This article was original published in LexLatin, here.
There’s a scene that’s becoming increasingly common in law firms: a lawyer discovers that a task that used to take 12 hours—reviewing a contract, classifying and reading hundreds of documents, drafting a first version—can now be done in 20 minutes with the help of a large language model (LLM). The reaction is one of euphoria and bewilderment, because at the very moment they celebrate the efficiency gain, their own firm’s billing system is telling them they’ve just lost 11.5 hours of revenue. This unease signals that AI isn’t just another tool to add to a lawyer’s arsenal, but a force pushing them to rethink their entire business model.
The arrival of agents and the constant improvement of models mean that this rethinking of the business is no longer a strategic option but has become unavoidable.
It’s worth clarifying a technical distinction: a chat is not the same as an agent . The former is a co-pilot who responds, summarizes, and drafts, but always within the existing workflow and under the guidance of the lawyer who questions them on a case-by-case basis. The latter chains tasks together from beginning to end (planning, consulting sources, executing intermediate steps, reviewing their own work) and, in doing so, doesn’t accelerate the workflow: it reorganizes it. The difference is not one of degree but of nature, and it’s the line that separates individual efficiency from organizational change . The chat makes each lawyer a little faster; the agent forces a rethinking of who does what, when, and with what structure. Most firms stick with the former because it’s convenient, fits within the existing framework, and doesn’t require any changes.
In reality, much of what is presented as organizational use of AI is individual efficiency disguised as transformation. The use of AI that truly reorganizes is what transforms the business, and it’s also what causes discomfort.
There’s a tension running through this entire conversation: the gap between what technology promises and can already do, and what organizations are actually willing to change . Models are advancing at breakneck speed; the economic and cultural structures of firms, on the other hand, are moving with the slowness of those who have much to lose.
1. AI and the business model of firms
The contradiction facing lawyers is the efficiency paradox. Under the billable-hour model, every productivity improvement translates into a lower bill. In other words, firms that bill by the hour have a structural incentive against being efficient. For decades, this contradiction remained latent, masked by comfortable margins, but AI has brought it into full-blown crisis. The question (currently being debated by consultants, legal economists, managing partners, and even the firms’ own clients) is whether we are witnessing the structural end of the billable-hour model or simply its most serious challenge to date.
The evidence points to the second answer: the most recent market data from Thomson Reuters shows that the vast majority of legal fees are still paid on an hourly basis, and that firms closed their last fiscal year with record profits. The hourly billing model remains very much alive, but it is heading toward gradual erosion.
To counter the skepticism of those who think this is futuristic, it’s worth remembering that in legal practice, AI use cases are already concrete and measurable: massive document analysis (reviews that previously took months are now resolved in days), expansion of the scale of cases a firm can handle, commoditization of services previously tailored to individual needs, and systematic reuse of accumulated knowledge and precedents. At the same time, more and more firms are charging flat fees for services that were previously billed hourly.
In this seemingly alarming situation, it’s worth recalling Jevons’ paradox . In the 19th century, William Jevons observed that improving the efficiency of coal use didn’t reduce consumption, but rather increased it; that is, as the cost of coal decreased, new uses emerged. Something similar could happen with legal work, since if AI reduces the cost of legal services, it not only lowers the bill per case, but also makes it profitable to handle cases that were previously unprofitable. There is a huge latent demand for legal services (small businesses, individuals, prevention, etc.) that is currently going unmet because the hourly rate is prohibitively expensive, or because the small firms that currently handle these cases are limited, by their size, in taking on more.
Lowering that price doesn’t necessarily shrink the market; in fact, it could expand it. This certainly doesn’t apply to sophisticated matters where the client is willing to pay whatever it takes. Even so, the trend points toward a larger and cheaper market, not a smaller one, and it reframes the question of surplus: perhaps the winner won’t be the one who charges the most per case, but rather the one who can handle many more.
The ability to capture this previously untapped market is part of the organizational change that AI necessitates, and this change cannot be achieved without impacting business operations (business development, proposal submissions, pricing intelligence, and workflow optimization). Organizational change here, too, is difficult, uncomfortable, and scarce, but it is the only kind that truly transforms the business.
Now, if the hour is no longer a useful measure, what do we replace it with? Productivity has to be measured differently, and the only proxy can no longer be billed time.
One option already under discussion is measurement by tokens (the computational units consumed by a model) as a complement to billable hours, which makes even more sense for measuring agent deployment. But tokens structurally redefine many premises about the nature of legal work, its allocation, and its value. And they carry a fundamental problem: they measure cost , not value . A token doesn’t distinguish brilliant work from mediocre work (just as the hourly rate never did, only the hourly rate was better at disguising the difference). Changing the unit without changing the logic would be repeating the same mistake under a different name.
Under this logic, the return on investment in AI should not be measured by hours saved, but by its impact on profitability . True ROI is reflected in variables such as new pricing models, expanded capacity to serve customers, additional revenue and improved margins, better customer experience, and a stronger competitive position (all ideally without significantly increasing staff, which raises questions about professional development and the creation of new roles that warrant separate analysis).
The mismatch lies in practice: most companies still measure performance by usage and time, and many AI pilot programs fail to demonstrate any impact on the bottom line. The promise is profitability, while the reality is still measured by the stopwatch.
And here comes the most uncomfortable question: if AI reduces the cost of producing legal work, who gets the surplus: the firm, through profit margins; the client, through price; or does the competition squander it? The conventional answer assumes there is a surplus to be distributed. Whether there is one is debatable.
Firms have to invest considerable sums in models, data infrastructure, new roles, and training to implement AI seriously, and that cost doesn’t behave as a one-time expense, but as a new structural and recurring overhead .
Industry spending data published by Thomson Reuters already hints at this: investment in technology and knowledge management is growing well above inflation, while customers are pushing for cutbacks. Perhaps the real question isn’t who captures the profit, but who survives the investment.
2. We need to talk about the data
If the return isn’t (or isn’t yet) in cost savings, where is it? The answer points to an asset that firms have right in front of them, yet almost none are leveraging it: their data . Lawyers are sitting on a goldmine (precedents, briefs, negotiated clauses, transaction histories, pricing and business intelligence). However, very few organizations are seriously working to build the infrastructure that would allow them to classify and utilize all that legal and operational information. The point is both technical and strategic: without organized and structured data, AI agents only deliver a fraction of their potential. You buy the engine and leave the tank empty.
For decades, large law firms won clients and cases through talent and brand recognition. In a world where everyone has access to the same frontier model (the same one used by the competitor across the street), the competitive advantage shifts to those with their own data and best-structured processes . A boutique firm with clean, well-governed data and processes is positioned to outperform a disorganized giant. However, it’s important not to idealize this: building such an infrastructure is expensive and still partially favors those with more capital to invest. The direction is clear: the defensive moat of the future is not prestige, it’s data. This, by definition, is the one thing a technology provider cannot sell, because it belongs to each firm.
3. The way forward
The promise is that AI will transform the practice of law; however, prevailing behavior tells a different story. The most real risk isn’t that firms will ignore AI because almost all of them already use it, but rather the false transformation : adopting it to do exactly the same things a little faster, declaring victory, and then leaving the business model untouched. When everyone automates the same tasks simultaneously, no one gets ahead. Most organizations will remain at the surface level, but that’s the best news for the minority who dare to delve deeper. The opportunity lies not in using the tool, but in redesigning the business around it.
What separates that minority from the rest? It’s not about buying more licenses, but about developing a specific skill set, organized in four phases:
- Diagnose : Identify high-impact opportunities, build business cases, and manage change.
- Design : redesign workflows, define ROI metrics that measure profitability and not time, and align stakeholders.
- Deploy : take those flows end-to-end, integrate them with existing systems, and train the teams.
- Measure : track the return on investment and expand into new use cases.
This sounds simple, but the reality is that since the vast majority of organizations don’t do this, they end up with AI pilot projects that stall before generating value. However, this doesn’t contradict this framework; rather, it justifies it: precisely because execution fails so often, deliberate expertise is needed in each of the four phases.
The conclusion, then, is not a call to enthusiasm or resignation, but to realism. This isn’t a race to save hours: it’s a capital gamble with an uncertain outcome, whose true prize lies in an asset (data) that companies already possess but haven’t learned to leverage. Those who use AI to do the same old things, just a little faster, will lose. Those who rebuild their business model, processes, and data infrastructure around it will define the next decade of the profession.
Patricia Villa Berger is Legal Advisor in Legal Ops and Legal Tech at KermaPartners ; Dan Safran is CEO of Unbiased Consulting ; and Adolfo De Unánue is Director of the Data Science Center ( ITESM ).