AI in E-Discovery: Cost, Risk and Accuracy Set the Agenda
-
September 23, 2026
-
Ask any legal team that has piloted generative AI in e-discovery why a tool stayed in use, and the answer usually centres around three key factors: what it costs, what risk is involved and whether the result can be trusted as accurate. More than any feature list or model benchmark, these three considerations continue to determine whether AI is treated as a promising demonstration or a defensible part of the e-discovery workflow.
The Scalability Test
The most common failure point is accuracy at the volume most matters now demand. A tool demonstrated against a clean sample of 2,000 documents will behave very differently once asked to work across hundreds of thousands of documents. Beyond that point, processing slows, the variety of data likely increases and relationships between documents continue to grow exponentially.
This is in part due to technical constraints that are easily overlooked during a demonstration. Clean documents being prompted against comfortably fit in context windows, retrieval augmented generation (RAG) powered tasks don’t breach a top-k (the number of active, allowed or visible tasks stays within the defined limit), and the text inputs are generally clean. Recognising limitations and matching the right tool to the right task, rather than assuming a single solution scales linearly, is therefore an essential component of building AI-powered workflows and choosing specific tooling.
In practice, a flexible, fit-for-purpose approach can combine the use of broadly available generative AI models for low-stakes, exploratory work alongside deployment of a sophisticated platform, such as IQ.AI by FTI TechnologyÔ, for complex tasks and higher stakes work.
Validating Accuracy
While AI results aren’t always identical from one cycle to the next, the standard is to reach a level of accuracy and consistency that can be tested, demonstrated and defended. This can be achieved through sampling, calibrated to the stakes of the matter, and using AI to rank documents along a scale of relevance, with supporting details for the given ranking.
In practice and based on observations from experts, generative AI is currently more accurate at the extremes. More uncertainty is concentrated in the middle tier, which is not dissimilar to humans delivering accurate results when confident, and less so when unsure. This is where additional sampling and expert review add significant value.
The Cost Equation
Where cost is concerned, generative AI is introducing more options for managing it, particularly in matters with large volumes of data; for instance, using AI to support collection and initial review in-house to reduce the overall population sent for further review with outside counsel. However, the approach for each case will be unique and will continue to depend on the specifics of the matter, such as how quickly review must be completed, what data sources are in scope, the size of the review, etc.
Teams are gaining confidence comparing AI review exercises versus traditional approaches, however it can be challenging to predict the cost of the prompt refinement stage, which varies based on prompting experience, review protocol complexity, data quality and complexity, and several other factors. Increasingly, parties in litigation should continue to factor in costs around agreeing and defending the use of AI as usage becomes more commonplace.
Getting Ahead of the Argument
Attorneys are now regularly facing e-discovery negotiations or opposition related to how AI is used in the process. Legal teams that understand their tools and how they are used before deployment will be best positioned to withstand any scrutiny.
Conversely, where decisions about AI use in e-discovery were made without careful attention to defensibility and process documentation, issues are likely to arise, especially if collections or other phases were handled outside of the legal team’s oversight. Any team applying AI to e-discovery must fully understand the legal parameters and pitfalls, and the outputs should not be relied upon without being tested.
The same discipline that has always underpinned defensible discovery applies no matter what technologies are in place: a clear, contemporaneous record of what was done, at what stage and why, how prompts were designed, and what oversight was in place to validate results.
As the use of generative AI in discovery matures and case law related to its use accumulates, the legal teams that will fare best will be the ones treating cost, risk and accuracy as a framework to revisit every time the scale, the data or the stakes of a matter change.
Related Insights
Related Information
Published
September 23, 2026
Key Contacts
Senior Managing Director