AI Document Review: How It Works, What It Costs and the Tools Firms Use

Last updated: 7 November 2026. Case law dates are from the judgments and contemporary law firm reports. Product details and prices are from vendor pages and announcements as read on 7 November 2026, the latest from July 2026. Accuracy figures are mostly vendor-reported, and we say so where they are.

Short answer: AI document review uses software to decide which documents in a disclosure or investigation are relevant or privileged, so lawyers read fewer of them. Older predictive coding learns from lawyers’ coding decisions; generative AI reads each document against a written review protocol and explains its call. Courts have accepted predictive coding since 2012. Prices are mostly quoted per matter, not published.

Key takeaways

• Technology-assisted review (TAR) has had court approval in the US since Da Silva Moore in 2012 and in England since Pyrrho in 2016. By 2015 a federal judge called its use “black letter law”.
• Generative AI review is newer. Vendors report recall and precision figures that match or beat first-pass human review, but most of the evidence comes from the vendors themselves.
• Few vendors publish AI review prices. Relativity folded aiR for Review into its standard RelativityOne rate, DISCO charges per document on a quote, and GoldFynch publishes per-GB hosting prices.
• ABA Formal Opinion 512 applies the usual duties of competence, confidentiality, supervision and reasonable fees to these tools. The signing lawyer stays responsible for the production.

What is AI document review?

Document review is the stage of litigation, regulatory work or an investigation where lawyers decide which collected documents are relevant, responsive to requests, privileged or sensitive. On large matters that can mean millions of emails and files. AI document review uses software to rank, classify or summarise those documents so that people read a smaller, better targeted set.

There are two main families:

• Technology-assisted review (TAR), also called predictive coding. Senior reviewers code a sample. The system learns from those decisions and scores or ranks the rest. Continuous active learning keeps retraining as reviewers code more.
• Generative AI review. A large language model reads each document against instructions written in plain language, such as the review protocol, and suggests a code with a short explanation.

Both sit inside wider e-discovery work. For how generative AI changed that wider workflow, see our analysis of eDiscovery and AI. Before any review starts, the documents must be preserved, which our guide to the legal hold covers.

How did courts come to accept predictive coding?

Through a short run of decisions between 2012 and 2016.

• Da Silva Moore v. Publicis Groupe (S.D.N.Y., 24 February 2012). Magistrate Judge Andrew Peck approved computer-assisted review in a case with over three million documents. District Judge Andrew Carter upheld the order on 25 April 2012, writing that “there simply is no review tool that guarantees perfection.”
• Irish Bank Resolution Corporation v Quinn (Ireland, 2015). The Irish High Court endorsed predictive coding even though the parties had not agreed to it.
• Rio Tinto v. Vale (S.D.N.Y., 2 March 2015). Judge Peck wrote that “it is now black letter law that where the producing party wants to utilize TAR for document review, courts will permit it.” He left open how far parties must share the documents used to train the system.
• Pyrrho Investments v MWB Property (High Court, 16 February 2016). In the Pyrrho judgment, Master Matthews approved predictive coding in England for the first time, so far as he was aware. About 3.1 million documents remained after de-duplication, manual review would have cost several million pounds, and the parties had agreed the method.

The reasons in Pyrrho remain a useful checklist: no evidence that the software is less accurate than manual review, more consistency, a very large document set, proportionate cost, enough time before trial and agreement between the parties.

How is generative AI review different from TAR?

The practical difference is the explanation. A TAR score tells a reviewer that a document is likely relevant. A generative tool says why, which makes quality control faster and makes errors easier to spot. The cost is that each document passes through a large model, which is why vendors often price generative review per document.

The point of validation is the same for both. You test a sample to estimate recall, the share of relevant documents the process found, and precision, the share of documents marked relevant that really are.

How accurate is AI document review?

The best known study of TAR is Grossman and Cormack’s 2011 article in the Richmond Journal of Law and Technology. Using data from the TREC 2009 Legal Track, it found that two teams using TAR achieved higher recall, precision and F1 scores than the official human assessors could have achieved by reviewing every document. That finding is a large part of why judges accepted TAR.

For generative AI, published evidence is thinner and mostly comes from vendors:

• Everlaw (September 2024). Testing its Coding Suggestions feature on four real litigation datasets of 7,737 documents in total, Everlaw reported precision of 0.67 and recall of 0.89. On one dataset with a direct comparison, recall was 0.82 against 0.6 for first-level reviewers. Everlaw called the results preliminary.
• DISCO (June 2025). DISCO said Auto Review exceeds 90% precision and recall in many cases, against an industry standard of 75% for human review, and that one client, Hueston Hennigan, saw 95% recall and 91% precision on a matter.
• Lighthouse (March 2025). Lighthouse said its responsive review reduces document volumes by up to 40% and that its privilege model is 60% more accurate than keyword-based models.

Read these as claims, not benchmarks. They use different data, different measures and no independent auditor. The fair conclusion is that well-run generative review can match first-pass human review on recall, but you still need your own validation on your own matter.

What does AI document review cost?

Few vendors publish prices for AI review. Here is what we could verify:

• Included in the platform rate. Relativity announced at Relativity Fest in October 2025 that aiR for Review and aiR for Privilege would become part of RelativityOne’s standard pricing. On 1 July 2026 it added aiR for Case Strategy to the standard rate. The RelativityOne rate itself is not published.
• Per document, on a quote. DISCO’s Auto Review pricing page describes a per-document charge that includes tag suggestions and explanations, but asks buyers to contact sales for the rate.
• Free for a period. Reveal launched its aji generative review engine in 2025 with full access at no cost until 31 December 2025, LawSites reported. We could not find its published price since then.
• Per GB, published. At the small-firm end, GoldFynch lists $27 a month for a 3 GB case and $76 a month for 10 GB, with cases above 300 GB at $275 per 50 GB. Its page lists machine learning assisted searches, not generative AI review.

In practice, ask for the total cost of a matter: processing, hosting per GB per month, the AI charge per document or per GB, user licences, project management and the human review that remains. Then compare that with the cost of a linear review at your reviewers’ rates.

Which AI document review tools do firms use?

The table covers the tools named most often for review. Facts are from vendor pages and announcements; our e-discovery software directory lists more options.

For more on Harvey, see our Harvey AI review. If you are choosing between tools, a structured trial on a closed matter with known answers tells you more than any demonstration; our guide to running a legal AI pilot sets out how.

What do the ethics rules say about AI document review?

In the US, the main guidance is ABA Formal Opinion 512, issued on 29 July 2024. It does not create new rules. It applies existing ones:

• Competence (Rule 1.1). Lawyers must understand the benefits and risks of the tools they use well enough to rely on them sensibly.
• Confidentiality (Rule 1.6) and communication (Rule 1.4). Lawyers must protect client information, and the opinion says boilerplate consent in an engagement letter is not enough where informed consent is needed.
• Supervision (Rules 5.1 and 5.3). Firms should set clear policies on AI use and make sure staff and outside vendors are trained.
• Fees (Rule 1.5). Lawyers billing by the hour may charge only for time spent, generally cannot charge clients for learning a tool, and may bill only the direct cost of using it.

On the court side, a US lawyer who signs a discovery response certifies under Federal Rule 26(g) that it was made after a reasonable inquiry. Using software does not move that duty to the vendor. Our guide to AI-assisted work and the billable hour covers how to record and bill the time.

How do you run a defensible AI review?

1. Agree the method early. Pyrrho shows how much weight courts give to agreement between the parties. Raise the use of AI review at the first discovery conference.
2. Write the protocol first. Clear definitions of relevance, issues and privilege help human reviewers and generative tools alike.
3. Test on a sample with known answers. Measure recall and precision before running the full set.
4. Keep people on privilege and close calls. Have lawyers check privilege calls and a sample of the AI’s decisions.
5. Validate at the end. Sample the documents the process marked not relevant to estimate what was missed, and record the result.
6. Document everything. Keep the protocol versions, prompts, sample results and decisions, so you can explain the process if challenged.

Frequently asked questions

Is AI document review accepted by courts?

Predictive coding is. US courts have approved it since Da Silva Moore in 2012, and the English High Court did so in Pyrrho in 2016. We found no published ruling dealing directly with generative AI review, so validation and agreement with the other side matter more.

What is the difference between TAR and generative AI review?

TAR learns from documents that lawyers have coded and then ranks the rest. Generative AI review reads each document against written instructions and suggests a code with a reason. Both need validation by sampling to show how many relevant documents were found.

How much does AI document review cost?

Most vendors quote per matter. Relativity includes aiR for Review in its standard RelativityOne rate, DISCO charges per document by quote, and GoldFynch publishes hosting prices from $27 a month for a 3 GB case. Always compare the total matter cost with linear review.

Is AI document review more accurate than human review?

Research on TAR found it can beat exhaustive human review. For generative AI, vendor studies report recall of about 0.8 to 0.95, often above first-level reviewers, but the evidence is vendor-reported. Test it on your own documents before relying on it.

Can lawyers bill for AI document review?

Yes, within the fee rules. ABA Formal Opinion 512 says hourly billers may charge only for time actually spent, generally may not charge for learning the tool, and may pass on only the direct cost of using it unless the engagement terms say otherwise.

Does AI document review replace contract reviewers?

It reduces the number of documents people read, but does not remove people. Lawyers still write the protocol, review privilege and close calls, check samples and sign the production. The work shifts towards supervision and quality control.

This article is general information about legal technology as of 7 November 2026. It is not legal advice or a product endorsement. Vendor features and prices change, so check current terms with each vendor.

Featured image: Photo: Kampus Production via Pexels (cropped)