Is AI Contract Review Reliable? What the Evidence Says
Yes, for one job: spotting issues in standard contracts. AI review tools now match or beat junior lawyers at finding problems in routine agreements such as NDAs, and they do it in minutes instead of hours. But reliability has limits. The same tools miss issues, struggle to rewrite clauses, and cannot take responsibility for a bad call. For a UK small business, the honest answer is this: use AI review as a fast first pass, and keep a solicitor for anything that carries real risk.
This article looks at what the published studies actually found, where AI review works well, where it fails, and how to use it safely. It is general information, not legal advice.
The short answer
AI contract review is reliable at reading contracts and flagging issues. It is less reliable at deciding what to change, and it is not reliable as a replacement for a lawyer. The distinction matters, because most marketing claims blur the two jobs together.
Think of it this way. Finding a problem in a contract is a reading task. AI is good at reading. Deciding how to fix that problem, weighing the commercial risk, and negotiating the change is a judgement task. That is still a human job.
The evidence backs this split. In the independent benchmarks published so far, AI tools beat lawyers on reading tasks such as answering questions about a document, summarising it, and pulling out key terms. Lawyers beat the tools on redlining, the task of marking up a contract with proposed changes. So the short answer is: reliable reader, unreliable editor.
What “reliable” means for contract review
Reliable is a vague word, so it helps to break it into four parts. A reliable review should be accurate, complete, consistent, and accountable.
Accuracy means the issues it flags are real issues, not false alarms. Completeness means it finds all the issues that matter, not just the easy ones. Consistency means it performs the same way on the tenth contract as on the first, even late on a Friday. Accountability means somebody takes responsibility when something goes wrong.
AI scores well on consistency. A tool does not get tired, bored, or rushed, and it applies the same checklist to every document. It scores well on accuracy for standard clauses, and less well on completeness, because it can miss unusual wording. It scores zero on accountability. If an AI tool misses a liability clause and your business signs a bad deal, the tool cannot be sued, and the vendor’s terms will almost certainly exclude that liability. You carry the risk.
That is why the right question is not “is AI review reliable?” but “reliable enough for which contract, used in which way?” An NDA from a trusted supplier is a different decision from a ten year commercial lease.
What studies actually found
Four studies tell most of the story. Read them together and a clear picture emerges.
The LawGeex study, 2018
The most quoted number in this field comes from a 2018 study by LawGeex, a contract review company. Twenty experienced US lawyers and the company’s AI each reviewed five non-disclosure agreements and flagged the legal issues in them. The AI averaged 94% accuracy against 85% for the lawyers, and took 26 seconds against an average of 92 minutes for the lawyers.
The number is real, but the caveats matter. LawGeex ran the study on its own product. The test used only NDAs, which are among the most standardised contracts in business, so they suit pattern based review. The answer key was built partly from the best answers of all participants, including the AI itself. The study was never peer reviewed and has never been replicated. It is good evidence that AI reads standard contracts well, and weak evidence for anything beyond that.
“Better Call GPT”, 2024
A 2024 academic study compared six AI models with junior lawyers and outsourced reviewers on ten procurement contracts. The best model, GPT-4, scored an F-score of 0.871 at finding issues, slightly ahead of the junior lawyers and level with the outsourced reviewers. It took about 4.7 minutes per contract at a cost of about 25 cents, against roughly an hour and about 75 dollars for the human reviewers.
The study also found the models favoured precision over recall. When they flagged something, they were usually right, but they sometimes missed issues, occasionally by misreading specific wording. Ten contracts is a small sample, and the authors worked for a legal technology company, so the result is encouraging rather than final.
Vals Legal AI Report, 2025
The Vals Legal AI Report tested legal AI tools against a blind lawyer control group across seven tasks. The tools beat the lawyers on document questions (94.8 against 70.1), summarisation (77.2 against 50.3) and data extraction (75.1 against 71.1). The lawyers won on redlining, scoring 79.7 against the best tool’s 65, and on research in public filings.
This is the most useful study for a small business, because it splits the work into tasks. AI wins the reading tasks. Humans win the changing tasks.
ContractScrub, 2026
A 2026 benchmark called ContractScrub tested ten leading AI models on the final proofreading pass: catching misused defined terms, broken cross-references, wrong party names, and contradictions planted by experienced lawyers. Only one model caught 75% of the planted errors. On a 60 page agreement, a 75% catch rate means some errors survive.
The lesson is consistent across all four studies. AI review is fast and strong on standard material, and it still misses things. A tool that catches three quarters of errors is a useful safety net, not a substitute for a careful read.
Where AI review works well
AI review earns its keep in five situations that most small businesses meet regularly.
Standard, repetitive contracts. NDAs, supplier terms, freelancer agreements, and standard terms of business follow familiar patterns. AI tools are trained on thousands of similar documents, so they spot missing clauses and unusual wording quickly.
First-pass triage. When five supplier contracts arrive in one week, AI review can rank them by risk in minutes. You then spend your time, and your solicitor’s time, on the one that actually needs attention.
Extracting key terms. Pulling out renewal dates, payment terms, liability caps, and termination clauses from a pile of contracts is exactly the kind of structured reading AI does well. The Vals study put extraction accuracy above the lawyer baseline.
Checking against your own standards. If your business has a playbook, for example “we never accept unlimited liability” or “payment terms must be 30 days”, an AI tool can check every incoming contract against those rules and flag the exceptions.
Volume work. Reviewing dozens of similar agreements, such as onboarding paperwork for new clients or suppliers, is where the speed advantage matters most. A task that takes a person days takes a tool an afternoon.
These are the same strengths that make AI workflow automation useful across a small business: the tool handles the repetitive reading, and the human handles the decisions.
Where it fails (and the risks)
The failures matter more than the successes, because a missed clause can cost real money. Here are the main risks.
Missed issues. The studies agree that AI tools miss things. ContractScrub’s best model missed a quarter of planted errors. “Better Call GPT” found models missed issues by misreading wording. A missed auto-renewal clause or an uncapped liability clause is exactly the kind of error that only shows up later, when it is expensive.
Hallucination. AI tools can invent confident sounding but false content. Stanford researchers found that leading AI legal research tools hallucinated on 17% to 33% of test queries. That study tested legal research rather than contract review, but the warning carries over: never accept an AI summary or citation without checking it against the document itself.
Weak redlining. Finding a problem is not fixing it. In the Vals study, the best AI tool scored 65 on redlining against 79.7 for lawyers. Tools tend to insert standard text instead of making the nuanced change a contract needs, and they struggle to tell a redline apart from the original text on complex amendments.
No commercial judgement. A clause can be legally unusual but commercially fine, or legally standard but commercially dangerous for your business. AI tools cannot weigh your relationship with a supplier, your cash flow, or your appetite for risk. Only you can.
Confidentiality and data protection. Uploading a contract to an AI tool means sharing it with a third party. Some tools train on user data, some store documents on servers outside the UK, and some keep copies after you delete them. For contracts containing personal data, that raises UK GDPR questions. Check where your data goes before you upload anything sensitive, and read our plain English guide to what GDPR means for small businesses.
Nobody to blame. If a solicitor misses a clause, professional indemnity insurance and regulation give you a route to redress. If an AI tool misses one, you have a software licence with a liability cap. The Law Society of England and Wales has been clear that solicitors remain responsible for work they delegate to AI. For a business using AI directly, that responsibility sits with you.
How small businesses should use it safely
The safe way to use AI contract review is as a junior assistant, not as a decision maker. A simple workflow keeps the benefits and controls the risks.
1. Keep a human in the loop. Never sign a contract on the strength of an AI review alone. Use the tool to produce a list of issues, then a person reads the flagged clauses in the original document and decides what to do about them.
2. Sort contracts by risk first. Use AI freely on low risk, standard documents such as NDAs and routine supplier terms. Send anything high value, long term, unusual, or heavily negotiated to a solicitor. A rough rule: if the contract could cost you more than a solicitor’s fee to get wrong, get the solicitor.
3. Write a short playbook. List the terms your business will and will not accept: liability caps, payment terms, termination rights, governing law, and auto-renewal. Give that list to the tool and to the person checking its output. A tool checking against your own rules is far more useful than one checking against generic ones.
4. Check the data handling. Before uploading a contract, confirm the tool does not train on your documents, stores data in the UK or the European Economic Area where possible, and deletes your files on request. If a contract contains personal data, this is a GDPR requirement, not a nice extra.
5. Verify every citation. If the tool quotes a clause number or a legal rule, open the document and check it. Hallucinated references are a known failure, and a wrong clause number in a negotiation is embarrassing at best.
6. Keep records. Save the AI report, the version of the contract you reviewed, and a note of who checked it and when. If a dispute arises later, that record shows you took reasonable care.
This workflow fits naturally alongside other automation in a small business. If you already automate client paperwork, our guide to automating client onboarding shows how contracts and e-signatures fit into the same pipeline.
FAQs
Is AI contract review accurate?
For finding issues in standard contracts, yes, broadly. Published studies put AI accuracy at finding issues between about 75% and 94%, depending on the task and the contract type. Accuracy drops on unusual wording, complex amendments, and redlining. Treat the headline numbers with care: the famous 94% figure comes from a vendor’s own 2018 study of five NDAs.
Can AI contract review replace a lawyer?
No, not for anything that matters. AI tools beat lawyers on reading tasks but lose on redlining, negotiation, and judgement. A solicitor also carries professional insurance and regulatory accountability that no software tool offers. Use AI to prepare and triage, and use a lawyer to decide.
Is it safe to upload contracts to an AI tool?
It depends on the tool. Check whether it trains on your documents, where it stores data, and how it handles deletion. Avoid uploading contracts with sensitive personal data to free consumer tools with unclear terms. For anything confidential, use a business tool with clear UK GDPR compliant data handling.
How much does AI contract review cost?
Costs vary widely. Some tools charge a few pounds per document, others charge monthly subscriptions from around 20 pounds to several hundred pounds for business plans. In the 2024 “Better Call GPT” study, the best model reviewed a contract for about 25 cents in computing cost, against about 75 dollars for a human reviewer. The real cost to budget for is the human time spent checking the output.
Which contracts should small businesses never give to AI alone?
Property leases, loan agreements, shareholder agreements, employment contracts with restrictive covenants, and any contract with unlimited liability or a value that could seriously hurt the business. These need a solicitor’s eyes, every time.
Conclusion
AI contract review is reliable enough to be genuinely useful, and not reliable enough to be trusted alone. The studies tell a consistent story: AI reads standard contracts fast and well, beats lawyers on extraction and summarising, and still misses issues and struggles to rewrite clauses. The 94% headline from 2018 is real but narrow, and the 2026 benchmarks put the honest figure for a final proofread closer to 75%.
For a UK small business, the practical rule is simple. Let AI do the first pass on routine contracts, check its findings against the original document, and send anything high value or unusual to a solicitor. Used that way, AI review saves hours and catches things a rushed human might miss. Used as a replacement for legal advice, it is a risk your business does not need.
