What Makes AI Defensible in Ediscovery?
Insights from BDO and Everlaw
Here's something interesting: when legal professionals logged into a recent webinar on AI in ediscovery, the opening poll showed most of them are already using or at least experimenting with generative AI in their work. That's a big shift from even two years ago, when most people were still in wait-and-see mode.
The webinar, "Ethical and Practical Strategies for AI in Ediscovery," brought together Daniel Gold, Principal and Forensics Ediscovery Managed Service Leader at BDO USA, and Cal Yeaman, Senior Strategic AI Advisor at Everlaw, for a practical discussion about what actually works and what defensibility really means when AI enters the discovery workflow.
Nobody Voted for Billable Hours (But Everyone’s Thinking About It)
Those same viewer polls revealed something else: defensibility and accuracy concerns are the top barriers to AI adoption, with training and capability gaps close behind. Billable hour incentives? Not a single vote.
Gold didn’t let that slide. "If I cut my doc review time down by 40%, my utilization's down now. I just lost all these hours," he noted, articulating the unspoken tension in many law firms.
But he challenged that assumption head-on. "If you get to the truth faster, if you get the information and you get it better and more robust, and if speed to outcome increases dramatically, clients notice. They’re happier.”
Higher client satisfaction means more work, more referrals, and ultimately, more opportunity. Firms often worry about AI eating into billable hours, but the deeper issue is that they haven’t learned how to quantify their value beyond simple time-tracking.
Where Defensibility Actually Lives
Yeaman was clear: "Defensibility lives in the hybrid."
What does that mean in practice? It means that generative AI doesn’t replace human judgement. It’s essentially amplifying it. And like any amplifier, it makes good processes better and bad processes worse.
"It is 100% a multiplier and amplifier," Yeaman explained. "If you add it as another tool in your toolbox to think critically and approach problems with a certain level of rigor, it's going to multiply what you can do. If you are instead just asking questions and not asking questions of the answers, that's going to multiply the impact of that decision point." Simply put, AI is a force multiplier for competence, not a substitute for it; it can accelerate your insights just as easily as it can automate your mistakes.
The workflow Yeaman described isn’t radically different from what legal teams already do with technology-assisted review (TAR). You craft parameters, test on samples, validate output, iterate, and refine. Critically, the evaluation criteria for performance in these workflows are long-established and court-approved. These standards are often directly transferable to generative AI applications, such as Everlaw’s Coding Suggestions, where the same principles of statistical sampling and validation apply.
The difference with generative AI is that instead of training a model on hundreds of coded documents, attorneys provide detailed prompts describing what they're looking for and why.
"You need to be able to concretely express what it is you're trying to accomplish," Yeaman said. "You need to be able to describe what constitutes the thing you're looking for.” Then you need to verify the performance.
In other words, just like you’d give clear instructions to a junior associate, you give explicit instructions to the AI tool. And just as you would review that associate's work product, you review the AI's output with a critical eye.
Human as the Krino
Gold introduced a framework he calls "Human as the Krino"; a deliberate pushback phrase "human in the loop."
"I am not a fan of that expression," Gold said. "When you say human in the loop, it suggests that as the attorney, you are not actually leading the process.”
Krino, a Greek word meaning to judge, discern, and separate the valid from the invalid, became an acronym for Gold's methodology:
K - Know what the AI did
R - Review the output as a judgment call
I - Interrogate the methodology
N - Normalize against your subject matter expertise
O - Own that determination with your professional license
The last one is critical. "If you are going to leverage AI to help you with your answers, whether it is legal research or ediscovery, you have to own that determination,” Gold said. “That's how you operationalize professional judgment."
This approach directly addresses ethical obligations under ABA rules like 1.1 (competence), 5.3(b) (supervision of non-lawyers), and 11 (certification of submissions). Some states are already amending their interpretations of Rule 1.1 to include understanding AI.
California, for instance, has proposed amendments that explicitly extend the duty of competence to include the benefits and risks of generative AI. Crucially, these changes mandate that a lawyer must independently verify and exercise professional judgment over any AI-generated output before it is used in a client matter; leaving no room for "low-stakes" exceptions.
However, this proposal has faced scrutiny from critics who argue that such a rigid, non-delegable duty of verification may create an onerous burden, potentially stifling the very efficiency that legal teams hope to gain from AI adoption.
What “Good” Looks Like
Gold offered a simple test: "Good means if a judge or regulator asks why you did this and how you know it's going to be reliable and you could answer that clearly, that's good."
He advocates for creating process documentation, essentially audit logs for AI usage. What prompts were used? What was validated? How were decisions made? Think of them as receipts for human judgment, serving the same purpose as privilege logs: demonstrating that the process was sound.
The February 2025 decision in US v. Heppner from the Southern District of New York drove this home. The court wanted to know the lawyer's reasoning, safeguards, and validation with respect to the tools, not how sophisticated the software was.
"Can you explain this to a judge who doesn't care about your tools?" Gold asked. "I think that's what it is."
The Behavior Problem
AI adoption, Gold argued, is fundamentally a behavior problem, not a tools problem. He brought in the COM-B framework from behavioral science which states that behavior alters when capability, opportunity, and motivation are present.
Capability means more than watching a demo. It means knowing how to use the tools defensibly, how to ask good questions, validate outputs, and document decisions. "A Gen AI tool feels risky as opposed to empowering" without proper training, Gold noted.
Opportunity means the organizational culture makes AI use easy and safe. Are there approved tools? Clear guidance on when you can and can’t use them? "If people have to guess whether something is allowed, the default is just going to be, nope. Not going to use it,” Gold said. “That's not resistance. That's rational human behavior."
Motivation is where the billable hour tension lives, but also where the reframing happens. If firms reward the behavior they say they want (client satisfaction, efficient outcomes, quality work), the motivation aligns naturally.
Yeaman shared what this looked like at Orrick, where he worked before joining Everlaw. The firm created creditable time for GenAI experimentation. "There was opportunity to provide insight and thoughts regarding different tools and what worked or didn't work. We had conversations going on about successful use cases in different practices."
That top-down culture of innovation let him handle fast-moving, overlapping cases more efficiently. "The necessity was the mother of invention at Orrick to be able to keep up with the rapid pace of things," he said. "The reward for me was more work, and I got more cases that needed that kind of touch."
Where Does Value Live Now?
If everyone has access to the same AI tools and the same knowledge, what becomes the differentiator?
"Knowledge is no longer the moat," Gold acknowledged.
Yeaman pointed to what remains irreplaceable: legal judgment itself. "The real true core of what it is to be an attorney, that mindset is still key to us and cannot be abrogated to technology ethically. The value you provide to the clients is in your insight and ability to understand and pare away the information to what truly matters."
The emerging role, Yeaman said, is the "forward deployed lawyer": someone who combines legal acumen, technical fluency, and consultative knowledge to tailor solutions to specific client needs.
In this new ecosystem, the technology provides the capability, but the implementation provides the control. Gold suggests that while the platform builds the engine, the role of the consultant, as in the case of BDO, is to help clients navigate the road responsibly.
Success here lies at the intersection of software and the human effort to design workflows and governance structures that actually stick. It’s about ensuring that the technology doesn't just sit on top of old habits, but actively reshapes how a team ensures compliance and adapts to change
"People who combine legal judgment,technical fluency, and leadership–that's really where we are going," Gold said. "It's about how we work together as partners to turn insights into action, action into outcomes, and outcomes into trust."
The bottom line, emphasized throughout the webinar: AI adoption is about equipping lawyers to exercise their judgment more effectively. And it's not about who has the best model. It's about who can operationalize these tools defensibly, with humans firmly in the lead.
