TAR vs Generative AI Review: ROI for Large-Volume Matters
What technology assisted review is
Technology assisted review, also called predictive coding, learns from documents that reviewers have already coded and uses that to rank the rest by likely relevance. The team reviews the most likely relevant documents first and can stop when the returns fall off. It is well understood and has been accepted by courts for over a decade.
What generative AI review adds
Generative AI review uses large language models to classify, summarise, and draft relevance, privilege, or issue calls with a stated reason. It can work without a large set of pre-coded examples, can explain why it reached a call, and can summarise long documents for a reviewer. It is newer, and the right validation around it still matters.
Where the ROI actually comes from
The cost of review is human time at scale, as set out in the real cost of document review. Both approaches save money the same way, by reducing the number of documents a person has to read or by making each review faster. The return is not a feature of the model, it is the volume it removes from the most expensive step.
When to use which
Technology assisted review is a strong fit for large, well-defined sets where defensibility through an established method matters. Generative AI review fits matters that need nuance, summarisation, or a faster start on a smaller set, and where the model's reasoning helps a reviewer move quickly. In practice they are often combined, with one cutting the bulk and the other handling the harder calls.
The defensibility of both
The standard is the same for either: you must be able to explain the method, show the validation, and document the sampling and quality control. A defensible process produces its own record as it runs. That requirement does not change because the model changed, and it is the backbone of the eDiscovery workflow automation guide.
Frequently asked questions
What is the difference between TAR and generative AI review?
Technology assisted review ranks documents by relevance after learning from coded examples. Generative AI review uses large language models to classify, summarise, and explain calls, and can work without a large set of pre-coded examples.
Which is cheaper, TAR or generative AI review?
Both save money the same way, by reducing the documents a human reviews. The cheaper option depends on the matter, the data set, and how each reduces the volume that reaches the most expensive step.
Is generative AI review defensible?
It can be, under the same standard as any review method. You must be able to explain the approach, show validation and sampling, and document the quality control.