AI-first discovery: Putting human judgement where it matters

AI is increasingly being used to carry much of the volume in discovery. But value is a different story: it’s won or lost in how senior expertise is used to keep review both lean and defensible. 

As soon as generative AI (GenAI) tools entered legal workflows, discovery stood out as one of the clearest use cases for AI—and with good reason. Earlier advances, from basic keyword lists through TAR and continuous active learning, helped with search and scale, but left most of the underlying work untouched. GenAI is a genuine step-change. It moves beyond incremental efficiency gains to a machine-first, human-verified model that can operate as the default in even the most data-intensive cases. 

Cutting down review time is the easy part—the real test is whether human judgment is brought to bear where it counts.

In our recent work helping dispute teams build GenAI into review design, one truth has become impossible to miss: as the tools remove more routine volume, the value of the review increasingly depends on how carefully that judgment is deployed.
This is where we’re seeing better reviews set themselves apart from merely faster ones. Senior involvement becomes more concentrated: translating the issues into instructions the model can be trusted to follow, deciding when its analysis needs to be challenged or recalibrated, and making the decisions that determine what’s surfaced, investigated and ultimately produced. Get those right, and teams move faster, go deeper where the evidence demands it, and deliver a review they can confidently defend. Get them wrong—or neglect to make them at all—and efficiency gains disappear fast amid omissions, delays, and difficult questions about the integrity of the disclosure. 

From capability to allocation: planning matters as much as platform 
The promise of shorter timelines and leaner teams still drives much of the commentary around AI in discovery. What we frequently see in live disputes is cost and delay locked in at the outset—in decisions about how broadly to scope data, how much to cull through search and filtering, and how far the team is willing to leave decisions to automation with minimal oversight. 

Those choices still tend to be treated as part of routine case management and project planning. In reality, they’re allocation calls: on where scarce senior attention will have the greatest impact, and where automation can safely carry the load. Make them early, document them clearly, and revisit them when the evidence or case theory changes. A new issue can then be addressed with a focused re-run, rather than reopening the document population for a second full review. 

Machine-first, human-owned review 
An AI-first, human-verified approach doesn’t just bolt GenAI onto a traditional linear review process; it rebuilds the flow around the tools. AI drives the first pass by organizing datasets across sources and languages, generating issue-specific views and summaries, and helping supervisors decide what needs closer scrutiny as new facts emerge or theories change. Human reviewers no longer spend their time acting as first-pass filters; they move into designing and steering the operating model—the higher-value work the AI-first model is meant to unlock.  

Faster insights—and faster errors  
GenAI doesn’t just reduce processing time—it accelerates conclusions that read as settled and authoritative even when they’re wrong. The risk is no longer only what gets overlooked, but what is accepted and repeated on the back of flawed output. In February, a U.S. federal appeals court sanctioned counsel over an AI-assisted brief containing fabricated quotations and citations that went unchecked before submission. Discovery is similarly exposed: if AI-generated summaries or issue tags pass into productions or case theories without a robust verification step, an error can compound quickly. 

This vulnerability means quality control cannot be limited to an end-of-process courtesy check. It must be designed in from the start—in sampling against ground truth, structured validation of AI-led classifications, clear escalation routes, and rules for deciding what happens when AI output and human reviewers disagree. Done properly, it puts senior time where scrutiny and challenge are most likely and most consequential, ensuring the process runs lean without defensibility running thin. 

Architecture as case strategy in cross-border work 
Much AI marketing still assumes a simple path: data goes to the cloud, the “best” platform processes it, and coded material comes back. We have yet to encounter a cross-border dispute or investigation that fits such a neat model, and the complexity only seems likely to increase. Data-sovereignty regimes are tightening, AI-specific provisions now increasingly appear in protective orders and ESI protocols, and courts are increasingly wary of confidential material being sent to public, cloud-hosted tools. The wrong platform, hosting model, or data route can leave an entire review approach unusable. 

In global disputes, review architecture has become a key part of case strategy: the hosting model and data route now determine which tools can be used at all. A “just upload it to our usual platform” approach is no longer feasible in the growing number of jurisdictions with data-sovereignty and localization rules that complicate multi-market discovery. China presents one of the more complex settings for cross-border review. We begin by mapping the data, the permitted routes and the review decisions that must stay local—with each choice shaped by PRC restrictions on data export and offshore processing—and then build the workflow around them. Our familiarity with the legal, technical, and  operational terrain is key to making a compliant cross-border model workable in practice.

Infrastructure that fits the jurisdiction
Having an impressive AI platform at group level isn’t enough; the work needs an in-country environment that can deliver machine-assisted efficiency and still meet local demands. AlixPartners has developed dedicated infrastructure for hosting setup, processing, and review entirely inside the jurisdiction and under the client’s control. The models do not send data to third-party clouds or use client material for training; data-handling, access and destruction protocols can be described and justified to courts or regulators when required.

Clients don’t need to see the engineering, but they do need to be confident that work can still be done within strict rules, without losing the benefits of AI. That combination— local deployment with globally consistent outcomes—is still the exception in cross-border matters. 

Overview: Turning AI capability into defensible speed
In discovery-heavy work, the hard part is no longer capability; it’s whether the workflow can withstand legal and regulatory scrutiny. Savings disappear fast if the AI-enabled process cannot be explained clearly. A credible decision trail has to go beyond search terms and hit counts, without turning into a second set of pleadings.

Maximizing value from a machine-first, human-verified model depends on four choices: 

• Use AI to focus senior attention. Push the tools wherever they can responsibly narrow the review surface to what genuinely warrants that judgment—within architectures designed for data-localization and protective-order constraints. 
• Let AI lead mapping and triage. Make AI-led triage the starting point for review design, with verification designed in from the outset. 
• Define human-owned calls early. Set out, in detail, which decisions cannot be delegated, concentrating reviewers’ time where it’s needed and making clear which judgments will need to be explainable.  
• Invest in in-country AI where the rules demand it. Deploy practitioner-designed local environments for China and other strict data-localization regimes—so cross-border work can run on the same operating model and judgment standards, even when the underlying infrastructure has to defer. 

We’ve seen the difference in practice: review teams that work this way stay busy, but they spend less time on data that was never material, and more on the decisions that shape outcomes—especially in the toughest jurisdictions.