Key Takeaway: In just 10 weeks, legal AI has shifted from tool selection to accountability. Epiq expert Jon Kessler reflects on his most recent roundtable and reveals the essential next steps for legal leaders: assign agent owners, validate outputs, strengthen data foundations, and redesign workflows. Acting now will position legal teams to govern agentic AI responsibly and turn experimentation into lasting value.

Not long ago, I reflected on conversations with legal leaders surrounding the use of agentic AI, highlighting that the legal industry was at an inflection point. Discussions had crossed from curiosity to execution and we were beginning to write the playbook together in real time. In just 10 weeks, that landscape has shifted enough to warrant a closer look.

In September, I sat back down with a group of legal, compliance, and eDiscovery leaders from some of the largest institutions in the world. Same table, many of the same people, only a couple of months since the last discussion. After moderating these conversations for years, I can usually anticipate where the conversation will go. This time was different.

In the last roundtable discussion, the most consistent dialogue was about access, and ownership was still fluid. Which tools, which providers, who gets a license, and who gets to say yes. This time around, nobody wanted to talk about access. They wanted to talk about accountability: Who owns the agent? Who validates the output? Whose data is this? Where does it actually live, and can we prove any of it a year from now?

For me, that is the headline. In 10 weeks, this group went from picking tools to governing them. The playbook is finally getting written. It just doesn't look like what anyone expected.

Everyone Has the Tools, Yet Almost Nobody Has Permission To Build

The "access to AI” conversation is over. Everyone at the table has the ability to use an organization-wide AI solution. Many are using legal-specific platforms, some have proprietary internal tools, and all note the AI features quietly showing up inside every product they already own. One leader summarized their relationship with AI tools in four words: we have so many.

But having the tools and being allowed to build with them are two very different things. More than one person described using AI every single day and still being locked out of the building layer. As one put it, "if I had access to those tools, I would start building tomorrow.” That frustration is not about technology. It is about the gap between what technology can do and what the institution has cleared it to do.

That gap produces exactly what you would expect. Someone builds something genuinely useful, brings it to the AI committee, and is told to go back through the approved route. The room was blunt about where that ends up: people are going to figure it out anyway.

Here is what I would say to that: Gatekeeping does not stop the behavior. It just moves the behavior somewhere you cannot see it. Running water will always find its way around the rock. If your governance model depends on people not being resourceful, it is not a governance model.

It’s Too Easy To Build an Agent

That was the line of the day, and it was not a compliment. It has never been easier to build an agent, but that ease is squarely the problem. Citizen development and the creation of personal agents comes with important questions about strategy, governance, and responsible use.

The governance work people described is serious and specific: agent inventories; named agent owners, with a process for transferring ownership when someone leaves; registration requirements for anyone building inside legal; recertification every time an agent changes. One organization has a decision tree in development for the moment agents stop being assistants and start carrying skills and taking action. Another has turned sharing off entirely while it assesses controls, which means every agent in the building is currently private, individual, and almost certainly duplicative.

Then came the comment that reframed the conversation. Two similar agents both doing something useful is not the worst problem to have. Purpose-built tooling for this is still thin, but agent instructions, knowledge, and triggers are text that AI can examine for redundancy and risk. So, the real question is not whether your agent estate is tidy, but what you are optimizing for. Are you running the technology as efficiently as possible? Or, are you encouraging your legal team to identify who is going to thrive in this environment?

Validation

Nobody disagreed on the topic of validation. Regular validation is a requirement, and yet the assumption that someone with an agent that works well is going to keep checking every word is, as one participant put it, not how human nature works. So, if you are not going to review everything, what is an acceptable amount? Nobody had a settled answer. Several are building toward one, teaching their teams how to test agents properly and pushing vendors for real validation tooling. 

That question, more than any tool selection, is what separates a pilot from a program. And the market seems to agree. The day before we met, a leading legal AI platform announced it had acquired Guardrails AI, a company built specifically around agent reliability, simulation and evaluation. Harvey's own framing of why was almost the exact same question our room had just spent an hour on: What will an agent do before we let it get close to real work? When providers start buying the answer to your governance question, that is a good signal you are asking the right one.

The Context Layer Is the Whole Game

If one theme ran through every part of this conversation, it was this one. The number I keep hearing, and heard again last week, is that the majority of what legal departments want to do with AI is blocked by the data, not the model.

Several organizations are responding by moving data to where the AI already is and treating that as an interim step while an internal context layer gets built. The honest estimate for the internal build was years. In the meantime, the operating principle someone offered was refreshingly practical: go wherever you get the best retrieval for the least token spend.

The best reframe of the day was about legal data itself. Most of what we hold is not data. It is documents and knowledge. And the first question that really matters about a document is whether it is still good. That turns out to be far easier to answer than people assume, because provenance and date do most of the work. If it is an opinion from a law firm, you know exactly what it is, and you can treat it as reliable as of the day it was written. It is not about finding the perfect record. It is about structuring things, so you have confidence in what was right at the time, and leaning on recent material over old material.

And the hardest part of that is not technical. It is the calendar. As one leader described it, the whole exercise depends on getting two hours of very senior time to tell you what actually matters to them. If you can get the two hours, the problem is solvable. That is the entire gate. I have watched data projects die for a lot of complicated reasons, and honestly, most of them come down to that.

One more piece of advice from that thread, and I would take it seriously. If you are still on license-based pricing rather than consumption, use this window. Push as much through as you can and turn it into structured data now, because the expectation around the table is that everything moves to consumption within a year or two. That is a real strategy, and the window is closing.

The People Making This Work Are Not Who You Would Expect

Nearly every organization at the table has some version of the same structure. Champions, catalysts, ambassadors, a center of excellence. Regular calls where people show what they built. Prompt libraries and agent libraries published internally. In one case, a friendly competition that has people building more than they otherwise would.

So, I asked what the successful ones have in common. Are they technologists? The answer came back immediately, and it was no. None of them. They are just very curious. They are the legal professionals in the weeds doing the work, and what makes them effective is the order they do things in: build the process first, then build the technology, then test it to death. Some of them are not the most senior people in the department, and this gives them a place to shine. And because they are peers, the message lands differently. It is not leadership saying you could build this. It is the person sitting next to you, saying “look what I did.”

The training conversation has matured too, and this is the part I would steal. It started where everyone starts, with prompt training and optional self-study. Now it is mandatory, and the curriculum has moved well past mechanics. Two topics dominated. First, what an agent actually is, because the word gets thrown around so freely that it is starting to mean nothing. Second, how to review work product created with AI. Not just your own output, but as a senior reviewing a junior's work, and as in-house counsel asking outside counsel what they actually did.

There was also a problem I had not heard before. The organizations furthest ahead are finding that new joiners arrive feeling behind. One leader is standing up a joiner program they never planned for, because lateral hires walk in without the fluency the department now takes for granted. That is a genuinely new kind of onboarding problem, and if you are moving fast, it is coming for you.

Outside Counsel Is Where the Economics Get Real

This was the most commercially interesting stretch of the day. One participant described simply committing to a reduction in outside counsel spend and treating how to get there as the problem to solve. The math is not demanding. At a large institution, reducing outside spend by a small percentage pays for the tools several times over.

What was surprising were the wins that nobody was looking for. Litigation and alternative provider work were the obvious places to start, and both are already reasonably efficient. The real money turned up in places legal operations has never touched. Doing one thing differently in an unexamined corner produced a saving of hundreds of thousands of dollars that nobody had forecast. The advice: go find the functions nobody has ever sat down with.

As for the firms, the reported blocker to fee reductions is that they do not yet have enough data to show the savings. I would gently push back on that, and so would the market. The Financial Times reported in early September that Goldman Sachs, Morgan Stanley and Citigroup have pressed their outside firms on exactly this point. Goldman has asked firms how much AI is saving them and expects to share in the benefit. Citigroup has amended its bidding process to require firms to explain those savings, with its global head of legal saying that if the hours on a matter come down because of AI, the expectation is that cost comes down significantly per transaction. Morgan Stanley expects most of its external legal work to move to competitive bidding and alternative arrangements such as fixed fees by the end of this year, and its general counsel described the traditional law firm compensation model as extraordinarily unstable. If you are in-house and you have been waiting for cover to have this conversation, you now have it.

The related claim, that clients will not let law firms use AI, got no sympathy in this room at all. The view was that it is usually one cautious instruction being over-read, not a real client position. Meanwhile the group is watching the AI-native firms closely, and watching serious people leave established firms to go build them. Changing the model from inside a partnership means changing incentives and billing at the same time, which is genuinely hard. The startups do not carry that weight. Whether they can attract enough talent, and for which kinds of work, is the open question.

One observation I raised, which the table broadly agreed with, is one I have not stopped thinking about: knowledge management is a discipline law firms have practiced for decades, and corporate legal departments largely have not. In a world where retrieval quality determines AI quality, that is a structural advantage sitting on the firm side of the table. If you are in-house, that should bother you enough to do something about it.

The Audit Is Coming, and the Framework Does Not Fit

Only one person at the table had been through an AI audit, and the account was sobering. The audit teams are staffed with technologists applying model risk frameworks built for traditional AI models, and the questions simply do not map. They want statistical evidence of model performance. What legal has is thin feedback data, because people do not leave feedback, and no way to show that a prompt or an agent was validated a month after the fact.

Someone else admitted, to laughter, that they had put theirs off. I understand the instinct. But the direction of travel is obvious, and the teams that come through this well will be the ones who started capturing validation evidence before anybody asked for it. Start now, while it is cheap.

The Basics Still Break

For all the sophistication in the room, some of the most striking moments were about ordinary things going wrong. One team tried to build a precise chronology and found the tool would not reliably pull the correct parent dates off email. They offered the obvious fix, a metadata export, and the platform could not consume it. The provider agreed to build the capability. Weeks later it still had not shipped. A competing platform would not entertain the request at all. As the participant said, “I thought that was basic.”

There is a pattern worth naming here, and it is the same one I flagged last time. These tools are consistently better at analysis than at ingestion. They reason beautifully over content and then stumble on the plumbing that eDiscovery teams solved 15 years ago. That is not a knock; it is a buying signal: ask about ingestion in the demo, not just reasoning. Features are now shipping faster than institutions can risk assessing them, which is why several organizations are deliberately leaving capabilities switched off. Not because the capability is bad, but because operationalizing it responsibly takes longer than the release cycle allows.

The biggest unclaimed opportunity raised all day sits right here, and it is the one I would chase. At a large bank, for every 100 matters that open in a year, 90 might never reach discovery. Data still gets collected for investigations, HR matters, and small regulatory inquiries, and it lands in tools that were never designed for the non-lawyers who need to ask questions of it. Give those teams a defensible review environment with AI over the top, respecting the permissions you already have, and you have solved a real problem that everyone recognizes and nobody owns. That is not a someday idea. That is right now.

What I Took Away

10 weeks ago, this group was comparing tools. Two weeks ago, they were building agent registries, negotiating outside counsel reductions, rebuilding knowledge management from scratch, and arguing about what an acceptable validation rate looks like. One participant said the volume of these conversations has gone up at least threefold in a year, and that the pace is awesome and scary and fun all at the same time. That is exactly right, and I would not change a word of it.

The most important comment of the day was also the quietest. What lawyers do today, which is a lot of doing the work themselves, is not what lawyers will do in three years. The work moves toward building the guardrails, designing the process, and checking the output. Departments that are only teaching fluency are solving last year's problem. The ones teaching people to redesign work from the ground up are solving the right one.
Despite how much has shifted in 10 weeks, the mandate is still clear. The playbook is maturing and being written in public by people who are figuring it out one hard question at a time. There is still no instruction manual, and everything will change again next week. But the departments that pull ahead will not be the ones with the most tools. They will be the ones willing to decide what good looks like, write it down, and then go get the two hours of senior time it takes to make it real.

My thanks to everyone who joined and spoke candidly. We are planning to regroup before the end of the year. 

The playbook is still being written. If you’re tackling these questions too, I’d welcome the chance to exchange perspectives.


Jon Kessler, Vice President and General Manager, AI Advisory, Data, and Governance
Jon leads the Epiq Global AI Governance and AI Innovation Advisory business, partnering with Fortune 500 organizations to transform legal and compliance through enterprise AI. A recognized leader in legal AI and AI governance, he focuses on helping organizations responsibly adopt AI while improving decision-making, governance, and business outcomes.