eDiscovery has an unusual position in legal technology. It was using machine learning at scale a decade before the rest of the profession discovered AI, and courts have been ruling on the defensibility of algorithmic review since 2012. So when generative AI arrived, this was the one area with an existing framework for the question everyone else was asking for the first time: how do you prove to a court that a machine did the work properly?
What was already true
Technology assisted review, usually called predictive coding, has been accepted in United States federal courts since Da Silva Moore in 2012 and in England and Wales since Pyrrho Investments in 2016. The mechanism is straightforward: senior reviewers code a sample, the system learns from those decisions, and it prioritises or classifies the remaining population.
That workflow already removed the bulk of linear review from large matters. Anyone claiming generative AI introduced machine assistance to discovery is selling something.
What generative AI actually changed
Three things, and they are real.
Classification without training data. Traditional predictive coding needs a seed set, which means senior time spent coding before anything happens. Generative models can classify against a written description of what you are looking for. On a small or fast-moving matter where building a seed set was never economic, this is a genuine unlock.
Summarisation as a first-class output. Older tools told you a document was responsive. Newer ones tell you what it says. For an early case assessment where the question is what happened rather than what is producible, that is a different kind of usefulness.
Language and question answering across the set. Asking a natural language question of a corpus and getting a cited answer changes how a case team orients on a new matter. Cross-language handling has also improved substantially, which matters in cross-border disputes.
What did not change
Defensibility still has to be documented. A court does not care that the tool was impressive. It cares about your process: what you searched, how you validated, what your recall and precision were, and whether the other side was told. Sampling and validation protocols did not become optional.
Proportionality still governs scope. Cheaper review does not mean broader collection is now appropriate. If anything, the opposite argument gets made against you, since lower cost weakens the burden objection.
Privilege review remains the hard part. Privilege is contextual, relational and frequently turns on who was in the room. It is the area where automated classification is weakest and the consequence of an error is highest. Inadvertent production of privileged material is not a problem you can technology your way out of.
Chain of custody and collection defensibility. Nothing at the review layer helps if collection was flawed. Most discovery disputes still turn on collection, preservation and spoliation rather than on review methodology.
The question you will be asked
Sooner or later, opposing counsel or the court will ask how the AI reached its conclusions. Answering “the vendor’s model determined responsiveness” is not sufficient.
What you need to be able to produce: what the system was instructed to look for, in writing; how the output was validated, with the sampling protocol and the numbers; the recall and precision achieved on a control set; who reviewed what proportion of the population manually; and what disclosure was made to the other side and when.
Firms that treat these as artefacts to be generated during the matter have a straightforward conversation. Firms that reconstruct them afterwards do not.
Buying considerations
- Hosting and data residency. Cross-border matters make this a threshold question, not a detail. Establish where data sits and under whose jurisdiction before anything is uploaded.
- Pricing model. Per gigabyte hosted, per document reviewed, per user, or a hybrid. These produce wildly different totals on the same matter. Model your actual expected volume against each, including the long tail where data sits hosted for months after active review ends.
- Whether AI features cost extra. Frequently they do, and frequently that is not in the headline price.
- Validation tooling. Does the platform produce the sampling and recall reporting you will need, or will you build it in a spreadsheet?
- Exit. Getting a production set out of a platform in a usable load file format is a real cost. Ask before you are locked in.
The service provider question
Most firms below a certain size should not run their own discovery platform. The expertise required is real, it is specialised, and it is not the same as being a good litigator. There is a reason that partnering with an eDiscovery service provider remains the standard model even for firms with substantial disputes practices.
What has changed is that the provider conversation now needs to cover AI methodology explicitly. Ask what tools they use, how they validate, and what they can produce if the methodology is challenged. A provider who cannot answer that in detail is a liability rather than a shield.
The honest summary
Generative AI made early case assessment materially faster and made small-matter review economic in ways it was not before. It did not change what a court requires of you, it did not solve privilege, and it did not remove the need for a documented, defensible process.
The firms getting value here are the ones that treated it as an improvement to an existing discipline rather than a replacement for one.
Where the legal industry reads first.
Enjoyed this article? Get the biggest legal industry updates, deals, appointments, insights and expert interviews in your inbox, free.
No spam. Unsubscribe anytime.