Key Takeaway: For organizations harnessing legal AI, the biggest risk lies in treating launch as the finish line. In this continuation of a series on roundtable reflections from conversations with legal leaders, Epiq expert Jon Kessler shares an urgent playbook revealed during a recent London roundtable. Legal teams hoping to capitalize on what AI has to offer must assign clear owners, validate outputs, and update source material, while tracking business outcomes and protecting portability before scaling.

Building a legal AI agent is getting easier. Keeping it reliable and useful is ongoing work. That's what remains top of mind for me after two London roundtable discussions this Fall. How do we keep a tool working well as policies change, teams evolve, and new uses emerge? 

Participants shared their experiences and the questions they're working through openly, creating space for meaningful reflection and a chance to strategize. Both sessions were held under the Chatham House Rule, so the following are general examples and nothing is attributed to any individual or organization.

The Demo Isn't the Finish Line

One participant described horizon scanning from curated, authoritative external sources, starting with a narrow geography and using SharePoint filters for jurisdiction and topic. Their team was working through ownership and sharing alongside the build. The combination of a defined scope and practical questions about the use of the tool was intriguing.

Another participant described a non-disclosure agreement (NDA) use case. Using a legal-specific AI tool, the team was able to develop the playbook. Copilot and Teams give the business an app to interact with it. The playbook rules set the boundaries, with red flags progressing to legal. The participant shared about testing against several years of previous NDAs and legal outcomes. What stood out to me was the check against work the team already understood, alongside a well-defined task and a familiar way in. 

The distinction between a routine NDA and an M&A NDA also matters. Different contexts call for different approaches. The discussion emphasizes the importance of focusing on the changes that matter, rather than generating more redlines. For an agent, that means understanding which changes are useful, when legal needs to step in, and how to help the business move forward.

That idea of a "familiar way in” also came up in the discussion of a simpler legal front door. Colleagues could describe a need in natural language or email rather than learn a complex form. An agent could ask for missing information, route a request to legal, HR, or procurement, point to approved guidance, or create a matter where appropriate. Teams are planning these workflows, with approved playbooks and clear escalation behind them. I was glad to see that leaders are focusing on making the first step easier.

Teams were taking different routes on governance: agents switched off by default, training before building, sandboxes, and checks for duplication. That raised questions about visibility and continuity. How can teams see which agents are running and how they've been tested? How can knowledge of policies, parameters, and maintenance carry across a change of owner? Easier building makes those questions worth addressing early. 

Finding the Advice Is Only Half the Job

This is an area I get genuinely excited about: bringing previous advice, precedent, and outcomes into the moment a decision is being made. Advice you’ve already paid for is a valuable starting point. There's a lot of opportunity to reuse that knowledge, with the context and professional review needed to establish whether it fits the question we're answering today.

Participants identified existing guidance on business SharePoint sites that need updating or clearer context. In some documents, jurisdiction was implicit because the humans using them already understood it. An AI tool doesn't automatically inherit that understanding. Making the context explicit is part of preparing existing knowledge for a new way of using it.

One participant described an AI answer to a legal question that relied on a repealed rule, which counsel identified. That was a useful example of professional verification doing its job. It also showed why a citation alone isn't enough. A real source may no longer reflect the law or apply to the question at hand. Checking currency and applicability remains an integral part of using these tools.

My takeaway is that an owner, a review date, and a clear scope, including jurisdiction, make guidance easier to reuse with confidence. A way to retire superseded material also helps. I want us to get more value from what we already know without losing the professional judgment that tells us whether it applies today. 

From Adoption to Outcomes

Participants were working with Copilot, legal-specific tools, and proprietary in-house systems, and exploring which tools best suited different tasks. I keep coming back to the same question: what tool do we use for what? A matrix of tasks, tools, models, and costs enables teams to make that choice. Questions about costs and caps also highlight the value of guidance on usage costs alongside prompting skills.

One participant described high Copilot usage and was wondering how to understand the outcomes it was producing. Another shared a contracts process where throughput and redirected capacity were easier to see. I took those as two ways into the measurement question: who's using the tool, and what changes in the work when they do? Both tell us something about progress. 

Several people highlighted the value of a baseline before rollout, so teams have something meaningful to compare results against. Are contracts moving faster? Is quality improving? Can the team take on work it previously had less capacity for? Is revenue moving through sooner? Time saved creates capacity, rather than simply a smaller headcount. Legal has a valuable perspective to bring to that business conversation.

Participants disagreed on what AI should change about outside counsel fees. Some wanted efficiency reflected in the bill. Others were buying judgment, accountability, and speed, not simply hours. I think that's exactly the conversation to have. If AI changes how work gets done, what are we paying for, what should change in the price, and who's still accountable for the answer?

Make Time for the Work Around the Tool

The build-or-buy discussion had a middle ground: try an interim agent while platform evaluations and requests for information (RFIs) and requests for proposals (RFPs) continue. A scoped pilot establishes what people need and identifies whether they'll use it before a larger commitment. I don't see that as a verdict for building or buying. Either route needs internal expertise, testing, change management, and ongoing maintenance. 

Whichever route a team takes, implementation isn't zero effort. Process design, adoption, and adjustments as the business changes still need attention. Faster builds make experimentation easier; we need time to make the result reliable in everyday work. 

For smaller teams, that may mean shared or fractional support, or making space for someone doing it alongside a day job. It doesn't automatically mean another hire. What matters to me is clear responsibility and enough time and resources, no matter what someone’s title is. 

The choice also needs to account for what happens if we move on. A participant's experience of a platform that didn't support exporting certain workflows prompted a discussion about portability. What comes with us, what stays behind, and what gets deleted? It's worth exploring data return and deletion, retention, subprocessors, training assurances, and the portability of playbooks and workflows before we need an exit. 

Keep Learning Alongside the Technology

People described internal workshops, champion demonstrations, and a website with examples. Informal monthly conversations across teams created space for wins and frustrations. Pairing an experienced colleague with someone seeking help offered a more personal approach. What struck me was how practical these ideas were: show what you've tried, make it easy to ask a question, and recognize that people start from different places.

We also talked about how junior lawyers develop judgment as automation changes some of the repetitions through which they learn. Apprenticeships and in-house training were among the approaches discussed. One suggestion was to let junior colleagues lead the conversation about how they're learning. Another example showed AI helping turn personal notes into quizzes and flashcards, with a senior colleague still providing context and guidance. The invitation was to ask juniors rather than assume we know. Faster answers still need critical thinking, verification, and discussion of why an answer fits.

Keeping Useful Things Useful

What encouraged me most in London was the progress underway, from a scoped NDA pilot to ideas for reusing advice and measuring outcomes. There isn't one approach for every team. Keeping these tools useful means giving people time and resources to test and adapt, with clear ownership and current content. I'm optimistic about what we're building together, and grateful to everyone for their time and candor.

What are you learning as AI becomes part of everyday work? I’d love to connect and hear your thoughts.
 

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.