Legal AI Works, Why Is the Return So Hard to Find?

By Alejandro Castellano, CEO, Caddi.

Two years ago the question in front of every management committee was whether the AI worked. That question is largely settled. The drafting tools draft. The review tools review. Firms that were running pilots in 2024 are running deployments now, and the people using them are not going back.

And yet when a managing partner asks what the spend returned, the answer is usually a slide about hours saved rather than a line anyone can find on a P&L. The gap between those two things has become the most uncomfortable conversation in legal technology.

It is tempting to treat that as a measurement problem: better telemetry, better adoption metrics, better attribution. I think it is something else. It is a location problem. The return is real and it is substantial. It is simply not in the place the industry has spent three years pointing at.

Every firm runs two economies

Inside a single firm there are really two businesses.

The first is the practice of law: billable, judgment-bound, watched obsessively. Realisation and utilisation are tracked to the decimal point, reviewed monthly, tied to compensation.

The second is the business of law: intake and conflicts, matter opening, docketing, time capture, billing, collections, and the document and vendor plumbing no client ever sees. It is non-billable, it is barely measured, and precisely because no one measures it, it is where the waste accumulates. A firm cannot recover what it has never counted.

Utilisation, realisation and collection each take their cut, and by the time a nominal day of work has passed through all three, a firm keeps roughly 75% of what it thought it sold. Very little of that missing quarter goes astray in the drafting.

Why the visible target struggles to produce a number

This is not an argument that practice-of-law AI does not work. It plainly does, and firms that ignore it will regret it. The difficulty is what happens when it works.

Compressing a billable task compresses a billable hour. Under an hourly model, a firm that succeeds at this has made its core product cheaper to produce and harder to charge for. The honest response is to re-price, moving towards fixed fees or outcome-based arrangements that let the firm keep the gain. That shift is coming and firms should prepare for it. But it is a pricing transition, negotiated client by client over years. It is not a return this financial year, and a management committee asking about last year’s invoice will not accept a pricing roadmap as an answer.

Automating a non-billable hour carries none of that ambiguity. There is no client to renegotiate with, no realisation impact, no privilege question, no malpractice exposure, because the work never touched legal judgment in the first place. The hour was pure overhead. Removing it removes cost. That is a number, and it arrives immediately.

Most firms cannot see the work in the first place

Here is the part that surprised me most. Before a firm can automate anything it has to know what it does, and at an organisational level most firms genuinely do not.

Ask a Chief Operating Officer to describe their departments and they will do it fluently. Ask them to name their systems and they will list them. Ask which processes run across the firm, how often, who performs them, how many tools each one crosses and what the whole thing costs in aggregate, and the confident answers stop. They have instincts about priorities. They do not have a source of truth.

Which means the first question of any automation programme, where do we begin, gets answered by whoever complained most recently. That is not a strategy. It is a queue sorted by volume of grievance, and it is why so many firms have automated three things nobody particularly needed automated while the expensive processes carried on untouched.

This is the deeper cost of leaving operational work unmeasured. It is not only that the waste accumulates. It is that the firm has no evidence on which to prioritise, so the spend goes wherever the noise is. The starting point is not a tool. It is an honest inventory of what your firm actually does, ranked by what it costs you.

It stayed manual because it is hard, not because it is trivial

The industry has a habit of calling this work low-value busywork. That framing is exactly why it never got fixed.

Filing an executed contract sounds like one step. In practice it is a decision tree. Did both parties sign, or only one. Is this the executed copy or another draft in the same email thread. Is there already a version filed against this matter, and does this one replace it or sit beside it. Did it arrive as a scan with no text layer. Are there two counterparties rather than one. Is the effective date missing, and is that a reason to stop or a reason to proceed.

Opening a matter is the same shape. So is a conflicts check, a records chase, or reconciling a day of client payments. The happy path is a handful of steps. The real process is the happy path plus forty exceptions, most of which exist only in the head of the person who has done the job for six years and has never written any of it down.

That is also why consolidation is only a partial answer. The hour is rarely spent inside the DMS or the practice management system. It is spent crossing between them, and every firm’s crossings are idiosyncratic: a product of which systems were bought, in what order, which arrived through a merger, and which associate designed the workaround in 2019 that everyone still follows. No vendor roadmap covers your particular seams.

So the work sat there. Too small to justify an IT project, too specific to buy off the shelf, too frequent to ignore. Firms that did try paid consultants to sit and watch, produced a specification, built against it, and then discovered the exceptions in production one incident at a time.

What has actually changed

The specification problem is the one that has cracked, and not because models got better at guessing.

It is now possible for the person who does the job to demonstrate it and talk it through, the way they would train a new hire, and for the AI to interrogate them while they do it. What if only one party has signed. There is already a version filed, do you overwrite it. What tells you an executed copy from a draft. Each answer becomes a rule rather than a note in a transcript.

That conversation is the whole thing. It is a deliberate attempt to extract in twenty minutes the same body of exception handling a traditional project discovers over six months of production incidents. The unit of automation drops from a project to a conversation, which is the first time the long tail of firm operations has been addressable at all.

One caveat worth stating plainly, because it is where a lot of current enthusiasm will break. Operational work needs to run the same way every time, and a model deciding every step of a long task does not do that. A step that is right 95% of the time is right about three quarters of the time across six steps, and these processes routinely run far longer than six. Use judgment where the work requires judgment, such as reading an unstructured document, and deterministic execution everywhere the work has to be exact. Then log every run, because in a firm whose regulators, clients and insurers can ask what the software did and when, the audit trail is not a feature alongside the automation. It is the precondition for deploying any of it.

Where to point it

Three questions do most of the work. Is the task bounded, with countable inputs and outputs. Does it recur often enough that the time adds up. Does it depend on legal judgment. Where the answers are yes, yes and no, you are looking at the highest-return and lowest-risk AI available to a firm today, and the only category where the business case can be settled with arithmetic rather than argument.

Professional services is the last large sector of the economy without an operations layer. Manufacturing got ERP. Sales got CRM. Software engineering got continuous integration. The operations floor of a law firm got a person with a checklist and an institutional memory nobody else shares, and in most firms it still has one.

The firms that show a number this year will not be the ones that put AI closest to the law. They will be the ones that put it closest to the work nobody was counting.

You can find more about how Caddi discovers and automates operational workflows for law firms here.

–

About the author: Alejandro Castellano is co-founder and CEO of Caddi, which builds AI agents for the back office of professional services firms. He was previously an entrepreneur in residence at the AI2 Incubator, founded by the Allen Institute for AI, and is a machine learning expert from Cornell.

—

[ This is a sponsored thought leadership article by Caddi for Artificial Lawyer. ]