AI Capability

Can AI review a construction contract?

Yes, partially. AI reliably catches missing clauses, deviations, dates, and flow-down gaps. It cannot tell you whether a risk is acceptable. Here is the full capability line, with sources.

On this page
  1. Can AI review a construction contract?
  2. What does automated contract review reliably catch?
  3. What does automated contract review catch unreliably?
  4. What can AI genuinely not do on a construction contract?
  5. How reliably can each part of contract review be automated?
  6. How does a general-purpose AI assistant fail differently from a purpose-built review tool?
  7. Is it safe to paste contract text into a general-purpose AI tool?
  8. What does a contract review workflow that actually works look like?
  9. What does this page not cover, and when do you need a lawyer?
  10. Frequently asked questions

An honest assessment of what automated contract review catches, what it misses, and where a human or a lawyer is still required.

Can AI review a construction contract?

Yes, partially. AI can reliably review a construction contract for what is present, absent, or inconsistent against a known standard: missing required clauses, deviation from your playbook, dates, dollar values, flow-down gaps, and defined terms used inconsistently. It cannot decide whether a risk is acceptable to your business, and it cannot give legal advice.

The gap between those two halves is where most of the confusion in this category lives. Automated review is genuinely good at the mechanical comparison work that eats the first several hours of any contract review: reading every page without fatigue, finding every date, checking every required clause against a list, and flagging what does not match. It is not good, and in some cases is actively unreliable, at the work that comes after, which is deciding what a flagged item actually means for this project, this counterparty, and this firm’s appetite for risk.

The reliability question is measurable, and the measurements are sobering. A 2024 Stanford RegLab and Institute for Human-Centered AI study, the first preregistered empirical evaluation of AI-driven legal research tools, tested leading commercial products that had marketed themselves as eliminating or avoiding fabricated output. The researchers found those tools “each hallucinate between 17% and 33% of the time.” Those are purpose-built legal tools built by large publishers, not general chatbots, and they were wrong on roughly one query in five at the low end.

That does not make automated review useless. It makes the distinction between tasks the whole ballgame. On a market where the U.S. Census Bureau estimated construction spending during May 2026 at a seasonally adjusted annual rate of $2,210.2 billion, the volume of contract paper moving through the industry is far past what anyone reviews line by line anyway. The realistic comparison is not automated review against a perfect human read. It is automated review against the read that actually happens on a Friday afternoon with three contracts due Monday.

What does automated contract review reliably catch?

Automated review reliably catches five things: missing required clauses, deviation from a known standard, dates and dollar values, flow-down gaps, and inconsistent defined terms. All five share one property, which is that they are comparisons against something already known rather than judgments about something new.

Missing required clauses. If you can list what must appear, a tool can confirm whether each item appears. This is the single most dependable thing automated review does, and it maps directly onto federal work. FAR 52.244-6, Subcontracts for Commercial Products and Commercial Services, enumerates 24 specific clauses at paragraph (c)(1) that a contractor must insert into subcontracts for commercial products or services. Twenty-four is more than anyone checks reliably by hand across a portfolio of subcontracts, and it is trivial for software.

Deviation from a known standard. Given your negotiated positions or a reference version of your subcontract, a tool can surface every place the counterparty’s paper departs from it. This is comparison, not interpretation, and it is the highest-value mechanical output in the whole workflow because redlines hide in language that reads as ordinary.

Dates, dollar values, and notice periods. Extraction is reliable. Notice windows, liquidated damages rates, retainage percentages, insurance limits, and milestone dates come out consistently and can be pushed into a calendar or a tracker.

Flow-down gaps. Comparing a prime contract’s required clauses against what actually landed in an executed subcontract is exactly the kind of set comparison software does well and people do badly, particularly at the third and fourth tier where the paper is thinnest.

Inconsistent defined terms. When a contract defines “Work” in one place and uses it differently in another, or when “Substantial Completion” carries two meanings, a tool catches it. Whether the inconsistency matters is a separate question.

What does automated contract review catch unreliably?

Three categories are caught inconsistently enough that you should never rely on a clean automated result: clause interactions, ambiguity that depends on course of dealing, and incorporation by reference to documents the tool was never given. In each case the tool may produce an answer, and the answer may be wrong while sounding exactly as confident as a correct one.

NIST names this failure mode directly. In its Generative Artificial Intelligence Profile, NIST AI 600-1, published July 2024, NIST defines the risk as:

“Confabulation: The production of confidently stated but erroneous or false content (known colloquially as ‘hallucinations’ or ‘fabrications’) by which users may be misled or deceived.”

NIST goes on to explain that confabulations “are a natural result of the way generative models are designed” and that the risk “may be especially important to monitor when integrating GAI into applications involving consequential decision making.” Signing a construction contract is consequential decision making.

Clause interactions are the most expensive of the three. An indemnity provision, a limitation of liability, a consequential damages waiver, and an insurance requirement each read as survivable on their own. Their combined effect is where money is actually lost, and evaluating a combination requires holding facts that are not in the document, such as what your policy actually covers.

Ambiguity that depends on course of dealing is not a text problem at all. When language is genuinely ambiguous, the meaning often lives in how the parties have behaved on prior projects. That history is not in the file, so no amount of reading the file resolves it.

Incorporation by reference is the failure mode most specific to construction, because construction contracts incorporate more than almost any other agreement: drawings, specifications, general conditions, exhibits, and the prime contract above them. FAR 52.236-21, Specifications and Drawings for Construction, states the operative rule for federal work:

“Anything mentioned in the specifications and not shown on the drawings, or shown on the drawings and not mentioned in the specifications, shall be of like effect as if shown or mentioned in both. In case of difference between drawings and specifications, the specifications shall govern.”

Read that as a capability statement about software. An obligation that appears only in the drawings binds you exactly as if it appeared in the contract text. A tool handed only the contract document will report a clean review and be wrong, not because it reasoned badly, but because the governing language was never in front of it.

What can AI genuinely not do on a construction contract?

Four things, and none of them are engineering problems that a better model will solve: decide whether a risk is acceptable, know your bonding capacity and financial position, know your history with the counterparty, and give legal advice. These are not gaps in the technology. They are questions whose answers do not exist inside the document.

Whether to accept a no-damage-for-delay clause is not a property of the clause. It is a function of your margin on the job, your schedule confidence, your current backlog, how much you want this client, and what your surety will say. A tool that has read only the contract has none of that. It can tell you the clause is there and that it differs from your standard position. It cannot tell you to sign.

NIST’s AI Risk Management Framework, NIST AI 100-1, released January 2023, places this boundary at the center of its guidance, stating that “human judgment should be employed when deciding on the specific metrics related to AI trustworthiness characteristics and the precise threshold values for those metrics.” The framework also cautions that “addressing AI trustworthiness characteristics individually will not ensure AI system trustworthiness; tradeoffs are usually involved.” Applied to contract review, the threshold for what counts as an unacceptable risk is set by your business, not by the software.

Legal advice is the hardest line of the four. Telling you whether a clause is enforceable in a given state, what a court has done with similar language, and what you should do about it is the practice of law. Automated review is a detection layer that tells a person where to look. The advice still comes from a lawyer.

How reliably can each part of contract review be automated?

The table below is the honest version. Each row states the review task, how reliably it can be automated today, what still requires a person, and why the boundary sits where it does.

Review taskHow reliably it can be automatedWhat still needs a humanWhy
Finding missing required clausesHighConfirming the clause list is correct for this contractThe tool checks against the list you supply. It has no way to know the list itself is wrong or out of date.
Deviation from a known standardHighDeciding whether the deviation is acceptableDetection is comparison. Acceptance is a commercial judgment about this deal.
Dates, dollar values, notice periodsHighConfirming which document governs when values conflictExtraction is dependable. Precedence between conflicting documents is a legal question.
Flow-down gaps in subcontractsHighConfirming what the prime contract and agency supplements actually requireThe required clause set comes from above you in the chain, not from the subcontract being reviewed.
Inconsistent defined termsHighDeciding which meaning was intended and whether it mattersDetecting inconsistency is mechanical. Resolving intent is not.
Clause interactions and combined effectMediumReading the risk-shifting clauses together against your actual coverageEffects compound in ways that depend on facts outside the document, including what your policy covers.
Ambiguity that depends on course of dealingLowThe entire analysisThe meaning lives in how the parties have behaved on prior work. That history is not in the text.
Incorporation by reference to documents not providedLowRetrieving and reading every incorporated documentSoftware cannot review a document it was never given. Construction contracts incorporate heavily.
Whether a risk is acceptableNot automatableThe decision itselfDepends on margin, schedule confidence, backlog, and appetite. None of it is in the contract.
Bonding capacity and financial exposure fitNot automatableWhoever owns bonding and marginThe surety's position and your capacity are external facts the document does not contain.
Relationship history with the counterpartyNot automatableYour team's memory and project recordsNot in the document, and frequently not written down anywhere.
Legal advice and enforceability opinionsNot automatableA licensed attorneyEnforceability is jurisdiction-specific and advising on it is the practice of law.

The pattern is consistent all the way down. Everything in the “High” band is a comparison against a known reference. Everything in the “Not automatable” band requires information that does not exist in the contract at all.

How does a general-purpose AI assistant fail differently from a purpose-built review tool?

They fail in opposite directions. A general-purpose assistant fails by overreaching, answering questions it has no grounding to answer. A purpose-built review tool fails by under-reaching, missing anything nobody configured it to look for. Knowing which failure you are exposed to determines what you have to check.

The overreach problem is well documented. In “Large Legal Fictions: Profiling Legal Hallucinations in Large Language Models,” published in the Journal of Legal Analysis in 2024, researchers found that legal hallucinations were “alarmingly prevalent, occurring between 58% of the time” and 88% of the time depending on the model, when general-purpose systems were asked specific, verifiable questions about random federal court cases. The same research found that these systems “often fail to correct a user’s incorrect legal assumptions” and “cannot always predict, or do not always know, when they are producing legal hallucinations.”

That last finding is the operationally dangerous one. A general-purpose assistant asked whether your indemnity clause is enforceable will answer. It will not stop to say the question is outside what it can determine from the document. If you ask a leading question, it will frequently agree with you.

A purpose-built review tool has the opposite profile. It typically checks a defined set of items against a defined standard and reports what it found. Its confident output is more trustworthy inside that scope, because the scope was constrained on purpose. The catch is that a clean result means only that nothing in the configured checklist failed. It is not a statement that the contract is fine. If your checklist never included a no-damage-for-delay provision, a clean report will never mention one.

Neither profile is safe to trust blindly, and the fix is different for each. With a general-purpose assistant, verify what it told you. With a purpose-built tool, ask what it was never asked to look for.

Is it safe to paste contract text into a general-purpose AI tool?

Not without checking two things first: what the tool does with the text you submit, and what your own contract says about disclosure. This is the risk most often skipped, because pasting is so easy that it does not feel like sending a document to a third party. It is.

Most construction contracts carry a confidentiality obligation covering pricing, terms, and project information. Pasting that text into a consumer tool can constitute disclosure to an outside party, and the fact that it was convenient is not a defense. Consumer and enterprise tiers of the same product frequently have materially different data handling terms, so the relevant question is what your specific account is subject to, not what the vendor’s marketing says generally.

Federal work adds a second, harder layer. FAR 52.204-21, Basic Safeguarding of Covered Contractor Information Systems, applies to systems that process federal contract information, which the clause defines as:

“Information, not intended for public release, that is provided by or generated for the Government under a contract to develop or deliver a product or service to the Government, but not including information provided by the Government to the public (such as on public websites) or simple transactional information, such as necessary to process payments.”

A federal subcontract and its pricing sit squarely inside that definition. Notably, FAR 52.204-21 is itself one of the 24 clauses required to flow down under FAR 52.244-6, which means the obligation travels down the subcontract chain. Where controlled unclassified information is involved, the requirements are more stringent still.

NIST also flags the commercial exposure directly, listing among generative AI risks the “eased exposure of trade secrets” in its Generative Artificial Intelligence Profile. Your bid strategy and your negotiated positions are trade secrets in every practical sense.

The workable rule is simple. Decide in advance which tools are approved for contract text, write it down, and make sure the people reviewing contracts at 6pm know what it says.

What does a contract review workflow that actually works look like?

The machine flags, and the human decides. That single division of labor is what separates a workflow that captures the speed benefit from one that quietly imports the tool’s errors into your contract file.

Step one: define the standard before the review. Write down the clause list, the positions you will not move off, and the limits your insurance and bonding actually support. Automated review measures a contract against a standard, so if the standard is only in a senior PM’s head, the tool has nothing to measure against.

Step two: gather every incorporated document. Drawings, specifications, general conditions, exhibits, and the prime contract above the subcontract. This step is skipped constantly and it is the one that produces false clean results.

Step three: run the mechanical pass. Missing clauses, extracted dates and values, flow-down coverage against the prime, deviations from standard positions, inconsistent defined terms. This is the part that turns a three-hour read into a thirty-minute one.

Step four: verify every flag against the source text. Open the contract at each flagged location and confirm the finding is real. Verification is quick precisely because the tool told you where to look. Skipping it means adopting the tool’s errors as your own.

Step five: decide, escalate, or negotiate. A person decides whether each item is acceptable, needs negotiation, or needs counsel.

NIST identifies the specific way this workflow degrades, listing “Human-AI Configuration” as a named risk category covering arrangements where a person experiences “automation bias, over-reliance, or emotional entanglement with GAI systems.” Over-reliance in contract review looks mundane. It looks like a reviewer who has seen the tool be right forty times and stops opening the source document on the forty-first.

For how this fits into day-to-day contract administration rather than a single review, see AI contract administration in construction→. For how Contracts Connected handles a vendor’s changes, see redlines inside the contract→.

Related reading:

What does this page not cover, and when do you need a lawyer?

This page assesses capability, not enforceability, and it is not legal advice. It does not tell you whether any specific clause will hold up in your jurisdiction, what any specific contract means, or whether you should sign it. Those questions require a lawyer licensed where the project sits, and nothing here substitutes for that.

Three more honest limits are worth stating plainly.

Capability is a moving target. Every reliability judgment on this page describes tools as they behave in mid-2026. The boundary between “unreliable” and “reliable” has moved before and will move again. The boundary around risk acceptance, bonding capacity, and legal advice will not move, because those depend on information that is not in the contract rather than on model capability.

We are not neutral, and you should read accordingly. Contracts Connected sells automated contract and compliance review. That is exactly why the concessions on this page are specific rather than vague. A vendor telling you the technology handles everything is telling you something the published research does not support.

This page does not evaluate specific products. It describes categories of tool behavior. Any claim about how a particular tool performs on your contracts should be tested on your contracts, with a set you already know the answers to, before it is trusted on ones you do not.

Where the answer depends on the law rather than the document, ask a lawyer. Where it depends on your capacity or appetite, ask the people who own those numbers. Where it depends on what the paper actually says, that part can be automated, and it is the part that wastes the most human hours today.

Review is one part of the job. To see the AI across the whole contract, from the first draft to the final lien waiver, see AI for construction contracts.

Frequently asked questions

Can AI review a construction contract?

Partially, and the boundary is specific. AI reliably catches what is present, absent, or inconsistent when measured against a standard you supply: missing required clauses, deviation from your negotiated positions, dates and dollar values, flow-down gaps, and defined terms used inconsistently across a document. It cannot decide whether a risk is acceptable to your business, and it cannot give legal advice. Treat it as a first pass that narrows what a human reads, not a replacement for the read.

Can AI replace a construction attorney?

No. Interpreting whether a clause is enforceable in a specific jurisdiction, advising on the consequences of accepting it, and negotiating on your behalf are the practice of law. Automated review can tell you a no-damage-for-delay provision is present and how it differs from your standard position. Whether that provision will hold up in your state, and what it will cost you if it does, is a question for counsel.

Is it safe to paste a construction contract into a general-purpose AI assistant?

Not without checking two things first. Most construction contracts carry a confidentiality obligation, and pasting the text into a consumer tool can be a disclosure to a third party. Federal work adds a second layer: FAR 52.204-21 requires basic safeguarding of federal contract information on contractor systems. Check the tool’s data handling terms and your own contract before any contract text leaves your environment.

How accurate is AI contract review?

Accuracy depends entirely on the task. Extraction and comparison tasks, finding dates, limits, and missing clauses, are reliable enough to act on with verification. Interpretive tasks are not. A 2024 Stanford study of leading commercial AI legal research tools found they produced incorrect information between 17 and 33 percent of the time, and general-purpose models tested on verifiable questions about federal cases did substantially worse.

What is the difference between a general-purpose AI assistant and a purpose-built contract review tool?

They fail in different directions. A general-purpose assistant will answer any question you ask about a contract, including questions it has no grounding to answer, and it will sound equally confident either way. A purpose-built tool checks a defined set of things against a defined standard and is more likely to stay silent outside that set. The first fails by overreaching. The second fails by missing what nobody configured it to look for.

What should a human always check after an automated contract review?

Check four things. First, that the clause list the tool measured against is actually the right list for this contract. Second, every document incorporated by reference that the tool was never given. Third, the combined effect of the risk-shifting clauses rather than each one alone. Fourth, anything the tool flagged as an exception, because the decision to accept or negotiate is yours, not the software’s.

Contracts Connected is the Back-Office System (BOS) for construction.

The AI agent is in beta. We are accepting teams now.

SOC 2 Type II · U.S. hosted · AI optional
Join the agent beta