Can AI Make Contract Review Faster?
Finding the clauses in a contract is slow, repetitive work, and AI tools now do it in seconds. One industry benchmark, in which vendors opted in, put the best tool slightly ahead of a lawyer baseline on extraction. Applying a standard to a clause is different: lawyers outscored the tools there, and no study we found has measured how much time is saved once the checking is counted.
On this page
Better: Note and confidence
Low confidenceClose to lawyers on extraction; behind them on judgementIn one industry benchmark of US legal AI tools, in which vendors opted in, the best tool slightly outscored a lawyer baseline on extracting terms from agreements, but lawyers clearly outscored the two tools that entered redlining against a standard. Grading was mostly automated, and the vendors chose which tasks to enter.Faster: Note and confidence
Low to moderate confidenceMachine speed on extraction; net time saved is unmeasuredAcross its seven tasks, the same benchmark found the tools six to 80 times faster than lawyers, whose times were self-reported. A vendor study from 2018 reported 26 seconds against an average of 92 minutes for lawyers. We found no independent study measuring time saved once a lawyer has checked the output.Cheaper: Note and confidence
Low confidencePlausible, but nobody we found has measured itThe benchmarks measure accuracy and speed, not cost. In an ACC and Everlaw survey, 12% of in-house legal respondents said they track technology return on investment.Safer: Note and confidence
Governance-gatedSafe when a responsible lawyer owns every risk callAustralian conduct rules require competent work and reasonable supervision, and the Law Society of South Australia says the supervising practitioner stays responsible for AI-assisted work. The UK regulator warns against trusting a tool to judge its own accuracy.The problem, in numbers
Australian solicitors carry a duty that pulls two ways: they must deliver legal services "competently, diligently and as promptly as reasonably possible".Source 1 Contract review tests both halves. Much of a first pass is locating clauses, such as a change of control or a limit on liability, before anyone decides whether they matter.
That work is slow because it is careful. When researchers built CUAD, a public dataset of 510 commercial contracts labelled for 41 clause types, their law student annotators first completed 70 to 100 hours of training, and the authors estimate each of the 9,283 pages was reviewed at least four times, at 5 to 10 minutes a page.Source 4 That describes their labelling method, not a typical law firm, but it shows how much expert time clause-finding absorbs.
The tools are arriving fast. England's solicitor regulator reported that by the end of 2022 three quarters of the largest firms were using AI, and that a growing range of products could help automate the review and management of contracts.Source 7 Measurement has not kept pace. In a 2025 survey of 284 chief legal officers, general counsel and legal operations staff, run by the Association of Corporate Counsel with Everlaw, a legal software vendor, 83% of respondents said they track outside counsel spend, but only 28% track time to resolution.Source 6
What improves, and why
In our reading, contract review splits into two jobs. The first is extraction: find every indemnity, termination right or most favoured nation clause and put it in a table. The second is judgement: decide whether a clause departs from the client's position, what that exposure is worth, and what to ask for instead. AI suits the first job. Extraction can be specified in advance: CUAD's 41 clause types are a fixed list that a model's answers can be checked against.Source 4 Judgement depends on the client, the deal and the counterparty, which no list captures.
The machine is plainly faster. In a 2018 vendor study, before generative AI, the contract review company LawGeex reported that its software spotted issues in five non-disclosure agreements in 26 seconds, against an average of 92 minutes for 20 US lawyers, and scored 94% accuracy to their 85%.Source 2 That is a vendor testing its own product, and its own report said the technology was not currently capable of being used as a standalone tool.Source 2
Finding the clause is the part AI does in seconds. Deciding what it costs the client is the part that still takes a lawyer.
What the evidence actually says
The most useful evidence we found is the Vals Legal AI Report of February 2025, an industry benchmark in which vendors opted in and chose which of the US law firms' tasks to enter. On data extraction, the best tool scored 75.1% against a lawyer baseline of 71.1%, while two of the four tools scored below the lawyers.Source 3 On redlining, which asks for a contract to be amended or checked against a standard, the lawyers scored 79.7% and beat both tools that entered, each tested through its general assistant.Source 3 Across all seven tasks, the report found the tools six to 80 times faster than the lawyers, whose times were self-reported.Source 3
In our reading, the extraction questions were the kind a first pass is made of: one asked for any text relating to most favoured nation clauses. The lawyers in the baseline were not told they were part of a benchmark; in our reading, that makes their scores closer to ordinary work than a contest would be.Source 3
Vals Legal AI Report, February 2025, an industry benchmark. US tasks, graded mainly by an AI evaluator; vendors opted in and chose their tasks. Best tool on redlining from the results table, which its task average supports; the report's text gives 59.4%.
In our reading, that pattern supports the angle of this review. The tools match lawyers on finding terms and fall behind on applying a standard to them. The benchmark has real limits: vendors chose which tasks to enter, an AI model did most of the grading, the number of lawyers is not stated, and the study covers the United States only.Source 3
Beyond that, the evidence is thin. A 2024 preprint from the legal technology company Onit compared large language models with junior lawyers and outsourced reviewers on ten procurement contracts, against answers set by senior lawyers.Source 5 It has not been peer reviewed, so we cite it only for its design. We found no peer-reviewed study that times a lawyer reviewing contracts with and without AI and counts the checking.
Where it falls short
Speed is measured on the machine, not the matter. Benchmarks time the tool's answer.Source 3 The lawyer still has to check it, and nobody we found has measured that combined time.
Judgement lags extraction. On redlining against a standard, lawyers outscored both tools tested.Source 3
The 2018 timing is old and vendor-run. The 26 seconds against 92 minutes comes from a contract review company testing its own pre-generative software in 2018.Source 2
Few teams can prove the saving. Only 12% of in-house respondents track technology return on investment.Source 6
Where governance decides the outcome
The risk call stays with a lawyer, and Australian rules say so in general terms. A solicitor with designated responsibility for a matter must exercise reasonable supervision over the solicitors and other employees working on it.Source 1 The Law Society of South Australia says practitioners remain personally responsible for their work, that the supervising practitioner stays responsible for AI-assisted work by juniors and support staff, and that output should be reviewed by a practitioner before it is used.Source 8 In England, the regulator says firms cannot delegate accountability to a provider, should not trust an AI system to judge its own accuracy, and need supervision systems that keep up with the speed of AI.Source 7 Data handling is its own question, covered in our review of client data.
How to think about adoption
Use AI where the task is extraction: building clause tables, checking a portfolio for a defined term, or triaging a data room. Keep the judgement steps, such as deviations from the client's position, risk ratings and negotiation advice, with a named lawyer. Before buying, run a pilot on your own contracts, time the whole job including the lawyer's check, and compare it with the same work done without the tool. Record which tool produced each table and which lawyer checked it, so the supervision the rules require can be shown later. In our reading, the first pass gets much faster; whether the matter does depends on how much of the output a lawyer can trust without redoing it.
BFCS.ai does not fabricate figures. The independent evidence on contract review speed is thin. The main benchmark is an industry study in which vendors opted in and chose their tasks, an AI model did most of the grading and lawyers reported their own times; the timing study most often quoted is a vendor's own test from 2018; and the two research papers cited are used only to describe their methods. We found no peer-reviewed study measuring how much lawyer time AI saves on contract review once the output has been checked, and no Australian study at all.
Sources
- 1
Law Council of Australia, "Australian Solicitors Conduct Rules," compilation dated 1 July 2026 (Legal Profession Uniform Law Australian Solicitors' Conduct Rules 2015, incorporating amendments to rules 8 and 13 applicable in NSW, Victoria and Western Australia). lawcouncil.au
Back to textWhat we took from it
Rule 4.1.3: a solicitor must "deliver legal services competently, diligently and as promptly as reasonably possible." Rule 37.1: a solicitor with designated responsibility for a matter must exercise reasonable supervision over solicitors and all other employees engaged in the provision of the legal services for that matter.
- 2
Tashea, J., "AI software is more accurate, faster than attorneys when assessing NDAs," ABA Journal, 26 February 2018, reporting a study by LawGeex, a contract review software company. abajournal.com
Back to textWhat we took from it
Vendor study (self-reported, not peer-reviewed). 20 US-trained attorneys and the LawGeex software reviewed five real non-disclosure agreements; the software scored 94% accuracy against 85% for the attorneys; the task was spotting issues; the attorneys took an average of 92 minutes for all five NDAs and the software 26 seconds. The report, as quoted, cautions that the technology is not meant to be, nor currently capable of being, used as a standalone tool. The study predates generative AI. We could not find or open the LawGeex report itself, so these figures are as reported by the ABA Journal.
- 3
Vals AI, "Vals Legal AI Report" (VLAIR), updated 27 February 2025, developed in partnership with Legaltech Hub; the February 2025 report covering contract and document tasks, not the separate Vals report on legal research of October 2025. vals.ai
Back to textWhat we took from it
Industry benchmark in which vendors opted in, not peer-reviewed. Tasks were proposed by participating law firms (including Reed Smith, Fisher Phillips, McDermott Will & Emery and Ogletree Deakins, plus four unnamed firms), with over 500 samples drawn mainly from Am Law 100 firms. Vendors (Thomson Reuters CoCounsel, vLex Vincent AI, Harvey Assistant and Vecflow Oliver) chose which tasks to enter; Vals AI states it has a customer relationship with one or more participants. A lawyer baseline was collected through Cognia Law from independent lawyers who were not told they were part of a benchmark; the number of lawyers is not stated, and they were asked to report the time taken to answer each question. Answers were graded pass or fail, check by check, by an AI evaluator model, with a second review of failing scores; the report says this method is not sufficient to replace an expert human grader. Data Extraction (30 questions, 29 documents, 204 checks): lawyer baseline 71.1%, Harvey Assistant 75.1%, CoCounsel 73.2%, Vincent AI 69.2%, Oliver 64.0%. Redlining (20 questions, 69 checks; suggesting amendments to match a standard, checking terms against a standard and explaining a redlined contract; only two vendors entered, and both were tested through their general assistant rather than a dedicated redlining feature): lawyer baseline 79.7%, Vincent AI 53.6%, Harvey Assistant 65.0% in the results table, while the report's text gives 59.4% for Harvey Assistant. The table's task average of 66.1% equals the mean of 79.7, 65.0 and 53.6 (with 59.4 it would be 64.2), and the extraction row's average is computed the same way, so we use the table figure (our calculation); the report itself has not corrected the inconsistency. "The AI tools were six times faster than the lawyers at the lowest end, and 80 times faster at the highest end" (across all seven tasks; tool times include file upload). The study covers the United States only.
- 4
Hendrycks, D., Burns, C., Chen, A. & Ball, S., "CUAD: An Expert-Annotated NLP Dataset for Legal Contract Review," NeurIPS 2021 Datasets and Benchmarks Track (conference paper; cited for its method only). arxiv.org
The Atticus Project, "CUAD" dataset page: atticusprojectai.org
Back to textWhat we took from it
510 commercial contracts labelled for 41 clause types considered important in contract review, with more than 13,000 labels; law student annotators completed 70 to 100 hours of contract review training, including video instructions by and live workshops with experienced lawyers, followed over 100 pages of annotation rules, and each annotation was verified by three additional annotators; in estimating the dataset's value, the authors state each of 9,283 pages was reviewed at least four times, at 5 to 10 minutes per page.
510 commercial legal contracts, 41 types of legal clauses, 13,000+ labels, labelled under the supervision of experienced lawyers.
- 5
Martin, L., Whitehouse, N., Yiu, S., Catterson, L. & Perera, R., "Better Call GPT, Comparing Large Language Models Against Lawyers," arXiv:2401.16212, 24 January 2024 (preprint, not peer-reviewed; cited for its method only). arxiv.org
Back to textWhat we took from it
Authors are from the AI Center of Excellence at Onit Inc., a legal technology company. Method: ten anonymised real-world procurement contracts from the United States and New Zealand; senior lawyers set the ground truth for whether each contract adhered to or deviated from a given standard; large language models were compared with junior lawyers and legal process outsourcers on determining and locating legal issues, time taken and cost. We do not rely on its reported results.
- 6
Everlaw (a legal software vendor) with the Association of Corporate Counsel, "81% of CLOs Say GenAI Accelerates Legal Work, Can Help Their Teams Demonstrate Value, But Lack the Metrics to Prove it ACC/Everlaw Survey Finds," press release, 17 November 2025, on the report "The Role of Generative AI in Proving Corporate Law Department Value." 284 chief legal officers, general counsel and legal operations professionals worldwide; 83% of respondents track outside counsel spend, 28% track time to resolution and 12% track technology return on investment. everlaw.com.au
Back to textWhat we took from it
Self-reported survey; question wording not published in the release.
- 7
Solicitors Regulation Authority (England and Wales), "Risk Outlook report: The use of artificial intelligence in the legal market," 20 November 2023. sra.org.uk
Back to textWhat we took from it
At the end of 2022 three quarters of the largest solicitors' firms were using AI; "A growing range of products are coming to the market which can help automate the review and management of contracts"; firms "remain responsible and accountable for the outputs from AI" and cannot delegate accountability to an IT team or external provider; "Do not trust an AI system to judge its own accuracy"; current AI "does not have a concept of truth"; firms should have supervision systems that can cope with the increased speed of AI.
- 8
Law Society of South Australia, AI Hub, "Frequently Asked Questions" (last updated 30 June 2026). lawsocietysa.asn.au
Back to textWhat we took from it
Practitioners "remain personally responsible for their work and conduct"; "The supervising practitioner remains responsible for ensuring that work is competent, accurate" and consistent with professional obligations, including AI-assisted work by junior lawyers, graduates and support staff; the page recommends clear practice controls on approved tools and reviewers; internal first drafts can be lower risk, but "the output should be reviewed by a practitioner before it is used."