AI Legal Document Review: Accuracy, Limits, and Professional Duties

AI legal document review is the use of machine learning and natural language processing to read a legal document, classify its clauses, extract the terms that carry risk, and compare them against a standard, so that a lawyer opens a flagged document rather than a blank one. What the category does well is narrower than its name suggests, and the gap matters most to law firms and in-house teams that spend their days on third-party paper. In the Vals Legal AI Report, legal AI products beat a control group of independent lawyers on document question answering, summarization, transcript analysis and data extraction, and lost to that control group on redlining. That split decides how AI legal document review works in practice, where contract review stops and eDiscovery review and general-purpose chatbots begin, and which confidentiality and court-verification duties attach once client documents go into a model. It also decides what to ask a vendor before a pilot, which platforms are built for which document, and how to roll a tool out without taking on the verification problem yourself.

[cta-1]

Key takeaways

AI legal document review is strongest at finding, extracting and summarizing what a document says and weakest at judging what it should say. Published benchmarks put legal AI products ahead of a lawyer control group on document question answering, summarization and data extraction, and behind that group on redlining. Contract review and eDiscovery review are separate jobs with separate tools, and a general-purpose chatbot does neither against your team’s own standard. ABA Formal Opinion 512 states that a lawyer should obtain the client’s informed consent before client confidences go into a generative AI tool, and that boilerplate engagement letter language does not satisfy that duty. A proposed Tenth Circuit rule would require a human who has independently verified every cited authority to review an AI-assisted filing before it is submitted.

How does AI review a legal document?

AI reviews a legal document in a sequence of steps, and knowing them tells you where the output can go wrong. The model parses the document into clauses, classifies each clause by type, extracts the operative terms inside it such as cap amounts, notice periods and governing law, compares those terms against a standard, and surfaces the deviations as flags. The standard is the step that varies most between products: it may be a generic model of market language, a set of benchmark provisions, or the position your own team has already agreed. A lawyer then reads the flags, accepts or rejects each one, and owns the resulting position.

The error concentrates in two of those steps, because clause classification and extraction degrade on documents that depart from a familiar shape: a bespoke instrument with renamed defined terms produces fewer and less reliable flags than a standard vendor agreement. The comparison step fails for a different reason, which is that the standard behind it is not the standard your team uses. Products that learn from the edits a lawyer accepts and rejects narrow that second gap over time, because each recorded decision becomes part of the position the next document is measured against. Neither step produces a legal conclusion. A flag is a prompt to look, and the lawyer who signs the document decides what it means.

How accurate is AI in legal document review?

Accuracy depends entirely on which part of the review you mean, and the published head-to-head says so plainly. The Vals Legal AI Report ran four legal AI products against a control group of independent lawyers across seven tasks. The products beat the lawyer baseline on document question answering, where the strongest scored 94.8% against the lawyers’ 70.1%, and the widest margin of all came on document summarization, at 77.2% against a 50.3% lawyer baseline. On redlining the lawyers won, scoring 79.7% against 65.0% and 53.6% for the two products measured on that task.

What the lawyer baseline measured

The lawyer baseline is what makes the report’s numbers usable. The lawyer control group was sourced through Cognia Law, an alternative legal service provider, and the lawyers received their assignments in standard format, did not know they were being benchmarked, and had two weeks to respond. Across the report’s remaining tasks, the best AI score against the lawyer baseline ran as follows.

1. Data extraction: 75.1% against 71.1%, to the products.
2. Transcript analysis: 77.8% against 53.7%, to the products.
3. Chronology generation: 80.2% against 80.2%, a tie.
4. EDGAR research: 55.2% against 70.1%, to the lawyers.

That is why an accuracy percentage on its own tells you nothing: a vendor figure is interpretable only with the task, the benchmark and the comparison group named beside it, and a score drawn from clause extraction says nothing about the same product’s redlining.

Contract review and eDiscovery document review are different jobs

Contract review and eDiscovery document review are both sold as AI document review, and they are different jobs. eDiscovery review sorts a large volume of electronically stored information, or ESI, to decide what is responsive and what is privileged, under a litigation schedule, and the tools that do it are built around culling, clustering and technology-assisted review, or TAR, in which a model trained on reviewer decisions ranks the remaining population. Contract review reads one agreement closely to decide whether its terms match the position your team is willing to accept, and the tools that do it are built around clause classification, playbook comparison and proposed language. A platform tuned for the first job will not propose a redline, and a platform tuned for the second will not cull a production set.

Which job produces which output

The practical test is what the output is for. An eDiscovery pass produces a coded set and a privilege log. A contract pass produces a marked-up document and a negotiating position. Transactional work has a third shape again, because due diligence document sets combine the volume of the first job with the clause-level questions of the second, which is why diligence platforms tend to sit apart from both. The volume changes what the reviewer is asked to do, since a diligence pass answers a question about a set of documents and a contract pass answers a question about one.

Reporting across a portfolio of signed agreements is a fourth job, closer to contract management than to review, and a tool that reads one document well does not necessarily hold thousands and report across the set. Getting this boundary wrong is an expensive scoping error, because a firm that buys a litigation review platform to speed up its non-disclosure agreement (NDA) turnaround has bought culling it does not need and lost the redlining it does.

Can ChatGPT review a legal document?

ChatGPT can review a legal document in the limited sense that it will summarize the document and raise plausible issues, and it cannot tell you whether a clause departs from the position your team accepts, because it has no access to that position. In the legal research edition of the Vals Legal AI Report, published in October 2025, every AI product tested scored within four points of the others on an average weighted basis, in a band from 74% to 78%, and within nine points of the 69% lawyer baseline, and ChatGPT was one of the four products in that band. Raw reading comprehension is not where the gap sits. What a general-purpose model lacks is the standard your team has already agreed to compare the document against, a traceable basis for each issue it raises, and a contractual commitment about what happens to the document after it is uploaded.

What a general-purpose model is missing

The three gaps a general-purpose model leaves carry different costs.

1. The missing standard produces output that is reasonable and generic, so it flags an uncapped indemnity but not the cap your team stopped accepting last quarter.
2. The missing basis means an issue cannot be checked without re-reading the clause, which removes most of the time saved.
3. The data terms are the gap that reaches the rules of professional conduct, because a consumer tier that reserves the right to train on uploaded content is a poor fit for client material.

There is a narrower use that holds up. A general-purpose model is a competent reader of a document you are allowed to share, which makes it useful for orientation on a public filing or a document with no confidentiality constraint, in the same way it is useful for first-pass legal research that a lawyer then verifies. It is a poor fit for the step where a deviation from a house position has to be caught.

Confidentiality and professional duties when AI reviews client documents

ABA Formal Opinion 512, issued July 29, 2024, maps the existing Model Rules of Professional Conduct onto generative AI rather than creating new ones, and the duty it draws most sharply is confidentiality under Model Rule 1.6. The opinion states that a lawyer should obtain the client’s informed consent before entering client confidences into a generative AI tool, and that boilerplate consent language in an engagement letter is not adequate for that purpose. It also treats technological competence under Model Rule 1.1 as a live obligation, which means understanding how a particular tool handles the documents it is given.

The three questions to answer before an upload

Read as a workflow requirement rather than as a rule, ABA Formal Opinion 512 produces three questions to answer before client material is uploaded: whether the vendor retains the document after processing, whether customer content is used to train models, and which subprocessors receive the data. That diligence sits inside competence rather than inside IT, which is how the ABA’s own practice guidance on AI document review frames it, and the informed-consent point is difficult to satisfy without the answers.

The opinion is guidance rather than a bright-line test, so the consent conversation it describes is shaped by the engagement. A standing client that has approved a defined set of tools for a defined category of work sits in a different position from a new matter involving a third party’s confidential information, and the second case is where a general instruction to use AI carefully does the least work.

Court rules requiring human verification of AI-generated filings

A federal appeals court has proposed writing human verification of AI-generated filings into its rules, and the court states the revisions respond to increased use of generative artificial intelligence by lawyers and pro se litigants alike. On September 18, 2026 the U.S. Court of Appeals for the Tenth Circuit published proposed revisions to Tenth Circuit Rule 46.5, whose new subsection (B)(4) would require that any document prepared with the assistance of generative artificial intelligence be reviewed before filing by a human who has independently verified the existence of all cited authorities, the accuracy of all quotations, citations and legal analysis, and the document’s compliance with all relevant rules and ethical obligations. The court invited comments from October 18, 2026, will post the final rules no later than December 1, 2026, and has set the revisions to take effect on January 1, 2027. The duty it describes attaches to the lawyer who signs the filing, not to the tool that drafted it.

What a verification duty means in practice

The Tenth Circuit’s revisions are proposed and out for comment, so the text is likely to move before it takes effect. The direction is the part worth planning around, because a verification duty expressed that way is satisfied by a record rather than by an assurance: someone has to have checked each authority, and the firm has to be able to say who. Transactional teams sit outside the Tenth Circuit’s filing rules, and the same expectation reaches them through the client rather than the court, in the form of outside counsel guidelines that ask who reviewed AI-assisted work product.

What to look for in an AI legal document review tool

An AI legal document review tool is judged on what it compares your document against, whether each flag carries a traceable basis, where the document goes after it is uploaded, where the review happens in your existing workflow, and what the vendor’s accuracy figure actually measured. A tool that compares a clause against a generic model of market language answers a different question from one that compares it against the position your team has already agreed.

1. What it compares against. Ask whether the standard is a generic market model, a vendor-supplied set of benchmark provisions, a clause library you can inspect, or your team’s own agreed positions, and ask who can change it. A standard you cannot read is a standard you cannot defend to a client.
2. Whether each flag carries a traceable basis. A flag that cites the provision it compared against and the position it departed from can be checked in seconds. A flag that offers a risk score cannot, and the verification burden lands back on the reviewer.
3. Where the document goes. The retention, training and subprocessor terms ABA Formal Opinion 512 makes relevant to a confidentiality analysis are contract terms, not features.
4. Where the review happens. A review that runs where the document is already being edited produces tracked changes a counterparty can read. A review that runs in a separate platform produces a report someone has to transcribe back into the document.
5. What the accuracy figure measured. Ask for the task, the benchmark and the comparison group. A number without those three is not a claim about the work you do.

Those five criteria are also the practical reason to run a pilot against a written review checklist: scoring the tool on the same provisions your team already checks by hand is how you see whether the flags match the judgments. A pilot without that written comparison measures speed and nothing else, and speed is the easiest thing to demonstrate in a sales conversation.

Across the wider legal AI tool landscape the surface the review happens on tends to decide fit, which is why a Word-native reviewer suits teams whose negotiation already happens in tracked changes, and suits litigation production work far less well.

[cta-2]

The tools lawyers use for legal document review

There is no single best AI software for legal document review, because the products lawyers use split by the document and the job. Transactional teams reviewing third-party contracts, litigation teams culling a production set, and research-heavy practices verifying authority are buying different products, and each entry below names which of those jobs the tool is built for.

Spellbook

Spellbook fits transactional teams whose review and negotiation already happen in Microsoft Word, because the Review feature surfaces risks and proposes redlines as native Word tracked changes, the Playbooks feature holds the positions a team has agreed so deviations are measured against that standard rather than a generic one, and the Compare to Market feature benchmarks a provision against a set drawn from 20M+ real contracts and 2,300+ benchmark standards.

Pros

1. Review happens inside the document being negotiated, so the output is tracked changes rather than a report.
2. Flags compare against the team’s own agreed positions through the Playbooks feature, not only against market language.
3. Zero data retention, SOC 2 Type II, HIPAA compliant and GDPR, with SSO and MFA available.
4. A free trial without a sales call, and no implementation project before the first review.

Cons

1. A commercial-contract specialist, not a litigation, eDiscovery or legal-research platform.
2. Built around Word and Google Docs, so a team that reviews in a browser-based repository is working against the grain.
3. Bespoke, heavily negotiated instruments give the playbook comparison less to match against than standardized third-party paper does.

LegalOn

LegalOn is a direct comparable for transactional contract review, covering intake through the life of an agreement and generating redlines in Word as well as in a browser application, which makes it a reasonable alternative for in-house teams that want intake routing in the same product.

Pros

1. Contract review, redlining and storage in one product, with intake arriving through email and Slack.
2. Redline output in Word, so the negotiation artifact stays in the document.
3. Positioned for in-house legal, procurement and sales as well as law firms.

Cons

1. The vendor’s current pages describe no eDiscovery, litigation review or legal research capability.
2. Intake and lifecycle features expand the footprint beyond review, which is more product than a pure review pilot needs.

Luminance

Luminance suits enterprise contract work that spans negotiation, portfolio discovery and workflow automation, and it is built for diligence and mergers and acquisitions (M&A) document sets, where volume and clause-level questions arrive together.

Pros

1. Negotiation support with precedent and playbooks, plus discovery across an existing contract portfolio.
2. Serves procurement, finance and compliance alongside legal, which helps where contracts are owned across functions.
3. Built for enterprise scale rather than for single-document review.

Cons

1. The vendor’s current pages describe no eDiscovery or litigation document review.
2. No Microsoft Word integration is described on the vendor’s current product pages, so review sits in the platform.
3. Enterprise positioning means a small transactional team may be buying more platform than the job requires.

CoCounsel Legal

CoCounsel Legal, from Thomson Reuters, covers research, analysis and drafting in one interface and grounds its output in Westlaw content and Practical Law guidance, which fits practices where document review and authority checking are the same task.

Pros

1. Output grounded in named legal authority rather than in a general model alone.
2. Microsoft Word, Outlook and Teams integrations, with drafting and review available in Word.
3. Tabular analysis across large document sets, plus a separate litigation document analyzer.

Cons

1. Research-led rather than playbook-led, so comparison against a team’s own agreed contract positions is not the organizing idea.
2. The vendor’s current product page describes no eDiscovery capability.
3. Value depends on holding the underlying Westlaw and Practical Law subscriptions.

Lexis+ with Protégé

Lexis+ with Protégé, from LexisNexis and announced as the next evolution of the product formerly sold as Lexis+ AI, pairs research with document analysis in a Vault that supports up to 100,000 documents per matter and links every output back to the exact passage it came from.

Pros

1. Outputs link back to source passages, spreadsheet rows and timestamps, which makes verification quick.
2. Analysis across mixed matter material, including PDFs, spreadsheets, images, audio and video.
3. Drafting in Word with linked LexisNexis sources.

Cons

1. Research and matter analysis lead the product, so contract redlining against a house playbook is a secondary use.
2. Announced in May 2026 as the successor to Lexis+ AI, so older documentation and reviews refer to the product under the former name.
3. The vendor’s pages describe no eDiscovery production workflow.

Harvey

Harvey is a firm-wide platform rather than a review tool, combining a document store, legal research, collaborative workspaces and contract analysis, and it fits firms consolidating matter work onto one system.

Pros

1. One platform across research, document analysis and contract work, run at the matter level.
2. Adopted at scale in large firms, with the vendor citing 3,500+ legal organizations.
3. Agent-style workflows for multi-step work rather than single-document passes.

Cons

1. No Microsoft Word integration is described on the vendor’s current site, so redlining in the document is not the default surface.
2. The vendor’s current pages describe no eDiscovery capability.
3. Platform scope and firm-wide rollout make it a heavier commitment than a review pilot.

AI legal document review platforms compared

The platforms above are grouped by the document each one is built to handle and the surface the review happens on, because those two facts predict fit better than a feature list does. Accuracy figures appear nowhere in the table, because they are only interpretable where a published benchmark names its scope, and no vendor publishes a comparable one.

How to roll out AI document review in a law firm

Rolling out AI document review is a governance exercise before it is a software exercise, and the decisions that carry the most risk are who verifies the output, which matters the tool is allowed to touch, and what the firm tells clients. Name the reviewer for each document type rather than relying on a general instruction to check the work.

1. Name the verifier per document type. Decide that a partner reviews flags on anything above a stated contract value and an associate reviews the rest, and record it. A verification duty of the kind the Tenth Circuit has proposed is met by a record of who checked, not by a policy that someone should.
2. Fence the matters. Start with high-volume standardized paper where the same provisions recur, and keep bespoke or politically sensitive instruments out of the pilot until the flags have been scored against manual review.
3. Settle the client conversation before the first upload. The informed-consent point in ABA Formal Opinion 512 is a conversation with the client and an update to outside counsel guidelines, and it is easier to have once for a defined tool and matter category than per matter.
4. Train on the standard, not on the prompt. The value sits in the positions the tool compares against, so the work of a rollout is writing down the positions the team already applies.
5. Score the pilot against manual review. Run both passes on the same set for a fixed period and compare what each caught. That is the measurement that tells you whether the tool fits your paper.

The groundwork in step four is what makes the rest compound, because the written positions that drive a review are the same asset document automation draws on when the team moves from reviewing third-party paper to generating its own. Teams that skip it tend to re-run the exercise a year later under a different budget line.

Sequencing matters more than scope here. Review is the right first project because its output is visible in a document a counterparty reads, which makes the pilot easy to judge, and broader workflow automation is easier to justify once that first measurement exists.

Where AI legal document review stands now

AI legal document review is now a measured capability rather than a claim, and in the published benchmarks the measurements point in one direction: across the products tested, the tools were better than a lawyer control group at finding, extracting and summarizing what a document says, and worse at deciding what it should say. ABA Formal Opinion 512 and the Tenth Circuit’s proposed rule both put a human reviewer at the end of the process.

Buying on that basis means asking what the tool compares against, what happens to the document, and where the review lands in the work your team already does. Try Spellbook free to see how the Review feature flags a third-party contract against your own positions in Word.

[cta-3]

Frequently asked questions

Can AI identify privileged information during legal document review?

AI can flag likely privileged material by pattern, and the privilege call itself stays with a lawyer. eDiscovery platforms score documents for likely privilege using sender, recipient, counsel names and language patterns, which narrows a production set substantially. The judgment about whether attorney-client privilege or work product protection actually applies to a given communication turns on facts the model does not have, so privilege logs and clawback decisions are reviewed by counsel.

How quickly can AI review a legal document?

AI can read and flag a standard commercial agreement in minutes rather than hours, and the time saved is review time, not turnaround time. A model parses, classifies and compares the whole document in a single pass, so the lawyer starts from a flagged draft instead of a cold read. The remaining work, deciding which flags matter and what to send back, still takes as long as the negotiation requires.

Can AI detect a missing clause or an altered definition in a contract?

AI can detect a missing clause and an altered definition, and it does so by comparison rather than by reading for meaning. A model checks the document against an expected set of provisions and reports what is absent, and it checks each defined term against its uses elsewhere in the document and reports where the two diverge. Both checks are only as complete as the expected set behind them, which is why the provenance of that set is worth asking about before a pilot.

Is it safe to upload legal documents to AI?

Whether it is safe to upload a legal document to an AI tool depends on what the vendor’s contract says happens to the document, not on the model behind it. ABA Formal Opinion 512 treats this as a confidentiality question under Model Rule 1.6 and calls for the client’s informed consent, which boilerplate engagement letter language does not supply.

Is there a free AI legal document review service?

Free AI document review tiers exist, and their data terms are usually what rules them out for client material. Free and consumer tiers commonly reserve the right to use uploaded content to improve models. A free trial of a paid legal product is a different arrangement, because the data terms of the paid service usually apply during the trial.

Which documents benefit most from AI review?

AI review pays off most on high-volume, standardized third-party paper where the same provisions recur: NDAs, vendor and procurement agreements, master services agreements with their statements of work, and commercial leases. These documents have a known shape, so a model can compare each provision against a position the team has already agreed. Bespoke, heavily negotiated instruments give the model less to match against, and the review stays mostly manual.

Can AI be used to manage legal documents beyond review?

Yes. The same extraction that supports review feeds obligation tracking, renewal and notice dates, and reporting across a set of agreements, which is contract management rather than review.