The Smartest Lawyers Have the Messiest Data (and AI Can’t Read a Hunch)

The best lawyers I know are almost always right… and fast.

Someone hands them a deal, a dispute, a policy question with a deadline attached, and very quickly, they’ve read the risk correctly. That read isn’t a guess; it’s years of pattern recognition compressed into a gut call, and it works because the environment taught them what to look for.

They’ve seen this deal structure before. They’ve seen how this kind of dispute usually resolves. Cognitive scientists have a name for this: recognition-primed decision making, the same mechanism studied in fire commanders and ER physicians making high-stakes calls under time pressure. Under those conditions, expert instinct isn’t the risky part of the decision. It’s usually the most reliable part.

The hours that come after the instinct

The initial read is fast, but proving it isn’t. So the smartest lawyer in the room spends the next two hours or two days reconstructing the evidence for a conclusion they reached in the first five minutes. Pulling precedent. Rebuilding the timeline. Digging through a sent folder, a legal pad, three OneNote tabs, and a memory that’s already moved on to the next matter. Not because the instinct was wrong, but because nothing was captured on the way to it.

Here’s the part that should make everyone a little uncomfortable: from the inside, a correct instinct and a confident guess feel exactly the same. The lawyer can’t always tell the difference in the moment (nobody can!). The only way to tell a real pattern-read from an overconfident one is to check it against what actually happened, across enough matters to see whether the pattern held. That check requires a record. Without one, you have an unverifiable decision rather than just an undocumented decision. And unverifiable looks exactly like unreliable to everyone outside that lawyer’s head, including, eventually, the lawyer.

I wrote recently about GC loneliness, the disconnect between what legal knows and what the rest of the business can see. This is the same gap, one level down. The organization doesn’t trust legal’s data because legal was never asked to produce it in a usable form. It turns out legal doesn’t trust its own instincts either, for the identical reason. The record of how a judgment call got made rarely exists anywhere durable. It exists in someone’s head, un-testable, even by them.

The AI conversation in legal is aimed at the wrong layer

Every conversation I’m in right now is about which tool, which model, which copilot. Sheila Dusseau said something on a recent Legal Maxxing podcast episode that’s stuck with me since: the real question isn’t which AI is best, it’s where human judgment is being wasted on repetitive work that shouldn’t require it in the first place. I’d push that one step further. AI doesn’t just fail to replace judgment. It can’t even see judgment that was never written down. You can’t amplify a signal that was never captured, and you can’t validate a pattern that was never checked. Every AI tool in legal is being asked to compress a record that, in most departments, doesn’t yet exist in a form worth compressing.

That’s a readiness gap. And readiness is solvable, but only if you treat it as an operational build, not a data-hygiene chore nobody owns.

ELM is where this becomes concrete

Enterprise legal management is usually described as the system of record for matters, contracts, and spend. That undersells what it does when it’s built right. A lawyer’s instinct gets trustworthy the same way any expert’s does: they see a pattern, and eventually find out whether they were right, and that feedback sharpens the next read. ELM, done well, is what lets that same loop run at the department level instead of staying locked in one person’s memory not just logging the call after the fact, but showing whether the call was good, so the next one draws on evidence instead of recollection.

Think about what ELM visibility actually answers, when the underlying data is trustworthy: Is the right person (internal or outside counsel) doing the right work at the right cost? Not a guess based on who seems busy. An actual read on workload, matter complexity, and spend against outcome.

Where is risk clustering, and did it cluster here before? Pattern visibility across matters is the thing that turns “I have a bad feeling about this vendor contract” into “here are the four prior disputes with this counterparty type, and here’s what they cost us”  the difference between a hunch and a hunch you can check.

Is outside counsel spend buying leverage or just buying hours? You cannot answer that from an invoice. You can answer it from a system that tracks matter outcomes against cost over time.

None of that is a legal-department-only question. Once that data is trustworthy, it feeds decisions well outside the legal function: procurement teams negotiating from real contract-term patterns instead of anecdote, product and service teams spotting where customer contract friction is actually a signal about market fit, sourcing decisions that account for legal risk exposure instead of discovering it after the fact. Legal becomes the department other functions pull insight from, instead of the department that gets looped in once something’s already gone wrong. That’s the difference between legal as a cost center and legal as a strategic voice and it’s built on the same data infrastructure, not a separate PR effort.

So how do you actually build It?

This is the same readiness sequence I use with legal departments evaluating any operational shift, and it applies directly here.

• Assess: Look at where judgment calls actually happen today, honestly, and where the evidence for them currently lives. Legal pad. Email thread. Someone’s memory. You’re not auditing performance, you’re mapping the gap between the instinct and the record.

• Plan: Define the data minimum. Not a sophisticated analytics build the smallest set of fields that make a matter, a contract, or a vendor relationship useful later: category, cost, timeline, outcome, risk classification. Trustworthy and boring beats comprehensive and abandoned.

• Execute: Pick one workflow and make it visible first, usually intake, or outside counsel spend, because that’s where the volume and the chaos both live. Prove the model on one process before asking anyone to trust it everywhere.

• Activate: Name the shift out loud. This isn’t instrumentation for oversight, no one believes that and it’s the one that kills adoption fastest. It’s the feedback loop that lets a GC’s judgment get verified instead of just trusted on faith, and that’s what lets it scale past the room they’re standing in. Say that plainly, to your team and to the business. People adopt a system faster when they understand what it protects, not just what it tracks.

The instinct was never the problem. What was missing was a way to tell the good instinct from the confident one for the lawyer, and for everyone relying on their read. Build that, and the two days it takes to reconstruct the evidence stop being the cost of good judgment. They become proof you already had it.