Since the Industrial Revolution, automation has swept through factory floors, decimated assembly lines, and robotized logistics sorting centers with relative indifference from what is, somewhat imprecisely, called the elite. That was the course of history. But as soon as language models begin to touch the prerogatives of knowledge workers (lawyers, consultants, journalists, academics), automation ceases to be a simple technical efficiency gain and becomes an existential civilizational crisis.
Because we intellectuals have a platform and know how to write. We project. We project our anxieties onto the entire social body. But these anxieties are not new; they have simply reached a segment of the population that shares them on LinkedIn.
A further observation, ironic in nature: the fact that writers are panicking does not mean their jobs are disappearing. To date, from a macroeconomic and empirical standpoint, no negative causal shock to net employment has been demonstrated. I know — this is not what our social media feeds, saturated with alarmism, typically report. Yet it is the strict truth from a scientific perspective: we are confusing two distinct phenomena:
- Correlation (the public availability of language models and shifts in the distribution of labor income)
- Causation (the attribution of those shifts to the availability of those models).
Recently invited to a panel of legal practitioners on these issues, I observed just how thoroughly dominant diagnoses depart from sound premises to reach erroneous conclusions.
The Historical Analogy Trap
At that panel, my colleague Ricardo Rodriguez raised a fundamental point: current foundation models cannot be compared to prior technological waves. He is entirely correct. Computing over the past forty years operated through rigid codification of formal rules (deterministic systems); LLMs operate through probabilistic inference and language manipulation. The cognitive levers they engage, and the qualification segments they affect, are therefore not the same.
That said, this shift should not be made to carry more meaning than it actually holds. Establishing that a tool is of a new nature does not authorize predictions of massive job destruction or a magical explosion in productivity, however insistently AI vendors repeat those claims. A different tool requires, above all, a different analytical framework. When the moderator asked me, “How do we measure the productivity of this technology?”, the real problem surfaced: we are attempting to evaluate a qualitative leap with measurement instruments forged for Taylorist management.
Traditional Metrics
Applying classical indicators from the legal or service sector to LLM impact measurement risks systematic error. Consider a few examples.
-
Measuring productivity in time spent mechanically penalizes any algorithmic efficiency gain.
-
Adopting a deliverable-volume metric (pages drafted, memoranda, clauses produced) introduces a separate error: text volume measures not productivity but informational noise inflation. Drafting a forty-page opinion when a one-page briefing note would suffice represents a regression in real productivity.
-
Assuming that automating 30% of a role’s tasks allows reducing a person’s working time by 30% ignores the cost of human supervision.
If these metrics are obsolete, what should replace them?
The Myth of Junior Obsolescence
In law firms and legal departments, the refrain is now well rehearsed: “We will no longer hire trainees or junior associates; AI will handle their research and first drafts.”
When I hear that, I want to say: bite into generative AI as hard as you like, but do not be surprised if you break your teeth. Because these tools give the impression that they “do the work”, the assumption follows that they actually do. Real-world full-automation cases tell a different story.
This prediction of junior obsolescence rests on a double error of judgment.
First, organizations tend to reserve advanced tools for senior partners or experienced profiles aged fifty or sixty, on the grounds that they are “profitable” and their billed hour commands a higher rate. The reasoning appears sound: their time is more valuable, so it should be saved. Yet, counterintuitively, the reverse is true. If these profiles bill at the rates they do, it is because their legal judgment is the asset. They should not be the ones primarily operating these tools; on the contrary, the firm should fill their schedules so they continue billing at full capacity. Additionally, many of these profiles, having not grown up with the technology, engage with these interfaces reluctantly or clumsily. Junior associates and trainees, by contrast, handle these technologies with native fluency and agility.
This argument alone is sufficient to begin envisioning a new way to measure AI-associated productivity in law firms. A second point follows: the illusion of time saved by the senior.
Again, intuition misleads. One might assume that if a senior partner decides to stop delegating to a trainee and prompts the AI directly, work will be obtained faster. But consider: when a senior partner gives an instruction to produce a memorandum, there is generally no requirement to receive it within the next thirty seconds. The senior partner is, above all, a multitasker, because demand on their time is constant.
What happens, then, when the senior partner is left alone with their generative AI? The answer is straightforward. They spend valuable time formulating queries and, above all, tracking hallucinations and model approximations. Time spent correcting a machine is destroyed time, of zero value: the senior partner has trained no one, transmitted no tacit knowledge, and monopolized a profile with an exorbitant hourly cost on textual debugging.
David Autor’s Thesis: AI as an Expertise Compressor
This is where the work of MIT economist David Autor becomes relevant. Historically, computing technologies have over-valued hyper-specialization. Addressing a niche problem, cross-border tax law, complex patent litigation, or rupture proceedings, required mobilizing a scarce expert whose fifteen years of specialized practice constituted a scarcity premium.
Generative AI inverts this dynamic: it functions as a cognitive prosthesis.
It does not replace expertise; it lowers the cost of access to the technical scaffolding. It enables generalist and intermediate profiles to access, instantaneously, a corpus of procedural expertise previously inaccessible without years of siloed practice. Who are these profiles? Precisely junior associates and trainees. The task is therefore to configure these tools optimally so they serve as a bridge between senior and junior partners. And that is, fundamentally, a knowledge management challenge.
A New Metric: The Complexity Shift
The true measure of AI productivity in intellectual professions is neither time saved nor the number of junior positions eliminated. Those can be measured, but such measurements do not capture productivity in the economic sense of the term, that is, higher billing.
The relevant indicator is the following:
What proportion of highly sophisticated problems, requiring until recently the exclusive involvement of a hyper-specialized profile, can today be handled, instructed, and resolved successfully by a junior or generalist profile equipped with advanced AI?
This indicator is compelling for three reasons:
-
It connects directly to business model value: highly sophisticated problems generate the highest margins and the greatest client value.
-
It measures a cost arbitrage: resolving a complex problem through an agile junior-plus-AI pair, at a fraction of the cost of an external hyper-specialist or an overloaded senior partner, frees that senior for pure strategy, negotiation, and high-quality technical document review, enabling participation across more high-value matters.
-
It restores the learning chain: rather than being confined to routine document-sorting tasks already handled by the algorithm, trainees and juniors are deployed much earlier on intellectually intensive matters, under the straightforward final supervision of their seniors.
Start Measuring What Matters
Those who predict the eradication of new graduates by artificial intelligence commit the same error as those who conflated, two years ago, Big Tech layoffs driven by interest rate increases with a supposed algorithmic great replacement.
Artificial intelligence should not empty law firms of their trainees. If executive leadership and senior partners stop using these tools as souped-up typewriters for time-pressed seniors, AI will be the stepping stone that allows a generation of augmented generalists to dismantle the silos of hyper-specialists.
That is where the real productivity leap lies. It only requires the intellectual honesty to look in the right place.