Legora voices

Legal was the top AI user in the company. I couldn't prove it was working.

Legal was the top AI user in the company. I couldn't prove it was working.

Category

Legora voices

published

author

Mark Allen

At one of my last in-house roles, a meeting appeared on my calendar out of the blue. Our VP of AI engineering led it, and the question was simple: the CEO wanted to know how every org in the company was benefiting from AI. Using it was no longer enough. He wanted impact, team by team, and he wanted it quarterly. Engineering wanted it monthly.

Nobody had to tell them legal was using it. They had pulled usage data by department across the whole company, and legal came out on top.

Then they asked what it had actually done for us, and I did not have a good answer. “It helps us mitigate risk for the company,” I said. “What does that mean?” they asked.

Not having a great answer at the time felt like a personal failure. It turns out to be closer to the industry's default setting. Axiom found that 83% of in-house teams cannot measure whether their AI spending is working, and that only 7% have moved past piloting into actually using, optimizing and measuring what they bought.

The framework nobody asked me to build

I had seen the question coming a year earlier. When we made the case to give the whole legal department licenses to a general-purpose frontier AI assistant (the one you are probably thinking of), my GC never asked me to justify it. I built the justification anyway, because I wanted something on paper if the question ever came back around. Three categories: risk mitigation, cost savings, time savings. Every person who wanted access submitted their use cases against it.

But the question was back. And the framework did not save me.

The answer I gave in that room, it helps us mitigate risk, was category one of my own framework read straight back to them. The categories were right. What I never built underneath them were the baseline metrics. We had not measured where the work stood before the rollout, and we had never agreed with finance on how a risk rating converts into a dollar figure. So I could name what we were doing and I could not measure it.

Part of the problem was how the original business case got built. People estimated their own savings, because there was nothing else to estimate from. Tool consolidation was part of the case too. Someone would look at a the compliance and audit platform we paid six figures a year for, expect that AI could handle most of what it did, and conclude we could probably drop it. Sometimes that turned out to be true, and we did retire a few vendors. But "probably" is doing a lot of work in that sentence, and a year later, in front of engineering and finance, that business case does not hold.

This was not our first attempt, either. An earlier AI pilot went out to about thirty people with an instruction I will never give again: apply this to your day-to-day work and tell us whether it delivers impactful efficiency gains. Impactful efficiency gains is not a measurable thing. People came back and said it worked here, it missed there, the output was thinner than they expected. Out of those thirty people, only a few had use cases that were genuinely ripe. We did not move forward.

Adoption stops counting

Here is what took me too long to accept. High adoption was not evidence that the strategy was working. It was evidence that people liked the tool.

Adoption is the right metric while you are experimenting. You want to know whether anyone will touch the thing. It stops being sufficient the moment the spend gets big enough for someone outside your department to notice, and AI spend is getting there fast. Nine in ten people logging in tells a CFO nothing, in the same way that telling an engineer a risk dropped from a level five to a level four tells them nothing.

The gap shows up across the market. ACC and Everlaw found that 52% of in-house counsel now actively use generative AI, more than double the 23% reported in 2024, yet 59% say they haven't seen noticeable savings from their outside counsel's use of it. Usage climbed. Savings did not follow it.

That second one is a translation failure, and it is fixable. Legal measures in a language the rest of the business does not speak. We would say we had reduced a risk rating. We would say we could review thirty percent more documents than before. Both true, neither denominated in dollars or hours, so neither one lands.

The cost of that is measurable. Thomson Reuters found that 86% of GCs consider legal a significant contributor to business success. Among C-suite executives, 17% agree, and 42% say legal contributes little or nothing at all. A gap that wide is not a performance problem. It is a reporting problem, and legal owns it.

So do the translation yourself. Risk mitigation becomes value protected: this control keeps the company clear of a category of regulatory exposure worth many millions of dollars. Time savings becomes headcount not hired: the department absorbed a year of business growth without the dozens of FTEs the old model would have needed. Cost savings was always the easy one, since it already arrives in the right units.

I went and did that work. It helped, and most departments skip it entirely.

If you're building that case now, start before you roll anything out. Set a baseline for the work you expect AI to touch. Pick one or two workflows and measure them in hours and dollars, not logins. And agree on the math with finance up front, so the number holds up when someone outside legal asks for it. Legora's in-house ROI report is a good place to see how other legal departments are framing it.

But underneath the translation there was less than I expected, and the reason took me longer to admit.

We changed the tooling and nothing else

We had given everyone a license. That was the entire intervention. Everyone got access to a very good tool, then went back to the same job, in the same order, with the same handoffs and the same approvals. The work still started and ended with a person producing it. AI sat beside the work instead of inside it.

And the work was already split across systems that didn't talk to each other: one for intake, another for contracts, another for matters and spend. Each one decided what the work could look like. AI became one more box on the shelf.

There was no number to find because nothing structural had happened.

What eventually changed my thinking came from outside legal. A VP of product on the business side sent me a blog post written by a solo practitioner. One lawyer, no budget, no team. He had gone through everything he did in a typical week, built a skill for each repeatable piece, and set those skills to producing first drafts. His day was no longer drafting. His day was reviewing what his own system produced and deciding what was good enough to go out.

We had more money, more people and a better tool than he did. He had more to show for it.

Yes, it is easier alone. He had nobody to convince and no careers to disrupt. But the thing that made his gains legible was not his size. His role had changed. He had stopped being a lawyer who produced work and become a lawyer who owned the system that produced it.

That is the line between AI-assisted and AI-native, and most legal departments are on the wrong side of it. AI-assisted means people do their existing jobs faster. AI-native means the jobs are different.

Different in specific ways

Senior attorneys stop producing and start owning skills. They become responsible for the context, the source material and the logic that make an output trustworthy, whether or not they build it themselves.

Attorneys and paralegals move from drafting to reviewing. That sounds like a demotion and isn't. Deciding whether something is right is harder than producing a first version of it.

Compliance stops checking outputs one at a time and starts owning governance. If everything the department produces flows through AI, the controls have to sit upstream of the output.

Legal ops stops doing manual operational work and starts designing workflows, for itself and for the rest of the department. Ops sits closest to the tooling and will always have an easier time building than attorneys will.

The hard middle part

None of that is a training problem, which is what most departments turn it into.

A few months before I left, I presented our AI strategy twice. The first room was the C-suite. They loved it. Immediate buy-in, go forth and conquer. The second room held the legal leaders who would actually have to live it, and it was a completely different meeting: real skepticism, and good reasons for it. We were asking people to rethink roles they had spent entire careers building.

The gap between a C-suite that says yes to a deck and a team that has to change how it works is the hard middle part. It is where pilots quietly die, and it is almost never about the technology.

The only place I watched a role genuinely change was legal operations, and only because it was my own team. I had opened a req for an AI operations manager, which is the normal move: spot the gap, hire for it, hand it over. Then it became clear the gap was not one person wide. Every person on the team needed to become that role.

That is a hard thing to say out loud. Telling one person their job is different now is hard enough. Telling an entire team at once that the thing they built their expertise on is changing is a different order of difficulty, and you cannot rehire your way through it. We saw it in ops before the rest of the department did, so we started moving earlier.

Six months later I left for Legora.

Ninety days or eighteen months

I was on a call with a group of legal ops leaders this week and someone asked whether we had a framework for deploying this. I said we recommend an eighteen-month arc. Someone pushed back straight away: eighteen months is far too long. Their own framework ran ninety days.

We were both right, and we were not talking about the same thing. Ninety days is about how long it takes to prove out one narrow use case. Eighteen months is what it takes to change how a department works. Neither number is wrong until you use one to plan for the other.

That mismatch is worth looking for in your own organization. Most departments have a ninety-day plan and an eighteen-month expectation. They scope a pilot and then wait for an operating model to show up.

The case against me

There is a version of this argument that cuts against my own position, and it deserves saying out loud.

The solo practitioner could rebuild his operating model over a weekend because he was one person with nobody to convince. Doing the same across a department of several hundred attorneys is a different category of project, and you could reasonably conclude the squeeze isn't worth the juice yet. Legal AI was not very good for a long time. It has only recently become genuinely good, and there is still real distance between where the technology is and where a department could push it. If you believe that, waiting is defensible.

I don't, and the reason is demand. The ACC's 2026 CLO survey found 63% of chief legal officers expect headcount to stay flat while their responsibilities expand. The work arrives whether or not the operating model is ready to absorb it. Waiting is a decision to meet it with the model you already have.

How this actually ends

I never got to a number I fully believed. We had just presented the strategy and were about to start the real change management work when I left, so I saw the reaction in that second room and not what came after it.

I am not sure that weakens the argument. The reaction was the finding. Adoption was never going to be enough. Translation was necessary and not sufficient. The part that actually produces measurable value was the part we had not done yet.

That sequence is the work I came to Legora to do, across a lot of departments instead of one. There is a general playbook and it holds up. But every legal team has a different shape, different constraints, and different people who need to hear that their job is changing.

Which leads to the part I still find hardest. Law has trained people the same way for centuries. You do the work, someone senior reviews it, eventually you become the person who reviews it. Asking a department to stop producing and start supervising is not a tooling problem and not a training problem. It is behavior change against a few hundred years of practice, and I don't think anyone has solved it yet.

Choosing the platform is the easy question. The harder one, which has to come first: what do you want your legal team to actually be?

Answer that first, and the tools stop being the boundary. Instead of fitting your work into whatever each system was built to do, you can build the way your team works from pieces, keeping what you need and leaving what you don't. Contracting, intake, spend, whatever the work is. The only real limit left is how creatively you're willing to rethink it. That's what legal work without limits means to me.

Next in this series: what it looks like when a legal team stops fitting its work into the boxes its systems came in, using one contract request as the example, and why legal operations should lead that redesign.

The author

Mark Allen is Senior Director, Legal Innovation at Legora, bringing extensive experience from leading operations, strategy, and AI at Zillow, Netflix and Fenwick & West. He serves on the Board of Directors of CLOC and advises multiple legal technology companies. He holds a J.D. from Berkeley Law, as well as degrees from the University of Southern California and American University

Stay in the loop

Our monthly newsletter is bringing you the latest updates, fresh insights, and stories from our customers.

Meet a collaborative AI for lawyers.

Work will never be the same.

Meet a collaborative AI for lawyers.

Work will never be the same.

Product

▼

Solutions

▼

Certified

▼

Resources

▼

Company

▼

Legal

▼

Social

▼

© 2026 Legora. All rights reserved.