AI contract review
Also called AI contract analysis, AI redlining, legal AI review
AI contract review reads and flags clauses in minutes, but independent tests show real error rates. What the evidence says for small firms.
AI contract review is quick, but independent tests show it works as a first pass, not as a lawyer. In a 2025 benchmark, two tools beat lawyers at pulling data from legal documents, yet lawyers beat every tool at redlining, scoring 79.7% against 65.0%. Specialist legal research tools gave hallucinated answers on 17% to 33% of test queries. Courts in England and South Africa have referred lawyers to regulators over fake AI citations. Use it to triage, then check every flag.
At a glance
| Status | In pilots |
|---|---|
| What it does | Reads contracts, pulls out key terms (parties, dates, payment terms, renewal, liability caps), flags clauses that differ from your standard positions and suggests redlines. |
| Typical cost | From $59 per month billed annually for self-serve AI review; AI review checked by a lawyer from $999 per month billed annually (US dollar list prices, checked Oct 2026) |
| Data it needs | Contracts and templates as Word files or text-based PDFs, plus a written playbook of the terms you accept and reject. |
| Setup effort | Medium: the tools work out of the box, but useful flags depend on writing your own playbook. |
| Evidence | Moderate. One peer-reviewed hallucination study, one independent benchmark, one law-review study and court judgments; contract-specific accuracy data is mostly vendor-run. |
What is AI contract review?
AI contract review uses large language models to read a contract and report on it. Typical tasks are pulling out key terms (parties, term, renewal dates, payment terms, liability caps, termination rights), answering questions about a document, comparing it with your standard positions and suggesting edits as tracked changes.
For a 30-person business, the realistic uses are triaging supplier and customer contracts before they go to a lawyer, building a register of renewal and notice dates from contracts already signed, and checking a counterparty's redlines against your template. It is different from AI legal research, which answers questions about case law and statute. The two share the same underlying models, though, and the same weakness: confident answers that are wrong.
How does it work?
You upload a contract or open it in a word-processor add-in. The tool splits the text into sections, then prompts a language model to find and classify clauses, compare them with a playbook (a list of the positions you accept, reject or want to negotiate) and draft comments or redlines. Better tools pull the relevant clause into the prompt so the answer is tied to the document rather than to the model's general knowledge.
Pricing splits two ways. Self-serve tools charge a monthly subscription, offer a free tier and leave the judgement to you. Hybrid services run the AI first, then have a lawyer verify the output before it comes back, at a much higher monthly fee. The existence of the second model says something useful: someone still has to check.
What it needs to work
AI contract review only works when the tool can read clean contract text and compare it with positions you have written down. Without a playbook and a register of what you have signed, it produces generic comments.
- Signed contracts saved as Word files or text-based PDFs (not phone photos or image-only scans) in one shared folder or document library
- A one- to two-page playbook: your accepted, fallback and rejected positions on payment terms, liability caps, indemnities, termination, renewal and governing law
- Your current templates (terms of trade, supplier agreement, NDA) as the baseline for comparison
- A contract register in a spreadsheet or CRM with counterparty, start date, renewal date and notice period, so extracted dates have somewhere to go
- A named person who checks every flag, and every cited case or statute, before anything is signed or filed
- A written agreement with the tool's vendor covering storage, security and breach reporting for uploaded documents
Most use cases in this guide succeed or fail here, on the data underneath them, before any AI is involved.
What does the evidence show?
Independent evidence is narrower than the marketing. The Vals Legal AI Report (2025) tested commercial legal tools, whose vendors chose to take part, against a lawyer baseline; Vals discloses a customer relationship with some participants. Two tools beat lawyers at data extraction (75.1% and 73.2% against 71.1%) while two fell short, and on redlining lawyers scored 79.7% against 65.0% for the best tool. Tools struggled with multi-clause extraction and with questions spanning several documents.
A peer-reviewed Stanford study of 202 legal queries found leading legal research tools hallucinated on roughly 17% to 33% of them. A study of law students doing realistic legal tasks found GPT-4 brought large, consistent speed gains but only slight, inconsistent quality gains.
One vendor-run study (researchers at legal software company Onit, ten procurement contracts) reports GPT-4 matched junior lawyers at spotting issues, in under 5 minutes against 56. We found no independent study of AI contract review in small businesses.
AI first-pass review vs lawyer review
| Compared on | AI first-pass review | Lawyer review |
|---|---|---|
| Speed per contract | Minutes; one vendor study reports under 5 minutes per contract | About an hour; the same study reports 56 minutes for junior lawyers and 43 for senior lawyers |
| Error profile | Confident errors, including hallucinated authority; weakest on redlined documents | Slower and dearer, but still ahead of AI tools on redlining in one benchmark |
| Cost model | Monthly subscription, from $59 per month on one vendor's list price | Hourly or fixed fee per matter |
| Best for | Triage, term extraction, renewal registers and first comments on standard paper | High-value, unusual or disputed contracts, and anything you are about to sign |
Status and evidence
Rated pilots: the tools are cheap and sold off the shelf, but independent tests show mixed accuracy, with lawyers still ahead on redlining and specialist tools hallucinating on 17% to 33% of research queries. Evidence is moderate: a peer-reviewed study, an independent benchmark and a law-review study, plus court judgments on fabricated citations. None measures results in small businesses.
Where does it go wrong?
- Invented or misread authority. The High Court of England and Wales found that 18 of 45 citations in one case did not exist, said freely available AI tools cannot do reliable legal research, and referred lawyers to their regulators. In Pietermaritzburg, only two of nine cases cited could be found and the judgment went to the Legal Practice Council. One public tracker listed 2,225 court decisions involving AI hallucinations by late October 2026, six of them in South Africa.
- Redlined and formatted documents. Vals found tools did best on clean plain-text clauses and struggled with formatted redlines.
- Deals spread across documents. A master agreement plus schedules and side letters is exactly the multi-document question tools handled worst.
- No playbook. Without your own positions written down, the tool compares your contract with a generic norm that may not suit your business.
- Confidential uploads. Contracts carry personal information and pricing; check where the vendor stores them before you upload.
Numbers it should move
Ask the demo business
Northgate Supply is a fictional wholesale distributor of catering and packaging supplies: 28 staff, R46.6m in sales over the last year. Its data is generated for this demo, to 30 September 2026.
Customers took about 42 days to pay at the end of September, up from 32 at the end of January. The biggest slowdowns: Hout Bay Coffee (30 to 60 days), Wheatfield Kitchen (37 to 62 days), Vineyard Catering (21 to 46 days).
How this was worked out
Money owed by customers at month end, divided by the last 91 days of sales, multiplied by 91 (fewer days in the first months, while history builds up).
Where the data usually lives
Where the records this use case needs are usually kept in a small business. It is not a list of recommended tools.
Frequently asked questions
How much does AI contract review cost for a small business?
Self-serve tools start at $59 per month billed annually on one vendor's list price, with a free tier that locks exports. Services that add a lawyer's check to the AI review start at $999 per month for 72 reviews a year on another vendor's list price. Both checked October 2026. Add the hidden cost: someone senior still has to read every flag.
Can AI review a contract instead of a lawyer?
Not on current evidence. In a third-party 2025 benchmark, lawyers beat every AI tool at redlining, scoring 79.7% against 65.0% for the best tool. Courts also treat unchecked AI output as the user's responsibility: the High Court of England and Wales referred lawyers to their regulators over fake citations. Use AI to prepare for a lawyer, or to review low-value standard contracts you would otherwise sign unread.
Do I need a developer or IT team to use AI contract review?
No. Self-serve tools are sold as monthly subscriptions with a free tier to try, so there is no integration project to start. The work that matters is not technical: writing a playbook of your standard positions and deciding who checks the output. If you want extracted renewal dates to flow into a CRM automatically, that part may need some setup help.
Is it safe to upload contracts to an AI tool under POPIA and UK GDPR?
It can be, if the vendor contract is right. Contracts usually contain personal information such as names, contact details and signatures. Under POPIA, a business that has a vendor process personal information for it needs a written contract requiring the vendor to keep it secure. The vendor must also tell the business immediately if it suspects unauthorised access. UK GDPR sets similar processor contract terms, which the ICO explains. Check the vendor's terms and whether your documents train its models. This is not legal advice.
What happens if AI invents a case or clause in a legal document?
The person who relies on it carries the consequences. In January 2025 a Pietermaritzburg High Court judge found only two of nine cited cases could be found and sent the judgment to the Legal Practice Council. A South African court also referred another matter, Northbound Processing, to the Legal Practice Council for investigation. One public database counted 2,225 such decisions worldwide by late October 2026.
More in risk, compliance & security
Sources
- 1Hallucination-Free? Assessing the Reliability of Leading AI Legal Research ToolsJournal of Empirical Legal Studies (Magesh, Surani, Dahl, Suzgun, Manning, Ho; Stanford), 2025Study
- 2Vals Legal AI Report (VLAIR)Vals AI, 2025Survey
- 3Lawyering in the Age of Artificial IntelligenceMinnesota Law Review (Choi, Monahan, Schwarcz), 2024Study
- 4Better Call GPT, Comparing Large Language Models Against LawyersOnit Inc. AI Center of Excellence (arXiv preprint), 2024Vendor
- 5Frederick Ayinde, R (on the application of) v The London Borough of Haringey; Al-Haroun v Qatar National Bank QPSC [2025] EWHC 1383 (Admin)High Court of England and Wales, King's Bench Division (Divisional Court), via The National Archives Find Case Law, 2025Government
- 6Judge finds lawyers may have used AI to source bogus citationsBusinessDay, 2025News
- 7Ghost authorities: AI hallucinations in the courtroomENSafrica, 2026News
- 8AI Hallucination Cases DatabaseDamien Charlotin, 2026Survey
- 9GenieAI PricingGenie AI, 2026Vendor
- 10Pricing - SuperlegalSuperlegal, 2026Vendor
- 11Protection of Personal Information Act 4 of 2013, section 21: Security measures regarding information processed by operatorActs Online (South African statute), 2013Government
- 12What needs to be included in the contract?Information Commissioner's Office (ICO), 2026Government
This entry is educational. It is not legal, financial or professional advice, and it does not recommend any vendor. Figures attributed to vendors are their own claims.
- 1Hallucination-Free? Assessing the Reliability of Leading AI Legal Research ToolsJournal of Empirical Legal Studies (Magesh, Surani, Dahl, Suzgun, Manning, Ho; Stanford), 2025Study
- 2Vals Legal AI Report (VLAIR)Vals AI, 2025Survey
- 3Lawyering in the Age of Artificial IntelligenceMinnesota Law Review (Choi, Monahan, Schwarcz), 2024Study
- 4Better Call GPT, Comparing Large Language Models Against LawyersOnit Inc. AI Center of Excellence (arXiv preprint), 2024Vendor
- 5Frederick Ayinde, R (on the application of) v The London Borough of Haringey; Al-Haroun v Qatar National Bank QPSC [2025] EWHC 1383 (Admin)High Court of England and Wales, King's Bench Division (Divisional Court), via The National Archives Find Case Law, 2025Government
- 6Judge finds lawyers may have used AI to source bogus citationsBusinessDay, 2025News
- 7Ghost authorities: AI hallucinations in the courtroomENSafrica, 2026News
- 8AI Hallucination Cases DatabaseDamien Charlotin, 2026Survey
- 9GenieAI PricingGenie AI, 2026Vendor
- 10Pricing - SuperlegalSuperlegal, 2026Vendor
- 11Protection of Personal Information Act 4 of 2013, section 21: Security measures regarding information processed by operatorActs Online (South African statute), 2013Government
- 12What needs to be included in the contract?Information Commissioner's Office (ICO), 2026Government