Can AI Review Our Contracts? What It Catches and What It Misses

Can AI review our contracts?

Yes, AI can review your contracts for a first pass, and the honest version of that sentence has a boundary in it. A current model will read a 40 page master services agreement, list the clauses it contains, compare them against a checklist you give it, and tell you what is missing, in about a minute. That is real work, and it is work most businesses currently either pay for or skip.

What it will not do is own the answer. A contract review has two halves. The first is extraction: what does this document say, where is the indemnity clause, what is the notice period, does it auto renew. Machines are strong here. The second is judgment: is this term unusual for this kind of deal, what does it expose us to, what should we push back on, what happens if this goes wrong in our state. Machines are weak here, and they are weak in a way that reads as confident.

The distinction matters because vendors sell both halves under one word. "AI contract review" in a product name usually means extraction plus a comparison against a clause library. That is useful. It is not a legal opinion, and nobody involved is carrying insurance for it.

For most businesses the practical result is a change in who reads what. The AI reads everything. A person reads the flagged sections and the whole of anything that matters. Your outside counsel sees the contracts above a threshold you set rather than whichever ones happened to reach them. That is more review than most small businesses do today, at a lower cost than reviewing everything by hand, and it is a different claim from "we replaced the lawyer".

What can AI actually catch in a contract review?

AI catches six things in a contract review reliably, and they are the ones that eat the most time.

Missing clauses. Give it the list of terms you expect in a subcontractor agreement and it will tell you which ones are absent. Absence is the easiest thing for a machine to spot and one of the easiest for a tired person to miss at 6pm.

Deviations from your standard. If you have a position on limitation of liability, payment terms and termination for convenience, it will compare each contract against those positions and show you the gaps. This is the single highest value use for a business that signs the same shape of agreement repeatedly.

Dates and obligations. Renewal windows, notice periods, delivery milestones, insurance certificate deadlines. Pulling these into a calendar is mechanical, and it is where the money quietly goes. World Commerce and Contracting's 2025 contract management research puts the average loss at almost 9% of value a year through poor contract management, with the best performers at about 3% and the weakest at 15% or more.

Inconsistency inside the document. Defined terms used before they are defined, a payment schedule in section 4 that contradicts the one in exhibit B, a party name that changes halfway through.

Plain language summaries for people who are not going to read it. A site supervisor will read six bullet points about what the contract requires of them. They will not read the contract.

Comparison across a set. Thirty signed leases, one question: which of these allow a rent review this year. That query is minutes of machine work and a week of human work.

What does AI get wrong when it reads a contract?

What AI gets wrong when it reads a contract is that it states things which are not in the document, in the same tone it uses for the things that are. The independent evidence on this is better than the marketing on either side.

Researchers at Stanford RegLab and Stanford HAI tested the major commercial legal research tools against more than 200 manually constructed legal queries. Their paper, Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools, appeared as a preprint in May 2024 and was published in the Journal of Empirical Legal Studies in 2025. The tools hallucinated between 17% and 33% of the time: Lexis+ AI and Thomson Reuters' Ask Practical Law AI produced incorrect information more than 17% of the time, and Westlaw's AI-Assisted Research was the weakest of the set. These were purpose-built legal products, retrieval-grounded, sold to law firms, and they still only did better than general-purpose models rather than well in absolute terms. Thomson Reuters disputed the design, arguing that Ask Practical Law AI was tested outside its intended use, and the versions tested were the products as they stood in 2024.

That study also named the failure that matters most for contract work. A wrong answer is one thing. A misgrounded answer is another: the system states the law correctly and cites a source that does not support it. In a contract review this looks like a flag pointing at the wrong clause number, or a reassurance that your indemnity is capped when the cap sits in a different section and does not apply to you.

The second failure is silence. A model will tell you what is wrong with the clauses in front of it. It will rarely tell you that the contract is missing the entire schedule that defines the pricing, or that the counterparty's entity name does not match the company you negotiated with. Absence of a thing the document never mentioned is hard for a reader that only sees the document.

The third is confidence, and the courts have been collecting the evidence. A database maintained by researcher Damien Charlotin tracks decisions in which a court found, or clearly implied, that a party relied on AI-fabricated material, and trackers reading from it put it past 2,000 recorded decisions worldwide by September 2026. The count rises most weeks. Those are filings written by people whose job is to check citations. The general handling pattern, including what to sample and when to stop trusting a flag, is in what happens when AI gets it wrong.

Is AI contract review accurate enough to replace a lawyer?

No, AI contract review is not accurate enough to replace a lawyer, and the study usually cited to argue otherwise says something narrower than the headline.

The LawGeex study from February 2018 is the one everybody quotes. It put an automated system against 20 US-trained corporate lawyers on five previously unseen non-disclosure agreements containing 153 paragraphs. The system averaged 94% accuracy at spotting the risks, against 85% for the lawyers, and finished in 26 seconds against an average of 92 minutes. The best lawyer matched the machine at 94%. The worst scored 67%.

Read the conditions before you read the result. Five documents, all NDAs, all of one familiar shape, with a defined list of issues to find, scored by the company that sells the software. That is a strong result for one narrow task: issue spotting in a standard agreement against a known checklist. It is not a result about a bespoke supply contract, a dispute, or anything where the right answer depends on what you negotiated verbally in March.

Vendor accuracy claims are easy to find and hard to compare, because the method behind them is usually unpublished. LegalOn's Contract Review Benchmark 2026 is the better documented example: 11 models across 3,282 head-to-head reviews on 21 provision types, each pair judged blind, with LegalOn placing first on all 21 and scoring more than 400 points above the best GPT model tested. LegalOn sells legal-specific contract review software, so the finding and the seller point the same way. The useful part is the direction rather than the ranking. A general model doing contract review with no legal-specific grounding and no checklist is the weakest configuration available, and it is the one most businesses are using.

Review path	What it reliably catches	What it misses	Sensible use
Nobody reads it	Nothing	Everything	The default in more businesses than will admit it
General AI model, no checklist	Obvious clause types, plain language summary	Missing schedules, jurisdiction-specific effect, misgrounded flags	Triage only, never a sign-off
AI against your own clause standard	Deviations from your positions, missing clauses, dates	Judgment on whether a deviation is acceptable	First pass on routine agreements
Internal reviewer with AI first pass	Most of the above, plus commercial context	Specialist risk outside their experience	The working setup for most mid-sized businesses
Outside counsel	The judgment calls, jurisdiction, negotiation strategy	Nothing you are paying them to catch	Anything above your risk threshold

Which contracts should we never let AI sign off on alone?

Five categories of contract should never get a machine-only sign-off, and the test is exposure rather than length.

Anything where the downside is larger than the deal. An uncapped indemnity inside a small routine agreement is a small deal with an unbounded risk attached. Machines read the clause correctly and have no sense of proportion about it.

Anything bespoke. A contract drafted for this transaction has no clause library to compare against, which removes the thing AI is actually good at. The more unusual the deal, the less an automated review is worth.

Anything in an unfamiliar jurisdiction. Enforceability of non-competes, notice requirements and consumer protection rules vary by state and by country, and a model will answer in the voice of whichever body of law dominated its training.

Anything regulated or licensed. Healthcare, financial services, insurance and the licensed professions carry contract terms that exist because a regulator requires them. The adjacent operational line, where scheduling ends and advice begins, is covered on the law firms and accounting practices board and the dental and medical practices board.

Anything where you are the one drafting. Review and drafting are different risks. A generated clause that reads well and does not do what you think it does will sit in your template for years.

One more thing belongs here, which is the question of what leaves your building. A contract is the counterparty's confidential information as well as yours, and plenty of them contain a clause about disclosure to third parties. Before a contract goes into any tool, somebody has to know where the text lands, whether it trains a model, and whether your own client agreements permit it. Is AI automation safe with my business data covers how to answer that properly.

How do we set up contract review that holds up?

Setting up contract review that holds up takes four steps, and the first one is not about software at all.

Write down your positions. What payment terms do you accept, what liability cap, what notice period, what happens on termination. Most businesses have these in one person's head. Until they exist as a list, there is nothing for an automated review to compare against, and the output will be generic observations about contracts in general.

Decide the threshold. Contract value, term length, or category. Below the line, an AI first pass plus an internal reader. Above it, counsel. The threshold is a business decision and it is the part that actually controls your risk, not the software choice.

Make the output a document a person can act on, not a chat window. A review that produces a dated file, listing each flag, the clause it points to and the position it departs from, is reviewable later. A conversation is not. When somebody asks in eighteen months why you accepted that clause, the file answers and the chat history does not.

Measure the misses, not the hits. Every quarter, have your reviewer read ten contracts the system cleared and record what it did not flag. That number tells you where the threshold should sit. Teams that skip this step have no evidence about their own process beyond the absence of a disaster so far.

Where BLACKSIG builds this, the extraction, the clause comparison and the obligation calendar run on our infrastructure and we operate them from there, with the thresholds and the sign-off staying with your people. The strategy work decides whether contract review is even the right place to start, which for a business signing forty agreements a year it usually is not, and for one signing four hundred it usually is.

What does contract review cost us today?

Contract review costs you more than the invoice from your lawyer, because most of the cost sits in the contracts nobody reviewed.

The visible number is simple to get: hours billed on contract work last year, plus the loaded cost of internal time spent reading, chasing and filing. The invisible number is the one worth finding. Count the agreements that were signed without a full read. In most businesses under 200 people that is the majority of them, and nobody is embarrassed about it because the alternative was a week of waiting.

Then there is the money that leaks after signature, which is the WorldCC loss above. Missed deliverables, renewal windows nobody watched, obligations nobody tracked. That is also the part automation is best at, because tracking a date is not a judgment call.

Your own arithmetic beats any benchmark. Take your three most recent renewals. For each, ask whether anybody knew the notice window was closing, and whether the terms you are now on are the terms you would have negotiated. That exercise usually finds more money than the review cost, and it points at the obligation tracking rather than the review itself as the first thing to build. The general method for pricing this sort of work, including how to avoid counting time that moved rather than disappeared, is in how to measure whether AI automation is saving money.

Frequently asked questions

How much does it cost to automate contract review?

What moves the number is how many agreements you handle, whether they arrive as clean documents or scanned PDFs, whether your clause positions are written down or have to be worked out first, how many systems the output has to reach, and how much human review stays in the path. A business signing 40 similar subcontracts a year is a different job from one managing 2,000 leases across a dozen states. BLACKSIG does not publish a rate card; scope is set on the call once the strategy work has established what is worth doing. The drivers are in how much AI automation costs.

How does AI contract review actually work?

The document is converted to text, split into clauses, and each clause is classified by type. The system then compares what it found against a checklist of positions you supplied, flags deviations and absences, extracts dates and obligations into a structured record, and writes a summary. Everything it flags points at a specific clause so a person can check it. The quality of the output depends more on the checklist than on the model.

Is AI contract review better than hiring a paralegal?

Different jobs, and the better setup usually has both. A paralegal carries context, chases the counterparty, and knows which partner to ask. A system reads everything in minutes and never gets bored on the fortieth lease. The combination, where the machine does the first pass and the person reviews the flags and owns the file, gets through more work than either alone, which is the point of doing it.

Can our contracts be reviewed by AI if they are scanned PDFs or paper?

Yes, with a step in front. Scanned documents go through text recognition first, and the quality of that step sets a ceiling on everything after it. Clean digital PDFs are near perfect. Faxed and photocopied documents are not, and signature pages, handwritten amendments and marginal notes are where recognition fails most often. Budget for a sampling check on recognition accuracy before trusting any review built on top of it.

Will a court accept that we relied on an AI review?

No, and nobody should design a process on the assumption that it might. Responsibility for a signed contract sits with the business that signed it. Treat an automated review as an internal control that catches more than the status quo, document who reviewed what, and keep a person's name against every signature.

Can AI Review Our Contracts? What It Catches and What It Misses

The short version