Last reviewed on October 3, 2026.
Document review is usually the most expensive part of eDiscovery. Every major review platform – including Relativity, Everlaw and DISCO – now offers some form of AI-assisted review. The methods differ, and so do the ways you prove the results are reliable.
The three main approaches
| Approach | How it works | Typical use |
|---|---|---|
| TAR 1.0 (predictive coding) | Experts code a training set; the model learns and ranks or classifies the rest; training stops once the model is stable. | Large, fairly stable collections where a one-time training effort is worthwhile. |
| TAR 2.0 / continuous active learning (CAL) | The model re-ranks documents continuously as reviewers code, pushing likely relevant documents to the front of the queue. | Most modern reviews, including rolling productions. |
| Generative-AI review | A large language model reads each document against written review instructions and suggests a relevance or privilege call with an explanation. | First-pass review, privilege review, issue coding and summaries; still maturing. |
How courts have treated AI review
US courts accepted technology-assisted review well before generative AI. Da Silva Moore v. Publicis Groupe (S.D.N.Y. 2012) is widely cited as the first judicial approval of predictive coding, and in Rio Tinto v. Vale (S.D.N.Y. 2015) the court described TAR as "black letter law" where the parties agree to use it. In England and Wales, Pyrrho Investments v MWB Property (2016) approved predictive coding. Generative-AI review is newer; the same principles – transparency about the process, reasonable validation, cooperation on protocols – are generally expected to apply.
Validation: how you show it worked
- Recall – the share of all relevant documents you actually found. Usually estimated by sampling.
- Precision – the share of documents marked relevant that really are relevant.
- Elusion testing – sampling the documents you decided not to review or produce to estimate how many relevant ones were missed.
- Documentation – record the protocol, training decisions, prompts (for generative AI), sample sizes and results.
For generative-AI review, treat the review instructions (prompts) as part of the protocol. Test them on a coded sample, refine them, freeze them, and validate the output with sampling just as you would for TAR.
Choosing e-discovery software: what to consider
When choosing a platform for litigation or an investigation, consider all of the following:
- Which applications and data sources the parties use or have used (email, chat tools such as Slack or Teams, mobile devices, cloud storage), and whether the platform can process them.
- Whether a cloud-based or self-hosted (stand-alone) deployment is appropriate for the data and any regulatory limits.
- Who controls the data and where it is stored (data residency, cross-border transfer rules, chain of custody).
- Which AI review methods are available and how their results can be validated and explained.
- Total cost: hosting per GB per month, processing, user licences and AI usage charges.
For a full selection process see the eDiscovery platform selection guide and the Everlaw vs Relativity comparison.
What the main platforms call their AI review tools
- Relativity: active learning (CAL) and the aiR family of generative-AI tools (for example aiR for Review and aiR for Privilege).
- Everlaw: predictive coding and the Everlaw AI Assistant for review, summaries and drafting; automatic coding suggestions appear within review workflows.
- DISCO: AI-assisted tagging predictions and generative-AI review features (marketed under names such as Auto Review).
Product names change frequently; confirm current features and how they are priced in a demo.
Frequently asked questions
What is TAR in eDiscovery?
Technology-assisted review (TAR) uses machine learning to rank or classify documents for relevance based on reviewers' coding decisions, so that fewer documents need human review.
What is the difference between TAR 1.0 and TAR 2.0?
TAR 1.0 trains a model once on a training set and then applies it. TAR 2.0, or continuous active learning, keeps learning from every coding decision throughout the review.
Is generative-AI document review defensible?
It can be, if the process is documented, the review instructions are tested and frozen, and results are validated with statistical sampling such as recall and elusion testing. Case law specific to generative-AI review is still developing.
What must investigators consider when choosing e-discovery software?
All of the following: what applications are being used or have been used by the companies involved, whether to use a cloud-based or stand-alone deployment, and who controls the data and where it is stored.