Article
Generative AI is moving into document review fast. Privilege screening, issue coding and privilege log drafting are the most common early uses, and GenAI review is often priced per document at roughly $0.11 to $0.50, according to the Winter 2026 eDiscovery Pricing Survey from ComplexDiscovery and EDRM.
Commentators also point out that there is not yet settled authority on using generative AI for responsiveness calls. Until there is, the safest position is a process you can describe, measure and repeat. These five practices get you there.
1. Write the protocol before the prompt
Start from your review protocol: the issues, the privilege rules, the examples of hard calls. Turn that into model instructions, and keep the two in step. When the protocol changes, the prompt changes, and both get a new version number.
2. Version and log everything
For every document, record:
- the prompt version and model version used
- the model's call and its written reason
- any reviewer override and who made it
- the date and time
If someone asks in six months why a document was withheld, you can show exactly what happened.
3. Measure against human calls first
Before using AI review on a live matter, run it on a closed matter where human calls already exist. Report recall, precision and the overturn rate by issue. This side-by-side test is the best evidence you have that the process works for your kind of data.
4. Route uncertainty to people
Ask the model for a confidence level as well as a call. Send low-confidence documents to human reviewers first, and sample a share of high-confidence calls for quality control. Track elusion, meaning relevant documents in the set you did not review, the same way you would with technology-assisted review.
5. Keep attorneys in charge
The model proposes. Attorneys decide. Make that explicit in your workflow and in any ESI protocol or 502(d) discussions. Reviewer overrides are also your best training data for the next calibration round.
Watch for health information
Litigation involving hospitals, physician groups and payers often contains protected health information. Build PHI detection and redaction into processing, not as an afterthought at production, and handle it under the appropriate protective order.
Want to test GenAI review on your own data? Our eDiscovery AI pilot runs side by side with your original calls on a closed matter and ends with an accuracy report and a defensibility memo.
This article is general guidance, not legal advice.