By Ralph Losey
Introduction: An Old Friend Retires and is Replaced by a New and Improved Version (I can relate!)
On September 1, 2026, EDRM released EDRM 2.0, the first total substantive rewrite of the Electronic Discovery Reference Model since EDRM was started in 2005 by George Socha and Tom Gelbmann. That was also when I first started to specialize in e-discovery. The EDRM became my bible, along with the only book on the subject then available, A Process of Illumination: The Practical Guide to Electronic Discovery. The author of that book was none other than Mary Mack, the current CEO and Chief Legal Technologist of EDRM. With those two sources, many e-discovery conferences, and a steady stream of cases, I started specializing in discovery in 2006,.
It has been virtually non-stop e-discovery ever since. Up to my retirement this year, I had the dubious pleasure of directly handling or supervising thousands of e-discovery projects. I did so as a partner in two major law firms, and, unlike many e-discovery specialists, I never worked for a vendor, although I became very familiar with the operations of old Kroll Ontrack. Like many, I had various versions of the EDRM chart memorized. I often used it with vendors, other lawyers, judges, and clients.
This new chart, version 2.0, is a big improvement over all of the prior versions of EDRM and really deserves to be called version 2.0. It reflects the substantial progress the profession has made in e-discovery over the past twenty years.
When I began writing weekly for my e-Discovery Team blog in 2006, Version 1.0 of the EDRM was a year old and already well established. I just counted: during those twenty years, EDRM was mentioned in 104 different articles on my blog.
EDRM Model 2.0 Was Worth the Wait
After more than twenty years, the EDRM was ready for a substantial redesign. Version 2.0 meets almost all of my expectations, and the few concerns I have are minor. Congratulations to the approximately 150 people who participated in this massive project. The effort was led by co-project trustees Shannon Lex Bales, Rian Kennedy, Stephanie Clerkin, and Brett Burney, with leadership support from David R. Cohen and Mary Mack.
Although EDRM has published many of my articles over the last few years, and I have sometimes served on its very large advisory board, I had nothing to do with this project. The views expressed here are, as always, entirely my own; and yes, AI did assist in writing this article, but not very much. As I always say, you must supervise and keep control of AI, and if you do, you may benefit from its assistance.
The Old Model Needed to be Replaced
I make this comment affectionately. Just take a look at the EDRM version in 2009 and see for yourself.
No competent e-discovery professional ever believed that the old EDRM diagram had to be followed literally from left to right. I would often make my own models to supplement the EDRM standard. For instance, I made a version that I called Electronic Discovery Best Practices (EDBP), which had COOPERATION as its own box. That did not go over well, even though I credited the original idea to Jason R. Baron. My work flow model had many similarities with the EDRM except that it only included legal services.
Anyway, we all knew that the apparently linear EDRM diagram was a little misleading because, in the real world, discovery almost always doubles back on itself. It was a recursive process, not linear. Preservation and Review often uncover new custodians. Processing finds unexpected file types. A witness interview sends everyone back to collect more devices. The old diagram was a model, not an instruction manual.
Still, pictures have power. Put a series of boxes from left to right and people naturally see an assembly line. Identification goes to Preservation, which goes to Collection, which goes to Processing, and eventually boxes of documents come out the other end. Discovery starts to look like an assembly line.
Modern information does not cooperate with that picture. A custodian may have email in multiple accounts but mainly use texts, Teams chats, or other forms of communication, even when management expressly prohibits them. In fact, upper-level managers are often the worst violators of technology restrictions and information-governance rules. And, of course, there are the documents! They can be in many places, and the most important ones are often where they should not be—on employees’ home computers or backup tapes stored in an IT employee’s basement. Today, important ESI may also consist of obscure system logs and other automatically generated data. Only a few people in IT may know those sources exist, and most are not motivated to volunteer that information to lawyers.
EDRM itself describes 2.0 as a conceptual model rather than a literal linear workflow. Practitioners may perform steps in different orders, repeat them, and return to earlier stages as their understanding changes. That is reality. Version 2.0 of the EDRM now reflects that. It encourages us to stop thinking of discovery as data being pushed down a pipe. Instead, EDRM 2.0 allows us to see e-discovery as a feedback system.
The New Diagram Correctly Pictures Today’s e-Discovery Practices
The new diagram has several major structural changes that correct the weaknesses of old models of EDRM. For one thing, there are now eight boxes, instead of nine, and two new bars at the bottom. The colors are improved to as is the overall look-and-feel of the model.
Information Management appeared as the far-left box in early versions, while later versions used Information Governance in that position. That box has now been removed. Information Governance is instead represented by the gray foundational band beneath the entire EDRM 2.0 model. A new Disposition box appears at the far right, where Presentation formerly ended the process. In my mind, Disposition is directly related to Information Governance: What do you do with the ESI at the end of the day?
Analysis used to be box seven of nine—a beloved spot, to be sure, but still just another step. It came after Review and before Production. In real-world e-discovery, however, Analysis was always constant, or should have been. I think it was the main thing e-discovery lawyers were supposed to do. EDRM seems to agree: Analysis has been removed as a separate box and converted into a blue band beneath the eight phase boxes. The band shows that Analysis informs and connects all the phases. As further explained later in this article, I think this is the most important change in Version 2.0. Congratulations to the EDRM writing team for coming up with this revision.
A close second in importance for me is the elevation of Preservation within the new Data Acquisition framework. Preservation is not new to EDRM, of course, but its new placement emphasizes its close interaction with Identification, Collection, and Processing. It is often about time and place and includes the famous question: Has the duty to preserve been triggered? To which I have often replied: Has the time arrived for protection of attorney work product?
Information Governance Now Moves Underneath Everything
Another change I support is placing Information Governance—and the Information Governance Reference Model (IGRM) shown below to the bottom of the EDRM model. It is, I assume, represented as the flat grey base at the bottom. I am not sure about anything having to do with Information Governance. I gave up trying to follow this complex subject many years ago after my embarrassing loss in a debate in London against Jason R. Baron on the subject.
I suppose it makes sense to include the IGRM and Information Governance as underlying the entire discovery process, much like Analysis. For one, you cannot preserve it if nobody knows where it is stored. Believe me that question can be incredibly complicated for a big organization. For them, only an AI could possibly understand it, and only the latest scary-type frontier models at that.
EDRM released IGRM Version 4.1 in June 2026 and specifically described it as addressing governance in an era shaped by AI, privacy concerns, and cross-functional data management. Mary Mack emphasized that information governance is not the job of a single organizational department. See EDRM, EDRM Releases Information Governance Reference Model v4.1.
Consider a familiar litigation problem. A complaint arrives alleging discriminatory termination. Legal identifies ten obvious custodians. Twenty years ago, we might have started with their email and network shares. Today, one interview may reveal Teams messages, shared cloud documents, HR workflow data, mobile communications, computer logs, and perhaps an AI system used to help evaluate employees.
At that point, information governance is no longer a records-management sideshow. It determines whether you can even find the evidence. Many times you cannot. Then you have to try to explain that to opposing counsel; and, after that fails, to the judge. Then comes my favorite kind of hearing—sanctions—with everyone probing states of mind for negligence, bad faith, or, under Rule 37(e)(2), an intent to deprive.
AI can help sort it out but also adds to the mess of too much information. Organizations are beginning to generate machine-created summaries, transcripts, recommendations, classifications, prompts, responses, embeddings, logs, and other information that may matter in future disputes. Deciding what should exist, how long it should exist, who controls it, and when it should disappear increasingly precedes the lawsuit.
The new EDRM gets that relationship right.
Data Acquisition: Four Boxes That Actually Talk to Each Other—and Disposition Finally Arrives
EDRM 2.0 groups Identification, Preservation, Collection, and Processing within a larger Data Acquisition structure. That too reflects what experienced practitioners already know. I imagine this was the subject of considerable discussion in the writing groups. I, for one, would like to hear the explanation: How is Processing part of Data Acquisition? It seems different to me.
The Data Acquisition collection boxes all grouped together into a larger box, which is something no prior version of EDRM has ever done before, I do see how the four activities would frequently overlap. Suppose litigation counsel learns during an initial interview that a key employee just resigned. That often means their computer is scheduled to be wiped, might already have been wiped. I handled a similar matter in which a routine process destroyed the ESI of a key player after a litigation hold had been issued. The client spent heavily defending sanctions motions, but we defeated them after testimony and argument persuaded a wise judge that the circumstances did not warrant sanctions.
Preservation cannot wait for Identification to finish. Collection of volatile or otherwise vulnerable information may have to begin immediately, sometimes before identification is complete. Processing may start as the first collections arrive while investigation of other sources continues. In fact, I have had cases in which we began Collection immediately: we simply came in and copied everything. Where applicable law, company policy, and the circumstances permit, collection from client-owned systems may even occur without advance notice to the employee. Routine automated processes sometimes do much the same thing. That can make for interesting interviews later, but it carries litigation, employment, and privacy risks too. Yes, Data Acquisition can be an exciting part of the EDRM.
Rian Kennedy, one of the EDRM 2.0 project trustees, explained during the public-comment process that modern tools permit Identification, Preservation, Collection, and forms of Processing or indexing to occur with far less separation than before. He also tied this “move left” to earlier visibility and better Early Case Assessment, something I was always trying to get attorneys to do. See Rian Kennedy, EDRM, EDRM 2.0 Public Comment: Three Big Changes.. I do not know Rian, but perhaps he will comment sometime on my Processing question. Knowing the complications of forensics and processing, however, I suspect the answer will require a whole webinar. After listening to many things I do not understand, I’ll probably end up nodding yes and agreeing. My knowledge of processing is very weak. I am no forensic expert like Craig Ball, with whom I often debated at LegalTech on preservation and analysis.
More on Analysis and AI
Analysis with AI assistance can play a role in Data Acquisition before anyone starts traditional review. Imagine an investigation involving 40 potential custodians. Instead of collecting everything first and thinking later, analytical tools can help lawyers examine communication patterns, likely date ranges, concepts, data concentrations, and unusual connections while acquisition is still underway. Human judgment then adjusts the scope. New information feeds back into preservation and collection. again.
The place of AI in Analysis is different from the old model, which treated Analysis as a phase separate from Review. To my eye, that placement fit the then-common assumption, which is still held by some, that the ESI to be reviewed was a giant haystack requiring armies of reviewers to find the needles. Many of us stopped using those armies after 2010. That is old-school, pre-predictive-coding, overpriced e-discovery. Today’s best AI-assisted search processes can often find the needles far more efficiently. We did that every day, even before generative AI, with small review teams and strategic use of AI.
Most of my e-discovery career has involved advocating for a hybrid multimodal approach to AI: humans working with multiple search and analytical methods, especially AI, but not limited to AI. Predictive coding taught us this years ago. Generative AI expands the possibilities dramatically, but the underlying principle remains sound. AI can improve the search; humans still decide what the search means. That human-machine “hybrid” partnership has been a continuing theme of my work.
Which brings us again, in recursive manner, to the most important line in the new EDRM.
Analysis Everywhere
Putting Analysis everywhere, the long blue bar, is a key improvement to the EDRM.
EDRM’s official definition of Analysis is concise: “Assessing ESI/documents throughout the lifecycle using human insight and/or technology.” See EDRM, Current EDRM Model. Now here comes a slight criticism, perhaps just nit-picking, but I think this wording is important. “And/or” is inclusive, but it also permits either human insight or technology to stand alone. I would require the combination: “using human insight and technology.” Also, I would go further and say “using human insight and technology, especially various forms of AI.” Sure, keywords and other technology methods have their place, but in my view, AI is the top of the search pyramid. If nothing else, AI can serve as an independent check on other methods. Many of us did that frequently. It strengthened our confidence that quick searches had not missed something important.
Notice what EDRM did not do. It did not create a “ChatGPT Phase.” Good. It did not build today’s vendor vocabulary into a framework meant to last for years. Better still, it made Analysis continuous and expressly broad enough to include human insight and AI.
The trustees explained that continuous Analysis is intended to encompass “legal judgment, analytics, data science, metrics, AI and emerging technologies” across discovery rather than treating analysis as a silo. See EDRM, Public Comments and Trustee Responses. Here the trustees use the conjunctive “and.” Very good. That follows my long standing multimodal approach.
In the past, many in e-discovery tended to associate analytics with review. It is the box that came right after Review and before Production, as if that were the only time e-discovery lawyers need to think. That was a mistake corrected here. Model 2.0 eliminates Analysis as a separate box; flattens it out, so to speak. Look again at the old and new models together to absorb the changes. While you are at it, notice how the yellow representing volume of ESI and the greenish area representing relevant ESI have changed in multiple ways. That alone would be worthy of a good webinar.
The Human Is Still in the Loop—and Should Be Driving the CAR
My enthusiasm for the new model’s treatment of AI comes with a qualification that my regular readers will expect.
Generative AI remains probabilistic. It can be astonishingly capable and still be wrong. I have repeatedly argued that lawyers should treat AI more like a very talented but untested consulting expert, and not like a super-intelligent oracle. Ask it questions. Challenge its assumptions. Demand sources. Check the answer. Cross-examine the machine before you rely on its work.
That does not weaken the case for AI in discovery. It strengthens it.
EDRM’s phrase “human insight and/or technology” leaves room for tools acting at different levels of autonomy, but high-stakes legal judgments still require accountable humans. Relevance is not merely semantic similarity. Privilege is not simply a pattern in text. Importance to a case may depend on witness credibility, procedural posture, a judge’s ruling, a client’s business priorities, or something the machine has never been told.
The enduring lesson is not man versus machine. It is the power of the combination.
AI finds patterns at a scale humans cannot. Humans understand why some patterns matter. EDRM 2.0’s continuous-analysis band gives us a good place to build that relationship.
Review Becomes an Intelligence Hub
I also like the way the new diagram treats Review as a nexus between decreasing data volume and increasing relevance.
Too often, document review has been treated as an unpleasant expense to be minimized. Hire reviewers. Code documents. Weed out junk. Produce what remains.
That view misses much of the value of technology.
Review is where lawyers often discover the case.
Suppose a reviewer finds three communications suggesting that the key meeting occurred a month earlier than everyone believed. That changes the chronology. The new date points toward another custodian. The custodian identifies another messaging system. Collection expands. AI finds a cluster of related conversations. One message changes how counsel prepares for a deposition.
That is not linear review. That is intelligence moving around a network.
The EDRM diagram now makes this kind of looping behavior explicit. The broad Analysis layer underneath Review is particularly well suited to a human-AI workflow. Lawyers provide feedback. Analytical systems identify more potentially important material. Lawyers assess it. A continuous-learning system incorporates that feedback; with other tools, the lawyers revise the search strategy. New understanding sends the team backward, forward, or sideways.
The technology is finally beginning to resemble the way good litigators have always thought.
AI Defensibility Will Be About Process, Not Incantations
The elevation of Analysis also has practical consequences for defensibility.
A future discovery hearing may involve questions that were unusual just a few years ago. What model did you use? What data did it examine? What instructions did you give it? Did the process change during the matter? Who validated the output? What did you do when the system produced inconsistent results? Did humans review the categories that mattered most?
The correct response cannot be, “The vendor said it has AI.” That will not work very well with most judges or clients. A work-product objection may fare better—but, as usual, it depends on the circumstances.
Lawyers do not need to understand every neural-network weight any more than they need to understand every transistor inside their laptop. But they do need to understand enough about the workflow to explain what they did and why.
I have written repeatedly about the “black box” problem. Modern generative models are trained into enormous structures of statistical relationships rather than programmed line-by-line in the old deterministic sense. That makes traditional explanations difficult. It also makes validation, documentation, testing, and human supervision more important.
EDRM 2.0 does not itself establish an AI evidence protocol, nor should we pretend it does. What it gives us is a conceptual home for the problem: continuous Analysis.
That should affect contracts too.
If a review provider’s statement of work assumes that analytics occurs once at the beginning of a project, the contract may not match the workflow anymore. A serious AI-era engagement may require repeated analysis, revised instructions, additional testing, model or tool changes, and human feedback over the life of the matter.
One recent implementation guide makes essentially this point, recommending that vendor statements of work accommodate continuous model refinement rather than a one-time TAR workflow. See Records Authority, Implementing EDRM 2.0 in Your Discovery Workflow. I would add one lawyerly word: document. Document what you did while you still remember why you did it. New AI can help you with that.
Disposition: The Box Lawyers Prefer to Forget
EDRM 2.0 also adds Disposition as a distinct core phase:, at the far right. EDRM defines it as “A systematic, defensible process to retain, delete, transfer or return data after use.” See EDRM, Current EDRM Model.
Although new, this may be the least glamorous part of the diagram. It may also save clients a lot of money and grief. Lawyers are very good at telling clients not to delete things. We are less famous for telling them when they can start deleting again. A matter ends, but collected data remains on a vendor platform. Copies persist in lawyer repositories. Legal holds stay active because nobody wants to be the person who releases them. Data accumulates. Storage fees continue. Security exposure increases.
The cloud has encouraged a dangerous illusion that keeping information forever is free because we no longer see warehouses full of banker boxes. The warehouse is still there. It merely has better air conditioning and invoices by the gigabyte. Vendors can make good money from storage fees.
There is also a longer-term cybersecurity dimension. Information with continuing value may need to be retained, of course. But retaining obsolete confidential information indefinitely can create risk without creating corresponding value. My recent work on quantum computing has made me especially sensitive to the danger of long-lived encrypted secrets. “Harvest now, decrypt later” is fundamentally a retention question as well as a cryptography question. The safest unnecessary secret is often the one you no longer possess.
EDRM 2.0 wisely closes the lifecycle and ties into the grey bar at the bottom, Information Governance.
Conclusion
The new model will not modernize a discovery program by itself. Changing the picture hanging on the conference-room wall accomplishes nothing—especially if the workflows, contracts, training, technology, and human habits remain frozen in 2020. Organizations still have to do the work: teach, train, and retool. That means studying EDRm 2.0 carefully and watching for EDRM’s forthcoming publications and training on Version 2.0.
It waited until the time was ripe and consensus could be achieved on the new realities of e-discovery; then it substantially redrew the map.
Different people will be impressed by different changes to EDRM 2.0. There are many. For me, the most significant is the placement of continuous human-and-AI Analysis across the lifecycle—the blue band at the bottom. EDRM 2.0 has created a map with enough room for the AI future that is already beginning to arrive. It is coming at all of us fast. As I have been saying for a long time, keeping humans in control is the only way to stay out of the TAR pit. With the new frontier models of generative AI, the pit is becoming far more dangerous, threatening to swallow the world. E-discovery professionals, many of whom have been working with AI for years, can help prevent that by bringing their hard-earned disciplines of testing, validation, supervision, and human judgment to the larger AI debate.
Educational commentary only. Nothing here is legal advice.
Ralph Losey Copyright 2026. All Rights Reserved,
Discover more from e-Discovery Team
Subscribe to get the latest posts sent to your email.
