Discovery Is Quietly Reorganizing Itself Around AI
Editor’s Note: For years, eDiscovery has been defined by a simple assumption: the larger the data set, the longer it takes to understand what matters. Generative AI (GenAI) is challenging that assumption, giving legal teams the ability to surface key facts, identify themes, and make informed strategic decisions far earlier in the lifecycle of a matter. This article distills key insights from the HaystackID® webcast where industry leaders explored how AI-assisted review is reshaping document review, early case assessment, validation, and governance. The discussion makes clear that the conversation has moved beyond whether AI belongs in discovery to how organizations can deploy it responsibly and defensibly. As the pace of adoption accelerates, success will depend not only on embracing new technology but on applying the legal judgment, validation, and governance that make its results trustworthy.
Discovery is Quietly Reorganizing Itself Around AI
By HaystackID Staff
Until recently, the playbook for a large data breach matter was familiar: millions of documents, a privilege log to build, and months of linear review before anyone understood what the data actually contained. Cases settled before review ever turned up the one email that would have changed the negotiation in week one, not because the document wasn’t there, but because nobody could get to it fast enough.
That kind of timeline used to be the cost of doing business in eDiscovery. Now it’s the exact scenario the panelists on HaystackID’s webcast, “From Hype to Workflow: Insights from Experts on the Impact of AI on eDiscovery,” argue no longer has to happen. The session brought together four practitioners who’ve watched GenAI compress that timeline from months to days.
The panel spent an hour on a single argument: discovery workflows built before these tools existed are already out of date, whether the organizations running them have noticed. Two years ago, letting an algorithm make final relevance calls would have triggered pushback from most legal departments. By the panelists’ account, that resistance has largely evaporated, and the panelists agreed the pace of the shift has outstripped anything they saw during the rollout of technology-assisted review (TAR) a decade earlier.
A Faster Curve Than TAR Ever Drew
Where TAR adoption crawled upward over years, GenAI jumped almost immediately from experimentation to default practice, according to Cristin Traylor, who spent more than two decades in Big Law before joining Relativity. Traylor cited her own company’s numbers to make the point.
“We’ve got hundreds of customers using AI for a review, for privilege. Over a hundred million documents have been analyzed,” said Traylor, Senior Director, AI Transformation and Law Firm Strategy, Relativity. “This is not something that just a few people are doing. We have law firms that are standardizing on it, using it, like I said, in every matter, which is really amazing because if we think about where we were before with TAR and active learning, there was adoption, but it was a very slow curve up. And here, we started over here, and we are way up here.”
Jim Sullivan, Chief Executive Officer, eDiscovery AI®, traced how the conversation itself has moved. In 2023, practitioners insisted that the technology would only augment reviewers and never replace them. A year or two later, that framing shifted to carving exceptions for privilege review specifically. Now, he said, organizations are running privilege calls as a full substitute for manual logging, a pattern Sullivan described as objections dissolving faster than the industry can restate them.
Reviewers Who Iterate Instead of Read
Review teams aren’t disappearing so much as review work is changing shape, according to the panelists. Large rooms of people reading line by line are giving way to smaller groups of subject-matter experts, the people who understand both the law and the data, spending their time refining prompts, sampling results, and signing off on quality rather than eyeballing every document that crosses a screen.
Read together, their accounts suggest the work hasn’t left human hands so much as moved to a different part of the process, one where legal judgment carries more weight than it did when that judgment was spread thin across a room of contract reviewers making the same binary call thousands of times over.
“I went in the last two years from ‘I never want to do a human-first level review. AI is going to completely replace contract review,’ to the point where I am now where it is that we have to completely reimagine what our discovery obligations are because of the way that we can access, understand, and read data,” said Esther Birnbaum, EVP of Data Intelligence, HaystackID.
The distinction matters: her point wasn’t that people step aside, but that the profession has to rethink where expertise gets applied, earlier, and with more scrutiny, rather than not at all. Later in the discussion, Birnbaum raised the possibility that linear, document-by-document review may eventually give way to natural-language querying of a full collection, though she framed that as a shift in method, not a case for removing lawyers from the process.
Find Your Smoking Gun, Sooner
One of the more concrete shifts the panel described involves timing rather than technology alone. Early case assessment (ECA) has traditionally meant a rough first pass before the work of review begins. With GenAI capable of indexing and classifying a full dataset almost immediately, organizations can surface key custodians, hot documents, and case themes in the first days of a matter instead of near the end.
Finding a decisive document at the outset, rather than months into linear review, changes the kind of decisions in-house counsel and corporate leadership have to make early rather than late: whether a case settles, whether someone gets terminated, whether a company has a disclosure obligation. This reordering can eliminate unnecessary review altogether, according to Sullivan.
“We can identify all the key documents, key themes, seeing everything very early on at a pretty low price point,” he said.
Teams can understand a data set thoroughly at low cost, decide whether the matter will actually proceed, and run a full relevance review only if the case doesn’t resolve first. Surprises that used to surface at the eleventh hour, such as an unidentified custodian or a missing data source, now tend to surface in week one instead.
Traylor pushed back gently on treating ECA as a single concept, noting the term means different things to different practitioners. For some, it’s shorthand for culling: deduplication, threading, narrowing a collection down to a defensible review population. For others, it means something closer to case intelligence: querying a data set in plain language to understand what happened before committing to a strategy. Keeping those two goals distinct, Traylor argued, helps teams choose the right tool for the right job rather than expecting a single workflow to accomplish both at once.
Birnbaum’s team defaults to running new collections through Case Insight™, HaystackID’s matter-intelligence engine, before anything else happens downstream, regardless of which flavor of ECA a given matter calls for, because the resulting overview of the data shapes every decision that follows it.
The Real Risk is Poor Governance
None of that speed comes free of obligation. When a webcast attendee asked whether ECA can be run safely through a general-purpose model like ChatGPT, or whether the work demands purpose-built tooling, Favro flagged it as the question of the day before handing it to Aleida Gonzalez, Global Advisory Managing Director, HaystackID.
Gonzalez laid out the actual diligence test a practitioner has to run before making that call.
“The tool that you’re going to use should be appropriate for the purpose of your search,” she said, explaining that whatever tool a team reaches for, someone has to have actually read what they signed up for.
She added, “Did you read the terms of use? Have you actually read your contract for ChatGPT? … Personally, I do not trust putting client information into an AI, using client names, any specific or sensitive material or information into an AI. Now, granted, a lot of legal tools probably and most likely have far more safety features than a broad general AI tool such as ChatGPT.”
The answer wasn’t a hedge; it was a framework, and one that lines up with how HaystackID’s governance practice approaches the same problem: classify the risk of a given AI use case before choosing a tool for it, not after.
Our company’s governance practice formalizes the same discipline Gonzalez described: inventory where AI is already running inside an organization, including the “shadow AI” that shows up when employees turn to consumer chatbots on their own, classify each use case by risk, validate a system’s security, privacy, and fairness characteristics before it touches real matters, and produce documentation that can hold up in front of a regulator, an auditor, or opposing counsel.
HaystackID aligns that work with frameworks including the NIST AI Risk Management Framework, ISO/IEC 42001, and the EU AI Act’s risk-classification requirements, so the resulting evidence isn’t built from scratch every time a new tool or matter comes along.
Sullivan added the practical checklist that flows from that same discipline to verify that:
- Vendor data isn’t used to train underlying models.
- Prompts aren’t retained by a provider.
- Content-filter triggers won’t route sensitive material to outside human reviewers.
- Data stays within the required region.
Birnbaum extended the point past infrastructure security into something more fundamental: whether the tool is actually built for the job.
“It’s also really important to talk about why these tools won’t necessarily be the best tools for you to use because they’re not fit-for-purpose tools,” she said.
Drawing on two years of experience refining AI review workflows, Birnbaum explained that purpose-built review platforms require guardrails to prevent models from “hallucinating answers” or becoming “forced to give answers when they can’t,” producing confident but incorrect classifications instead of acknowledging uncertainty.
That risk, she argued, sits outside the data-security checklist entirely: it’s not about whether a vendor can see your documents, but whether the tool’s output can be trusted in the first place.
Gonzalez tied it back to ABA Formal Opinion 512, which requires lawyers to read and understand the terms governing any AI tool they rely on, and closed the segment with a reminder that weaker governance means greater exposure for the organization.
Search Terms: Not Cashing Out the 401(k) Just Yet
While GenAI has gained traction, keyword search isn’t throwing in the towel. Sullivan argued that running an entire collection through an AI relevance tool is still too expensive at today’s per-document pricing to skip pre-culling altogether, which means search terms still serve as the first filter before AI-assisted review begins.
That tension surfaced directly in Schulte v. LinkedIn Corp., a case Favro raised partway through the discussion and later unpacked in more detail in his article, ‘“Search, Forward” with AI: Court Approves AI-Assisted Review Workflow for eDiscovery.’ LinkedIn had used twenty-five search strings to cull a multi-terabyte collection down to roughly 200,000 documents before running an AI-assisted review platform on what remained, and the plaintiffs argued the technology “had been designed to both ingest and then make responsiveness calls on an entire data set” without any pre-culling at all. Judge Laurel Beeler disagreed, finding that pre-culling with search terms satisfied “the reasonableness and proportionality standards of Rules 26(b) and 34(b)(2)” and noting that requiring LinkedIn to run its full universe of documents through the AI technology would create “significant costs related to processing, hosting, and human review.”
Traylor noted the ruling largely tracked what the parties had already agreed to in their ESI protocol; the real dispute wasn’t over whether AI could make responsiveness calls, but over how the reviewed population got assembled in the first place. Favro drew the parallel to TAR-era case law directly: courts had already worked out, more than a decade earlier, that pre-culling with search terms before running technology-assisted review was reasonable given the cost of processing everything, and Schulte simply extended that same logic to AI.
Same Metrics, Sharper Scrutiny
If there’s one place the panel insisted nothing needs reinventing, it’s validation. Recall, precision, and elusion testing remain the backbone of defensible review, whether the underlying engine is TAR or a large language model. Traylor described a workflow where teams iterate on a prompt against a sample set, check the resulting precision and recall, adjust as needed, and either validate before running the full population through the tool or validate afterward, but the metrics themselves haven’t changed.
“No one’s calculating recall and precision on keywords in 2009,” Sullivan said.
GenAI gets validated constantly and by design; Birnbaum treats that scrutiny as standard practice rather than a special exception.
“There’s so much more interest in it,” she said of prompt validation, pointing out the irony: keyword search built its reputation on decades of unchecked assumptions, while AI-assisted review gets checked at every stage almost as a matter of course. Favro connected this to Judge Peck’s warning from the Rio Tinto opinion, cautioning against holding newer technology to a stricter standard than the tools it replaced.
Less Admin, More Lawyering
The discussion closed on a question from an audience member wondering whether GenAI ultimately shrinks the need for lawyers altogether. Sullivan’s answer was immediate: “Yes.” When Favro turned to Traylor and Birnbaum for their own take, both offered more of a caveat. Traylor argued the technology frees attorneys to focus on judgment calls and case strategy rather than document logistics.
“We still need lawyers, but now we can get back to doing what matters. Actually lawyering, helping to advise our clients and, again, putting the strategy together for the case and not worry about reviewing documents,” she said.
The panel didn’t agree on everything; Sullivan, Traylor, and Birnbaum gave genuinely different answers on what’s left for lawyers to do. But on the core question of whether GenAI belongs in the workflow, all three were in agreement. The open question wasn’t if; it was how fast.
About HaystackID®
HaystackID® solves complex data challenges related to legal, compliance, regulatory, and cyber requirements. Core offerings include Global Advisory, Cybersecurity, Core Intelligence AI™, and ReviewRight® Global Managed Review, supported by its unified CoreFlex™ service interface and eDiscovery AI® technology. Recognized globally by industry leaders, including Chambers, Gartner, IDC, and Legaltech News, HaystackID helps corporations and legal practices manage data gravity, where information demands action, and workflow gravity, where critical requirements demand coordinated expertise, delivering innovative solutions with a continual focus on security, privacy, and integrity. Learn more at HaystackID.com.
Assisted by GAI and LLM technologies.
SOURCE: HaystackID
Advisory Note: As organizations adopt GenAI across the discovery lifecycle, the ability to understand a matter early is becoming a strategic advantage rather than a luxury. HaystackID’s Core Intelligence AI Case Insight, delivered in partnership with eDiscovery AI, is a GenAI-powered matter intelligence engine that transforms large, complex data collections into actionable case intelligence within days. By identifying key documents, mapping relationships, surfacing emerging themes, and highlighting potential risks at the outset of a matter, Case Insight enables legal teams to make better-informed decisions before investing in full-scale review. Combined with expert legal oversight and rigorous validation, the platform delivers rapid insight while maintaining the defensibility today’s matters demand. Whether informing early case assessment, refining review strategy, or accelerating fact development, Case Insight helps organizations move from reactive document review to proactive, intelligence-driven discovery.