AI in Ediscovery: Where It Works, Where It Exposes You
Insights from Everlaw’s CTO, Max Christoff
Everlaw Chief Technology Officer Max Christoff has a distinct childhood memory of his father, a patent litigator, spending 90 hours a week at a giant dining room table buried under stacks of bankers’ boxes, sticky notes, and highlighters. "I remember thinking, I don’t know precisely what a litigator does, but I definitely don’t want to do that," Christoff laughs. Years later, he finds himself solving that exact problem from the technology side.
We have officially passed the point where artificial intelligence in ediscovery is just a novelty. On defined litigation tasks, the math shows it's already meeting, and regularly beating, human review speeds and accuracy. But as the tech moves from simple acceleration to deep strategy, the line between an asset and a liability is thinning.
This exact tension was the catalyst for a recent fireside chat at LegalTechTalk 2026 in London, where Christoff joined Tom Whittaker, Director and Solicitor Advocate, Burges Salmon, and moderator Karina Morawska, Chief Operating Officer, Linklaters.
We recently sat down with Christoff for a deeper, one-on-one conversation. We wanted to unpack the operational realities of this shift, moving past the panel talk to look at exactly where AI delivers measurable value in ediscovery, and where it exposes teams to serious risk.
"Historically, ediscovery has operated like a waterfall," Christoff noted. The traditional process is familiar: negotiate a protocol, hand it off to a review team, and weeks or months later the litigator finally gets a readout. That linear model, he said, is breaking down. Attorneys are showing "real curiosity much earlier in the lifecycle," wanting an early read on the risks to shape their strategy, figuring out whether a matter is headed to trial or settlement before the traditional review engine even warms up.
Christoff was careful to distinguish this from traditional early case assessment. "I'm not describing what I would traditionally have called early case assessment, where your primary goal is to try to cull the material," he explained. "I'm really talking about understanding what the risks are in the content and starting to get a read on the case to help inform the strategy."
The shift changes how attorneys engage with their clients. "It allows them to show up as a more engaged partner throughout the process," Christoff said, "instead of saying, 'Well, we'll know more in a few months, and then I'll let you know.'"
AI Is Already Outperforming Human Review on Defined Tasks
The performance data is getting harder to ignore. When AI is applied to bounded, measurable tasks with clear criteria, it delivers strong, repeatable results.
First-pass document review is the clearest example because performance can be evaluated using established metrics. Across four live-litigation datasets, Everlaw's Coding Suggestions achieved an average recall of 89 percent. In one dataset where Everlaw could compare directly against first-pass human review, Coding Suggestions achieved 82 percent recall versus just 60 percent for human reviewers.
The results scale. In a government investigation, an Am Law 100 firm used Coding Suggestions on more than 126,000 documents, achieved over 90 percent accuracy, and coded the entire document set in approximately one day. This reduced the total review time by 50 percent to 67 percent with only a quarter of the personnel needed for a comparable managed review. Similarly, Orrick, a global law firm, used the tool for an intellectual property case and cut document review costs by more than half.
Another firm, Dinsmore & Shohl, took a layered approach by stacking strategic search, Coding Suggestions, Predictive Coding, statistical validation, and human review across 26,000 contracts. The result? They boosted their recall rate to an exceptional 98.2 percent.
Christoff sees ediscovery as a particularly strong fit for AI but emphasizes that fit matters more than capability. "It's not just, do we have any technology, and can we copy and paste it into every situation," he said. The real question is whether the technology matches the problem. In ediscovery, he argues, the match is strong: AI can "review documents and find needles in haystacks in a way that human reviewers just feasibly cannot do with any definition of proportionality."
These numbers give legal teams something concrete to evaluate and defend. The key to success lies in defining the task clearly, iterating on the prompts, validating performance against a controlled sample, and only then scaling up.
The Real Shift: Strategy-First Discovery
The bigger opportunity goes beyond faster document checking. Tools like Everlaw's Deep Dive allow legal teams to ask natural-language questions across an entire document set and receive citation-backed answers, shifting exactly when insight happens in a case.
Instead of waiting for a massive review team to trudge through hundreds of thousands of files before the core narrative emerges, core teams can pinpoint key actors, themes, and timelines at the absolute outset. Review becomes guided by what the team already knows, rather than a blind search for what they might find.
Scale is a massive differentiator here. Christoff explained that many legal AI tools remain constrained to tiny document sets or force users into fragmented third-party systems for analysis, breaking data governance. Everlaw, by contrast, has reported that Deep Dive has been used in cases with millions of documents.
However, Christoff warns that the usefulness of large-scale analysis hinges on engineering choices. Legal teams, he said, need to understand "what knowledge is embedded within an LLM’s training data versus what knowledge is actually in their documents." An LLM trained on the open internet will always produce a plausible answer to why the sky is blue, for example.
Christoff's team runs this test regularly: type that question into Deep Dive with a matter loaded, and the system will tell you it cannot find any reason why the sky is blue. "That's because we have designed a system to ensure that the answers are grounded in the document set, and not coming from the inherent knowledge of the training data," he explained.
Reliable tools, Christoff added, should ensure that every statement is supported directly by the documents, with citations for verification. "They should freely state when they don’t have enough information to answer your question. We know the models are trained to be overconfident and to please you, and that's something we take pains to avoid in our own system design."
Where AI Exposes Teams
The flip side of this efficiency is compliance risk. AI exposes teams to liability when it's treated as a series of disconnected, ad hoc shortcuts. Exporting sensitive evidence into standalone consumer tools fragments the record, breaks data governance, and leaves risk partners entirely in the dark about client confidentiality.
The most common mistake, Christoff said, is that "the word AI gets used as a fairly broad brush." Firms say they allow AI or they ban AI, but that covers vastly different safety profiles. There is, he noted, "a world of difference between consumer-grade tools, where dropping data into a public interface risks the loss of attorney-client privilege, and enterprise-grade systems built from the ground up with zero data retention and zero model training on client data."
There’s also what Christoff calls the “blinking cursor problem.” Consumer AI tools present a blank slate to the user that essentially asks “What you’d like to do today?” — a blinking cursor, capturing the imagination with apparently powerful technology placed directly in attorneys’ hands. The results can be stunning on a small scale.
"Fine for ten documents," Christoff said. "By the time you reach a hundred, or a thousand, or ten thousand, what you actually needed was a statistical, explainable approach.” He explained that you need to be able to divide up the work with a team of other legal professionals. You need an audit trail to keep track of who did what and what's already been reviewed and what hasn't. “And by then you are fairly deep in,” he said.
The lesson, he argues, is that the setup needs to be proportionate to scale. A single contract being examined carefully requires a very different process from 1.5 million documents dumped from a large company. “Same tool, different process. Get that mismatch wrong and you are in trouble before the technology has had a chance to fail you," Christoff said.
This risk becomes more urgent as legal tools become more agentic. The more systems can do on behalf of users, such as orchestrating workflows, making preliminary determinations, and organizing evidence, the more vital strict internal guardrails become.
Christoff described how this needs to work in practice: "It's not just an agent randomly going off and doing things. You can see exactly what they did, when they did it, there's an audit trail, there's logging. Maybe that work goes to another team for a second-pass review."
What Trustworthy Legal AI Requires
Trustworthy legal AI requires strict operational discipline, moving past abstract promises of security. True safety means giving law firms absolute control over their environment, including the clear choice to opt in or out of generative features. When the tech is deployed, it needs to inspire confidence by focusing on highly specific use cases where every single insight is anchored directly to verifiable evidence.
There is no room for black-box guesswork or loose data boundaries here. Third-party AI tools must undergo rigorous legal and security screening before they ever touch customer data, and that data must never be used to train generalized models. Ultimately, these principles only matter if they are backed by daily, practical guardrails like human checkpoints, precise user permissions, and unalterable audit trails.
The Platform Consolidation Question
As AI capabilities mature, the argument for keeping all data within a single platform gets stronger. The most valuable systems will capture repeatable review protocols, validated workflows, and user patterns that help future matters run more effectively. Legal teams benefit most when AI operates across the actual system of record, eliminating duplicate uploads or disconnected workspaces.
According to Christoff, this ecosystem approach is where the concept of "AI memory" becomes critical, operating at three distinct levels.
The first level is data. Christoff describes individual pieces of evidence such as an email, a PDF, or a text message as "atoms." From those atoms, teams build "molecules": established facts backed by citations. For example, based on three emails and a text message, a team might establish that the CFO knew about an audit finding by June 1. "That's a fact, and the fact has citations going back to those atoms," he explained. The platform should retain those facts so that deeper analysis can build on what's already been established, rather than regenerating everything from scratch.
The second level is team memory. In a collaborative workflow, the software captures how a team operates, the context of their matters, what's been done, what's left, who's working on what. "If someone's going out on leave, I know exactly what that employee was working on, and I can jump in where they left off," Christoff said. This is also how he sees agentic AI entering the picture safely, as agents whose actions are fully logged, audited, and passed through human approval gates, rather than operating autonomously.
The third level is institutional. "It's the memory of the law firm itself," Christoff added. "When a firm handles a specific type of litigation, the platform should retain those historical workflows, motion templates, and past strategies." The goal is a system where no team has to start from zero and where each matter builds on everything the firm has learned before.
Getting It Right
The legal teams that lead the next era of practice will be the ones deploying AI inside workflows that are measurable, repeatable, and defensible.
AI thrives when it is applied to bounded tasks with clear criteria and proper validation. It also thrives at the true scale of modern litigation, operating directly inside the secure system where evidence, permissions, citations, and audit trails already live.
Conversely, it exposes teams to risks when it's used without structure, without central governance, and without the human judgment that must remain at the center of the law. The technology is entirely ready. The only question left is whether firm discipline can keep pace.
