafts a directive or a change instruction may produce language that purports to commit a party or direct work in a way the drafter is not actually authorized to do under the contract, which is a problem the polished prose hides completely. The contract assigns specific authorities to specific roles, the architect issues an ASI within defined limits, certain changes require the owner's written authorization, and a drafted document that exceeds the drafter's authority is not just wrong, it can create disputes about whether anyone was bound at all.

Definitions matter because A201 gives precise contractual meaning to the very documents AI drafts. An RFI, an ASI, a Change Order, a Construction Change Directive each has a defined role and defined consequences in the contract, and they are not interchangeable. An AI that treats a Construction Change Directive as if it were a Change Order, or drafts something as an ASI that should be a different instrument, produces a document that is mislabeled in a way that carries contractual consequences, because the label determines the process, the pricing rights, and the obligations that attach. The model uses these terms fluently and does not reliably respect their distinct contractual meanings, so a human who knows what each instrument actually is must confirm that the AI-drafted document is the right instrument for the situation, correctly labeled, within the drafter's authority. The prose can be perfect and the instrument still wrong.

The Pay-App Narrative: Where Contract Meets Money

The AIA pay application, the G702 summary and G703 schedule of values, is where contract language and money meet, and the narrative around it is a place AI increasingly helps and must be carefully governed. A pay-app narrative explains and supports the payment being requested, and it sits on top of certified figures that are sworn statements about money. Two A201 dimensions govern this: the provisions about applications for payment and certification, which establish what the application represents and the obligations attached to it, and the provisions about the architect's review and certification role.

The AI failure here combines the dollars gate and the contract-authority gate. An AI-assisted pay-app narrative can misstate what the application represents, overstate completion in a way the certified figures do not support, or characterize the payment request in language that conflicts with the contract's actual payment provisions. Because the underlying figures are sworn and the narrative supports a request for money, an error here is not cosmetic; it touches both the truthfulness of a financial certification and the contractual basis for payment. So the verification is doubled: the figures get the dollars-gate check against the schedule of values and the work in place, and the narrative gets the contract check to ensure it accurately represents the application within the contract's payment provisions and does not assert more than the certified figures and the contract support. The narrative makes the request readable; it must never make the request say something the figures and the contract do not.

Where AI Truly Helps, Done Right

None of this means AI is useless on contract documents; it means its useful role is the language work, not the legal judgment, and keeping that line clean is the whole skill. AI is truly valuable at drafting the prose of a notice once a human has decided the contractual basis, at summarizing a long contract to point you at the clauses that need reading, and at assembling the factual narrative of a claim from the daily reports and the correspondence. In each of these, the human supplies the contractual judgment, which clause applies, what the deadline is, whether the notice is even warranted, and the AI supplies the drafting speed, turning the human's decision into a clean document fast.

The pattern to internalize is decide-then-draft, never draft-then-decide. A PM who decides "this is a concealed-conditions situation, the fourteen-day clock under our contract's version of 3.7.4 started when we hit the rock, and we are giving notice today" and then asks AI to draft that notice is using AI exactly right, because the contractual decisions were made by the competent human and the AI only rendered them into prose. A PM who asks AI "do I have a claim here and what is my deadline" and acts on the answer has inverted the order and handed the legal judgment to a tool that conflates clauses and invents windows. The same tool, the same notice, opposite safety, decided entirely by whether the human made the contractual call before the AI touched the keyboard. Get that ordering right and AI becomes a genuine accelerator of contract administration; get it wrong and it becomes a generator of confident, well-formatted contractual mistakes.

Why the Model Structurally Cannot Know Your Contract

It is worth being precise about why AI is so unreliable on contract clauses specifically, because the reason tells you exactly how to compensate. There are two layers of ignorance stacked on top of each other. The first layer is the training-cutoff and prediction problem you already know: the model learned contractual language as patterns, so it produces plausible clause numbers and deadlines the same way it produces plausible spec sections, by prediction rather than retrieval, which is why it conflates the three clauses and invents windows. Even on the standard A201 defaults, it is guessing in the shape of confidence.

The second layer is the one people forget, and it is decisive: your project is almost never on the bare A201. Owners and their counsel routinely modify the General Conditions through supplementary conditions, striking clauses, changing deadlines, altering notice recipients, tightening or loosening dispute provisions, so the rules that actually govern your project are the A201 as amended by your specific agreement. The model has never seen your supplementary conditions. It cannot have, because they are private to your contract. So even if the model somehow recited the A201 default perfectly, it would still be wrong wherever your owner modified it, and it has no way to know where that is. This double ignorance, unreliable on the default and blind to your modifications, is why the verification cannot be against the model's knowledge or even against a generic A201, but only against your executed agreement. The model is answering a question about a rulebook it half-remembers and has never seen your edition of.

This reframes the whole task. You are not checking whether the AI got the contract right; you are using your knowledge of your contract to verify a draft the AI produced, because the contract is the authority and the AI is the assistant. That ordering, your contract first, the AI's draft second, is the entire posture, and it is why a PM's contract literacy is not made obsolete by AI but made more leveraged, since it is now the verification layer that lets the firm safely use AI for the drafting volume.

The Applied Problem: Map the Clauses That Govern AI-Drafted Documents

Here is the exercise that turns contract literacy into a working safeguard. Using A201-2017, or better your project's actual executed General Conditions, identify and highlight the specific clauses that govern the AI-drafted documents you actually produce. Build two short lists, because they protect two different things.

First, the clauses that constrain an AI-drafted notice. Find the four that matter most for your notices: the provision governing the Contractor's review and reporting obligation, the Notice of Claims provision with its actual deadline, the concealed-conditions provision with its distinct shorter deadline, and the provision governing how and to whom notice must be given. Mark each with what it actually requires, the real deadline, the real recipient, the real trigger, so you have a reference that catches the AI's conflation the moment it appears in a draft. Second, the clauses that govern an AI-assisted G702 narrative: the application-for-payment-and-certification provision and the architect-review-and-certification provision, marked with what the narrative may and may not assert.

The deliverable is a one-page contract reference keyed to your AI-drafted documents: for each document type, the clauses that govern it and what they actually require, drawn from your real contract, not the model's memory. This is the contract-authority gate made concrete, the thing you check an AI-drafted notice or narrative against before it goes out, and it is built from your executed agreement precisely because the model does not reliably know even the A201 defaults and knows nothing of your project's modifications. A PM who has this page catches the conflated clause, the wrong deadline, the exceeded authority, and the mislabeled instrument before any of them reach a counterparty, which is the difference between AI that accelerates your contract administration and AI that quietly undermines it. Every contract-touching lesson in the next level produces a document that must clear this exact reference.

Key Takeaways

• Contract clauses control your rights, deadlines, and money, so the most dangerous AI error on the business side is a confident document citing the wrong clause or asserting a deadline that does not apply, which can forfeit the entitlement it was meant to preserve.
• A201 is the rulebook most AI-drafted documents (RFIs, ASIs, notices, pay-app narratives) live inside, defining authority, definitions, dispute resolution, and the notice timing that determines whether a claim survives. The model produces fluent contractual prose without understanding which clause governs.
• The signature AI failure conflates three clauses: A201 3.2.4 (contractor review and RFI duty, no fixed-day window), 15.1.3.1 (the 21-day Notice of Claims), and 3.7.4 (concealed conditions, 14-day notice in A201-2017, shortened from 21 in the 2007 edition). Knowing they are three different clauses with three different rules is the defense.
• AI also errs on authority (drafting language that exceeds the drafter's power to bind or direct) and definitions (treating a Construction Change Directive as a Change Order, or mislabeling an instrument), because the label determines the process, pricing rights, and obligations.
• Pay-app narratives combine the dollars gate and the contract-authority gate: the figures are sworn and the narrative must accurately represent the application within the contract's payment provisions, never asserting more than the certified figures and the contract support.
• Verify every clause citation and deadline against the actual executed contract, because A201 is only the default and your project's agreement may modify these provisions, and the model knows neither the default reliably nor your modifications at all.
• The artifact: a one-page contract reference keyed to your AI-drafted documents, listing the clauses that govern each and what they actually require, built from your executed agreement. It is the contract-authority gate made concrete, catching conflated clauses, wrong deadlines, exceeded authority, and mislabeled instruments before they reach a counterparty.