# I Turned a Legal Department Framework Into an AI Skill. It Still Needs a Lawyer.

By Wes Pomeroy. Published August 27, 2026. https://wespomeroy.com/writing/legal-department-ai-skill/

Last month I came across the Thomson Reuters Institute's [Value Alignment Toolkit](https://www.thomsonreuters.com/en/institute/value-alignment-toolkit). It's a method for a general counsel to show what their legal department is worth to the business, instead of only what it costs. Seven modules of guides and worksheets, plus a spreadsheet with 126 metrics in it.

The work is good, and following it is a real project. Reading the guides, working through a 126-metric library, deciding how each metric gets measured, assigning owners and targets, then turning all of it into something a board will read. None of that happens in an afternoon.

The framework is a catalyst. A general counsel who follows it ends up with something better than they would get working it out alone, and gets there faster. It's still a fair amount of manual work.

So I built one piece of it as an AI skill, which is a set of written instructions that lets an AI assistant carry out somebody's method on your own documents and answers. One skill out of seven, a first draft, as an experiment. I wanted to find out which parts a machine could take over and which parts still needed a person.

## The framework

TRI's model is called the Four Spinning Plates. A legal department is always balancing four kinds of value at once. Effective is the quality of the advice. Efficient is the use of money, process and technology. Protect is keeping the company out of trouble, and Enable is helping the business grow.

Most departments report on Efficient and little else, because cost and cycle time are easy to count. That is how a legal department ends up looking like nothing but an expense.

TRI's answer is a balanced set of metrics: 8 to 12 of them, two or three per plate, all four plates covered. Their quick-start checklist walks you through picking them, and a lot of it is typing.

## Before and after, step by step

| The job | Doing it by hand | Doing it with the skill |
| --- | --- | --- |
| Name the company's goals | You write down what the business is trying to achieve this year | It asks, and pushes back when you answer in legal terms. "Get outside counsel spend down" is not a company goal, it's legal's version of one |
| Audit what you measure now | You list your current metrics and sort them into plates by feel | It looks up each one against the library and takes the plate from the library's own column. You correct what it gets wrong |
| Find candidates for the thin plates | You read a 126-row spreadsheet | It searches the library and proposes metrics for the plates you're missing |
| Settle on your final set of metrics | You choose 8 to 12 and hope you've covered all four plates evenly | It drafts a balanced set, and a checking script flags any plate that came out empty before you see the draft |
| Check whether the data exists | You think it through metric by metric | It asks what systems you run, then marks each metric available, unreliable, or needing something new built to capture it |
| Write down what each metric means | You fill in the worksheet | It writes what exactly gets counted and in what form - a count, a percentage, a dollar figure - and who owns it. Targets it leaves empty unless you've told it what to base one on |
| Turn it into something reportable | You build the table and the slide | It gives you the table, a list of what you can't measure yet, a count of how the metrics fall across the four plates, and a list of everything still unresolved |

The right-hand column is not the tool working alone. You still name the goals, you still supply the systems, you still set the targets, and you still decide what goes in front of the board.

And what changes is bigger than the typing. You don't start at a blank worksheet. It gets you going, and it keeps you inside TRI's method while you work, so the first pass comes out closer to right. It finds the gaps in what you measure today and says what it would take to fill them. When you know your business cold but can't put a metric into words, it drafts the words and you correct them. It does the tedious parts. And it keeps everything you told it and every choice you made, so next quarter, or in the next module, you aren't reconstructing all of it from memory.

## Who this helps

I tested it against three legal departments I made up, because they don't all start from the same place, and the help each one needs is different.

One had clean documents and working systems. There the tool takes on most of the manual work - writing out what each metric counts, who produces it, how often it gets reported. It also keeps everything the team told it along the way, so the next piece of work doesn't start by asking all the same questions again.

One had nothing written down, because the company had just hired its first lawyer. There the tool asks about the business, a topic at a time, and builds a starting set of metrics out of the answers. That department needs a foundation before it needs anything else, and this is how it gets one.

One had a partial, inconsistent pile of numbers kept by hand over a couple of years, which is where most departments actually are. Cleaning that up takes judgment - which numbers still mean something, which ones to let go, what to do about the gaps - and those calls belong to the people who know the work. The tool goes through the pile with them, lays out what it found and which parts are unclear, and walks them through deciding.

## I could only get so far, because I'm not a lawyer

Playing the role of agentic business analyst and engineer, I could read the material, structure it, build it, and test it against three invented companies. Telling a company objective from a department's restatement of one is my day job in another setting. What I could not do was certify any of it - that TRI's method is applied the way a legal expert would apply it, or that what comes out is right.

Partway through, a review turned up a rule I had credited to TRI that TRI never wrote. I had the skill picking 5 to 8 metrics. TRI's actual number is 8 to 12. Nothing broke. A framework built on my number would have looked completely normal on its way to a board.

That's what happens when you're close to a field but not working in it. It didn't look wrong. It read like a careful summary of TRI's material. You'd have to do this work for a living to catch it.

I was experimenting and learning, so I put it down. Picking it up would take someone who knows TRI's framework and has run a legal department, both in the same head, working on it as it gets built. And it never finishes: the guidance changes, the library grows, and every revision is a call about legal departments, not about code.

A growing number of people will be able to build a first draft of something like this. But to get it right and keep it there, you need someone who knows the field. I didn't have that.

## This is one use case that worked

This started as a test. I've been building agentic workflows for my own projects - ways of getting an AI assistant to carry out a defined method from start to finish - and I wanted to know whether the same techniques would hold up on somebody else's specialist work. They did, and almost nothing had to change. One skill, one use case, and I was happy with how it came out.

The techniques aren't specific to legal metrics. Anywhere a good method exists and people are working through it by hand, the same approach applies, and the work gets faster and better at the same time.

And the better the underlying method, the more there is to gain. Everything this skill does well, it does because TRI's method told it what to do. Their thinking was already finished and sitting in a set of PDFs. This put it to work. There's a lot of good method sitting unused because working through it costs more time than anyone has. That just got a lot cheaper.