Key Takeaway: If there’s one thing recent roundtables on agentic AI with legal leaders have made clear, it’s that we have crossed a threshold from curiosity to execution, and hesitation is the biggest risk. The message is clear: discipline separates progress from noise. Pair governance with experimentation, keep your people in the loop, and focus technology on the work that truly matters to deliver repeatable, scalable outcomes.

I’ve spent the last 30 years in legal technology, mostly in forensics, eDiscovery, data governance, AI, and legal operations. I’ve learned to pay attention when a room of practitioners stops debating whether technology matters and starts arguing about how to deploy it. That’s exactly where we are with agentic AI right now.

Last week, I had the privilege of moderating a roundtable of senior legal leaders from some of the world’s largest financial institutions and corporations; the people running legal, compliance, and operations at real scale. They came ready to compare notes on what is working, what isn’t, and where the next round of value will come from. What follows are my reflections on the themes that emerged. I’ve kept the specifics high-level out of respect for the candor in the room, but the insights were too valuable not to share.

Three priorities set the tone before we even got started: a strong appetite for eliminating or reducing first-level eDiscovery review, a clear focus on agentic solutions and workflow automation, and an early emphasis on legal intake, contracts, and efficiency gains for complex tasks. These threads ran through everything that followed.

Everyone Is Experimenting, and Everyone Has Too Many Tools

The conversation opened with a simple survey of what’s in use, and the honest answer was: everything. Enterprise assistants like Copilot, ChatGPT, and Claude sit alongside legal-specific platforms like Harvey, CoCounsel, Legora, and Clio, and several teams are building proprietary tools of their own. The common thread wasn’t a shortage of options; it was the opposite. Most organizations now have multiple AI tools available at once, and the hard question has quietly shifted from “which tool should we buy?” to “which tool do we use for this task?”

A pattern emerged. Teams reach for legal-specific tools for substantive legal work. Platforms like Legora or Harvey are optimized for the terms and concepts lawyers use, while enterprise tools like Copilot are often better suited for broader, general-purpose tasks. But it isn’t that clean. In some cases, the enterprise tools outperform the legal-specific ones; in others, the reverse is true. Legal-first AI platforms often perform best where legal nuance and defensibility matter, yet they still show real gaps in integrating with the enterprise systems where legal work supporting the overall business lives. The candid takeaway is that the landscape is scattered, and the “right” choice changes task by task.

Governance, Access, Cost, and Ownership are Becoming the Next Bottlenecks

Two structural concerns surfaced again and again. First, in many organizations, IT has become the gatekeeper for agents, and the group was frank that this model may not scale. Second, governance is struggling to keep pace with how quickly legal and the business are putting these tools to work. Looming over all of this is an anticipated shift from license-based to consumption-based pricing, which several leaders flagged as the next real cost-management challenge.

Who owns the agents?

This was one of the liveliest exchanges. As teams start building custom workflows and applications, the question of who should own agent development is unresolved. Some organizations favor a centralized, IT-led model. Others were firmly in the camp that legal and business teams should lead, because they’re closest to the work. There’s no consensus yet, and I suspect the answer will look different in every organization. But this is also the moment to put some discipline around the lifecycle of agents, before experimentation turns into agent sprawl, orphaned workflows, or agents that probably should have stayed as prompts.

Knowing What’s Under the Hood

Those governance questions quickly led to a related issue I care a lot about: understanding what’s happening under the hood. As the discussion turned to how these leaders choose tools and providers, the difference between a semantic model and an enterprise AI layer, along with the differences in how platforms retrieve and reason over data, wasn’t academic. It determines whether you can trust the output. There was real frustration with the lack of transparency in some integrations, particularly with limited visibility into legal-research connectors.

What these leaders want isn’t a black box; it’s a partner. The strongest selection criterion in the room wasn’t a feature list. It was whether a supplier behaves like a true partner, willing to customize workflows and evolve as the client’s needs change. That’s the bar, and it’s the right one.

Data Quality and Access Are the Whole Game

The conversation deepened, as these discussions always do, into data. Retrieval-augmented generation (RAG) varies enormously from platform to platform, and its value depends almost entirely on the quality and structure of the underlying data and on how documents are processed. “Chunking” a contract is a different problem than chunking an email thread, and it shows.

So, teams are deliberate. Rather than opening broad integrations, many are restricting AI to approved repositories such as SharePoint sites, shared mailboxes, and vaults, and they are testing in controlled environments before opening the doors. The “vault” model on some vertical legal AI platforms, including practical repository-size limits, came up as one example. Many share the aspiration to combine an enterprise-wide semantic layer with a legal-specific one, so the system understands both the business and the law.

The frustrations were just as instructive. Model context protocol (MCP), along with other connector strategies used by curated data providers and document management systems, is promising. But in many legal environments, these connections still do not unlock the full value of AI. I often think of MCP as the USB-C of AI: a common connection point that should make it easier for AI tools to plug into the systems where the work and data already live. But as with any connection standard, having the port does not mean you get full access to everything behind the system. In legal, that distinction matters.

Too many connectors provide constrained access to the underlying system, operating more like a permissioned search layer than a true workflow or reasoning layer. That means an external AI tool may be able to find a document, run a limited query, or retrieve selected metadata, but still cannot fully analyze matter context, preserve attribution, take action inside the source system, or reason across repositories in the way lawyers need.

Some of this is prudent governance, especially given the risks of over-permissioned AI access to sensitive client, HR, billing, and litigation data. Some of it may reflect a commercial reality: legal platforms have strong incentives to keep users inside their own AI experiences rather than expose data and functionality to another provider’s model and risk their tool being relegated to a repository that could be replaced. The result is a growing gap between the promise of open connectors and the practical reality of partial integrations. 

A practical note that resonated with me is that closed ecosystems limit what you can build. Much of the industry conversation is about connectors, including MCP and other approaches, which leads to a simpler question: Can the tool reach your data, and is that data useful once it does? Connectivity is the means. Accessibility is the point.

The Value Is in the Use Cases

When the group turned to where they’re seeing returns, three use cases rose to the top:

• Legal Intake as the “Front Door”: There’s significant interest in agentic intake workflows that triage and route legal requests from the business as they arrive.
• Contracts Analysis: AI is delivering more value and handling greater volumes across simple and medium-complexity contracts. Highly complex contracts are where the payoff from human judgment is greatest. Expect this to evolve quickly as horizontal AI (OpenAI, Anthropic, etc.) focuses on contracts-related skills.
• Discovery and Investigations: With a clear focus on early case assessment (ECA) and fact development, the room saw this as one of the biggest opportunities for cost reduction.

The through-line is that the highest-value applications aren’t about volume. They’re about compressing the judgment-heavy work that previously consumed a significant amount of time.

What the Tools Still Can’t Do

I appreciated how honest the group was about limits. Today’s tools still struggle with email-heavy datasets and with understanding email chains and threading. Investigation results have been mixed and, tellingly, several of the poor experiences tracked back to an incorrect setup or approach rather than the technology itself. RAG, in particular, can be a poor fit for open-ended “did this ever happen?” questions, or any scenario that demands complete coverage of a dataset. And there are hard platform constraints, repository size caps, and tools that are far better at analysis than at ingestion.

The encouraging part is that none of this dampened enthusiasm. If anything, it sharpened it. Knowing what a tool can’t do is exactly what lets you deploy it where it can.

The Return on Investment Conversation Is Maturing

Time savings remain the headline metric, and the numbers that participants reported were significant: reductions of 30 – 70% on certain tasks. But the more sophisticated framing was about how you use the reclaimed time: reallocating people to higher-value work and reducing reliance on outside counsel. Fact development came up again as a high-impact target, where even incremental improvement can translate into meaningful savings given the scale involved.

And there was a point of near-universal agreement that’s easy to miss: workloads are increasing, and teams are overextended. Nobody in that room was talking about AI as a way to do less. They were talking about it as a way to keep up and augment capacity, rather than shrink demand.

Digital Workers and the Human in the Loop

Inevitably, we got to the harder questions. There are real concerns about job displacement, but the prevailing expectation was that AI will reduce hours and shift the focus toward higher-level work rather than eliminate roles outright. The most thought-provoking thread was the emerging idea of “digital workers”: when an agent starts to function like an employee, how should it be governed and classified? We don’t have the answer yet, but the fact that serious legal leaders are asking tells you where this is heading.

The room was unanimous on one point: keep a human in the loop, especially for legal decision-making. The enthusiasm for automation is real, and so is the discipline about where judgment has to stay, with the legal team.

Where This Leaves Us

The roundtable discussion adds to the several hundred hours of rich conversation I’ve had with senior legal leadership. If I had to condense it into a sentence, it boils down to this: adoption is rapid and experimental, not yet standardized, and the leaders getting the most out of it are the ones being curious and deliberate. They’re prioritizing high-ROI use cases, building controlled data environments, and investing in workflow automation and agent development. The barriers they’re wrestling with are consistent: data quality and access, tool transparency, and governance structure. None of them are unsolvable.

The direction is unmistakable. We’re moving from tools to workflows, experiments to agents, and curiosity to capability. The organizations that win won’t be the ones with the most licenses. They’ll be the ones that pair governance with experimentation, keep their people in the loop, and point this technology squarely at the work that matters most.
My thanks to everyone who shared so openly. Conversations like this are how our industry writes the playbook together, and we’re writing it in real time.

Learn more about Epiq AI Governance and AI Innovation.



Jon Kessler, Vice President and General Manager, AI Governance and AI Innovation
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.