"In God we trust; all others bring data." — W. Edwards Deming

The first time I was asked to put a number on AI efficiency, it arrived with almost no ceremony.

We were talking through a new engagement with the sort of client every practice group wants to believe it can turn into a permanent fixture: sophisticated, demanding, commercially important, and therefore capable of making everyone on the call sit up a little straighter. Close to the end of the call, among a harmony of "nothing from my end, thanks," someone added, in footnote fashion, that the client expected to see an AI efficiency reduction on the bill.

Talk about burying the lede.

The request was not ridiculous, not at all. Firms had spent the better part of a year telling clients that AI would make legal work faster and more efficient, and clients, having heard us say this, had committed the small social crime of believing us. If the work was going to be faster, why shouldn't the price be lower?

The difficulty was that no one could yet say, with the kind of confidence required to survive contact with a spreadsheet, what the efficiency actually was. I could not tell you how many lawyers would use the tools on a particular matter, how much time those tools would save, or how much of that saved time would be returned to the client after someone reviewed the output, corrected the errors, reworked the prompt, checked the citation, and performed the ancient legal ritual of deciding whether the machine had been helpful or merely enthusiastic.

Still, I built the model.

It had a lever. You entered an assumed AI efficiency percentage, the projected hours came down, and the price moved accordingly. It was neat, responsive, and faintly dangerous in the way all elegant models are dangerous: it made an uncertain thing look like it had finally agreed to behave.

Thankfully, we did not use it.

This is not a Frankensteinian abhorrence to my own creation, but a reluctance to introduce concepts as facts before they are tested against reality. No one wanted to turn a developing capability into a promise we could not yet measure, much less defend as a line-item discount. The client accepted a broader commitment that AI would be used where appropriate, and the matter moved on.

But I kept thinking about the lever.

Not because it was ready, or even because it was right, but because it felt like an early draft of a question the market was going to keep asking. Once clients begin to believe that AI changes the cost of legal work, law firms will need something more precise than enthusiasm and something sturdier than anecdotal evidence.

What clients are expecting and what firms can actually demonstrate are in vastly different fields right now. Thomson Reuters' 2025 State of the Legal Market reports that 79% of legal professionals say AI is having an impact on their work. Impact is not measurable efficiency, and measurable efficiency is not what most firms can currently show a client on an invoice. The legal AI companies are building real tools producing real time compression on specific tasks. What that translates to at the matter level, after accounting for review time, error rates, and the portions of the work AI does not touch, is a much murkier number than the headline suggests.

Meanwhile, clients are arriving to pricing conversations more informed than they were two years ago. The expectation (borderline imperative) of an AI discount is becoming standard, particularly among sophisticated corporate buyers who have watched their own operations adopt AI and seen cost curves move.

The firms that adopted AI early and have not adjusted their pricing yet are capturing margin. Clients are paying pre-AI prices for AI-assisted work, and the difference is staying inside the firm.

As clients get more sophisticated, the expectation is rapidly shifting from "confirm that AI is being used" to "show me where the efficiency showed up." The firms building internal measurement systems now, tracking AI-assisted time by phase, matter type, and task category, will be able to answer that question. The firms that are not will be improvising, with potentially high risk in doing so.

I think approaches worth building now, even if imperfectly, include creating AI efficiency line items in the budget as a discrete input to the scoping process, not a discount applied at the end. "Here are the phases where we expect AI to compress delivery. Here is how that translates to the proposed fee. Here is what happens if it does not compress as expected." This is more honest than a vague reference to AI adoption, and it gives the client something to evaluate rather than take on faith. Track AI-assisted matter completion data internally, even before you are ready to surface it to clients. The data does not exist unless someone builds it, and building it takes time you will not have when clients start demanding it urgently.

One of my favorite things about data sets is that within their absolute truth lives a world of immense creation. You can build whatever works for your firm, your operation, your practice area. Change it or add to it ad nauseum until there is a model that suits you. As long as the data is collected early, and the important metrics are identified.