The phantom ROI: why AI is costing many law firms a lot of money, and almost nobody wants to talk about it

Beyond the various articles I keep running into, I've been having the same conversation, more and more often, with partners at law firms. It's no longer about whether they should adopt AI, that's been settled for a year or two now. It's a more uncomfortable conversation: "we spent the money, trained the teams, signed the vendor contracts, and revenue hasn't moved." Nobody likes saying this out loud, because it sounds like admitting an investment mistake. But I think it's worth talking about openly, if only because the industry's own data confirms it isn't an isolated case or just my impression.

At small firms and in solo practice, AI is delivering exactly what it promised at the operational level: it saves time, improves the quality of the work, speeds up responses to clients. The problem shows up when you look at revenue. Only about a third of these firms saw AI translate into more billing, while at large firms that number nearly doubles. So the discomfort I hear on the ground isn't disguised resistance to change; it's a real pattern, and it hits hardest whoever has the least room to absorb an investment without immediate return.

The more comfortable explanation is usually "processes are still adapting, the return will come with time." It's tempting to believe that, but I think it's a half-truth. There's a more structural problem underneath: if billing is still by the hour and AI genuinely does the work faster, that reduces billable hours by definition. Fewer billed hours means less revenue and less profit. Not the other way around. It's simple math, and maybe that's exactly why so few people want to say it out loud.

There's one exception worth separating from the rest. When AI operates in back-office functions, administration, collections, tasks that were never billable to begin with, it cuts direct costs and frees people up for other work. That return is real and, I'd venture, is practically the only kind of AI ROI that's being calculated with any rigor in law firms today. The problem is that most of the money wasn't invested with the back-office in mind. It was invested in substantive legal work, billed by the hour, and that's where the math stops adding up.

There's one figure that changed how I look at this: according to Thomson Reuters, cited by CathCap, only about 18% of firms collect any return metric on their AI tools. Not out of disinterest, but because most simply never built a financial baseline that would let them compare before and after: cost per matter, margin by practice area.

What this means in practice is that "AI didn't bring in revenue" can sometimes be a disguise for "we don't know if it did, because we never built a way to measure it." Those are two very different things, and conflating them leads to bad decisions in both directions: either continuing to invest blindly, or giving up too soon for lack of proof.

There's something even more fundamental underneath, which I think gets talked about too little. The traditional business model of corporate law, senior partners billing for the work of junior associates, organized as a pyramid, was built, even if nobody admits it openly, on a certain amount of productive inefficiency. That inefficiency is what sustained profitability. AI attacks the pyramid right at its base: it does quickly, at almost no cost, tasks that used to be handed down to associates and paralegals and billed hour by hour. This isn't a question of the process not having matured yet. It's a tension between a technology that reduces working time and a business model built on selling working time. Until that tension gets resolved, through fixed fees, new pricing models, ways of capturing the value of efficiency instead of losing it, AI is going to keep looking, on the books, like a cost with no visible return.

Which brings us to the question I find most interesting in this whole discussion: can an AI investment justify itself purely through added quality, even without generating new revenue? I think it can, but only if it's measured with the same rigor you'd apply to a revenue line. One practical way to do that is to translate everything into equivalent value: hours saved times the hourly rate, plus avoided expense, outsourcing that's no longer needed, less software, external fees that no longer get paid. Through that lens, quality and speed are real value. It's just that they show up on the avoided-cost side, not the new-billing side. That's legitimate, as long as it's actually what gets measured and communicated to the partners, and not a vague excuse for a purchase made out of fashion.

At ROOX, this is a discussion we've taken seriously for several years, precisely because the products we develop for the legal sector have algorithms built in whose function is to save time and simplify processes. I don't feel comfortable promising a guaranteed return here without numbers to back it up, that would be falling into the same trap I describe above. What I can say is that the same discipline I'm arguing for in this article is what we try to apply to our own work: building, from the ground up, ways to measure cost per matter and time actually saved, instead of selling the promise of efficiency without proof to back it up.

In the end, I think the question "did AI bring in more revenue?" is badly framed from the start, because it assumes all value is measured in new billing, when a good part of it may sit on the side of avoided cost, quality, client retention, reduced risk. The question worth asking instead is: do we know, with concrete numbers, what the cost per matter was before AI, and what it is now? Without that baseline, any answer (yes, it did; no, it didn't at all) remains opinion dressed up as fact.