The Goldilocks Matter

Last week Relativity hosted a  lunch and learn about AI in discovery.  An attendee asked our panel why there is still such a gap between what these tools can do and what happens on live matters. The standard answer is a list of worries: hallucinations, validation, defensibility, cost. But that answer mistakes the symptom for the cause. I rarely hear someone say the tool is useless or the risk is unacceptable. I hear something more familiar: we like it, we see the value, but not on this case, and let’s revisit it on the next one.

That is how adoption stalls without anyone meaning to stall it. The decision is never really no; it is “not here,” “not yet,” or “after we finish this matter.” Sometimes that is completely reasonable. A population is already scoped. An ESI protocol is already approved. A review team is already moving. No one wants to reopen the plan two weeks before a deadline just to prove a point about innovation. The problem is what happens next: the deferral becomes the operating model.

What we are really waiting for is the Goldilocks matter: the right team, with the right knowledge, facing the right amount in controversy, on the right data set. The matter has to be large enough for the result to make a difference, but not so visible that nobody wants to be first. It has to be early enough that the workflow is not already locked, but mature enough that the team understands the population, the issues, and the risk tolerance. By the time the next possible matter opens, one of those variables has usually moved.

The industry used to have a rhythm that made those deferrals easier to manage. New products and new ideas arrived on a calendar legal understood: you heard about them quarterly, went back to your team, figured out what was worth piloting, and had a natural window to evaluate. That rhythm gave advisors time to do the job clients actually need us to do: pressure-test the tool, separate demo from reality, and translate the change into something a legal team could responsibly decide to use.

That world is gone. Legal is a deliberate industry, but AI has put product change on something much closer to a news cycle. New capabilities show up before we have had time to put them through their paces, test them on real matters, understand where they are strong and where they break, and help clients decide whether the cost, time, and disruption are worth changing the way they work. The trickle-down effect is real: if we are still learning what changed, our clients are not yet in a position to make a confident adoption decision.

Governance compounds the problem because every meaningful release can restart the clock. A new feature can mean a new security review. A changed model can mean another defensibility discussion. A change in data location, retention, prompts, audit logs, or human oversight can become privacy, security, and outside counsel policy all at once. Nobody in infosec cares that you want to build privilege logs faster, and they should not. Their job is to slow the thing down until the risks are named and mitigated.

That is why the Goldilocks matter is not just hard to find; it is hard to preserve. By the time the right team, right data, and right dispute are lined up, the product may have changed again.

That is why desire alone does not move the work. Outside counsel may want the experience, the in-house team may want the efficiency, and litigation support may be ready to run the workflow. But adoption still has to land before the window closes, on a matter where coordination, risk, timing, and ownership all line up. That is a much narrower target than “we should use AI more.”

Once that matter appears, the ask is narrower than it can sound. We are not asking anyone to abandon the old playbook or become an AI expert overnight. We are asking them to choose a discrete workflow, define the validation methods, agree on what would count as a successful result, and decide who has authority to keep going, stop, or adjust course.

That is still a real operating change, but it is not a mysterious one. The enterprise lawyer wants the matter to look like the last one because the last one was explainable. Counsel knows how to defend search terms, sampling, and reviewer decisions, but needs the same clarity around prompts, outputs, QC, and escalation points. Review teams need the protocol: what to review, what to sample, what to escalate, what to log, and when to stop.

But we are also starting to move past the Goldilocks problem. Not because every variable suddenly lines up, and not because the pace has slowed. We are moving because waiting for the perfect pilot matter has started to look like its own kind of risk. At some point, you stop treating every use case like a pilot and start building the process around the work that is already here.

That is where Schulte v. LinkedIn matters. Magistrate Judge Laurel Beeler treated LinkedIn’s use of Relativity aiR for responsiveness review as technology assisted review and declined to impose special scrutiny simply because the tool used generative AI. That is not a magic approval slip, but it is a lifeline. It gives counsel a familiar TAR frame, which makes it harder to keep saying “not this matter” as the default answer.

The conversations have changed too. A year ago, I spent most of a kickoff explaining what the tool was and how it used a model. Now the questions are narrower and harder: what population should we run, what should humans sample, how much time do we budget for exceptions, who signs off, and how do we measure impact? That is what adoption becoming practical actually sounds like.