What should legal AI actually cost you?

There is a conversation happening in legal technology procurement that does not make the conference coverage, because it is less exciting than agents and less flattering than adoption statistics. It is about the bill.

The reason it has surfaced now is structural. Systems that plan and execute multi-step work consume vastly more underlying compute than a person typing a single question. As firms move from asking questions to running whole workflows, cost stops scaling neatly with the number of lawyers and starts scaling with the complexity of the work. Pricing models built around a seat were never designed for that, and firms are right to be asking what happens to their software line in eighteen months.

There is a second problem sitting alongside it: very few firms can defensibly quantify what they are getting back.

Time saved is the usual metric, but it is difficult to measure well. It depends on someone estimating how long a task would otherwise have taken, and those estimates tend to become less reliable the further you get from the original work.

That leaves firms in an uncomfortable position. The cost of AI can become increasingly variable at exactly the point when its value is still difficult to express in a spreadsheet.

The answer, we think, is not to avoid consumption-based AI. It is to make the economics understandable before adoption rather than after it.

A firm should know what it is committing to, what happens when usage increases, and what it can expect to pay when a matter becomes unusually complex. It should also decide in advance what success looks like. Not simply whether a lawyer completed a task faster, but whether the firm can take on more work, respond more consistently, or spend more of its time on work that actually requires legal judgment.

Those measures matter because the most valuable effect of AI may not show up as minutes saved.

If a firm can review a matter it would previously have had to decline, turn around a transaction without adding extra days, or give a partner confidence in a larger body of material than they could realistically review themselves, the value is not “efficiency” in the narrow sense. It is additional capacity.

That is the conversation we think firms should be having with AI vendors now.

Not simply: What does it cost per user?

But: What does it allow us to do that we could not do before, and can we predict what that will cost us?