The Juniors Got Better. The Bench Did Not.
Co-Founder & CEO of CLRT
The report reaching the partners' meeting this quarter is a pleasant one. Since the firm rolled out its AI drafting assistant, first drafts from the associate pool read like the work of people several years more senior. Turnaround is shorter, review notes are thinner, and the people who used to need the most correction now need the least. It is natural to read that as the training problem solving itself. The first rigorous long-run test of that reading, published on 7 October 2026, says something more uncomfortable. The drafts did get better. The juniors producing them, measured three months later with the assistant switched off, did not. The seniors did. A firm that measures only what leaves the building will see a success, and will not see what is happening to its bench.
The study deserves a close reading because of how it was built. David Autor of MIT and a team from Google ran a pre-registered randomised trial with 133 practising patent lawyers at eleven US intellectual property firms, 91 of whom completed the full protocol. Two in three were given a custom AI drafting assistant and the rest worked without one, and every piece of work was scored blind by expert patent attorneys at an independent firm. Label it honestly: this is a working paper, not yet peer reviewed, the study was conducted and funded by Google, and every participating firm has a drafting relationship with Google. Within those limits, the first result is the one every vendor deck promises. AI access raised drafting quality by 0.34 standard deviations at ten days and 0.38 at ninety. Juniors, defined as under seven years of practice, gained most, 0.58 and 0.60, against 0.22 and 0.30 for seniors, a difference the authors note is not itself statistically significant.
Then the researchers did what almost no firm does. At ninety days, every lawyer redlined an existing patent application with no AI at all, a core task that rests on expert judgment rather than drafting fluency. Overall, the lawyers who had spent three months with the assistant did better, by 0.32 standard deviations. That gain belonged entirely to the seniors, who improved by 0.45. The juniors' effect was essentially zero, at minus 0.03. The paper's own sentence deserves a place above the partners' table: three months of AI-assisted drafting opened a junior to senior gap in unassisted judgment that is absent among controls. The people who gained most while holding the tool gained nothing measurable once it was taken away, and the people who gained least while holding it came out sharper. The order flipped the moment the assistant left the room.
The same tool sharpened one group and not the other, and the authors offer a reason. They suggest that foundational expertise may be a prerequisite for extracting durable skill from AI-assisted practice. A senior arrives with a model of what good looks like, so every AI draft is something to argue with, and the arguing is practice. A junior arrives without that model, so the draft is something to accept, and the step where judgment would have formed is the step the tool performs. A different profession shows the same shape. In Anthropic's own randomised study of 52 mostly junior software engineers learning an unfamiliar library, those with an AI assistant scored 50 percent on a comprehension quiz against 67 percent for those coding by hand, and finished only about two minutes faster, a difference that was not significant. It is small vendor research, but its texture matters: the patterns built on delegation averaged under 40 percent, while conceptual questioning scored 65 percent or higher.
The most important finding is the one an average conceals. The juniors' flat result is not a group standing still. It is a distribution pulling apart, and the authors cannot yet say which juniors rose and why. Against controls, the treated juniors' unaided redlining scores showed a fall of 0.42 in the share landing in the mediocre fourth quintile and a rise of 0.22 in the share landing in the poor fifth, with a smaller, imprecise rise in the good second quintile. The mediocre band of the junior bench emptied, mostly downward. The authors describe the assistant as a springboard for some juniors and a cushion for others, and the cushion is the more dangerous of the two because it is comfortable. A firm watching average quality sees nothing, because the risers and the fallers cancel. Nothing in an output report tells you which associate is which, since with the assistant switched on they all produce good drafts.
This is where the usual reading of AI and junior staff goes wrong. The debate is framed as replacement: will the tool do the work juniors used to do. The more serious question is manufacture. A professional firm is a machine for turning juniors into seniors through years of doing the work under supervision, and what it sells is the judgment that machine produces. If the tool performs the step where judgment forms, quality today rises while the supply of tomorrow's reviewers quietly thins, and the people promoted to check AI output a few years from now will include the ones whose unaided judgment never grew. No memo fixes that. It is a design decision about where to point the tool inside the apprenticeship, which tasks juniors do with it, which they must still do without it, and how the firm measures the difference.
The drafts got better. The juniors producing them, measured with the assistant switched off, did not.
A deeper dive
The second-order trap is that every instrument a firm uses to watch quality points at the wrong condition. Review comments, client feedback, turnaround times and error rates all measure work produced with the assistant switched on, which is exactly where juniors look strongest. The decline, where it exists, shows only in unaided work, and almost no firm observes its people working unaided once the tool is in place. The business case deepens the blindness. In this trial the assistant saved about ten minutes on a drafting task whose control group averaged 110 minutes, a real but modest gain that appears immediately on a timesheet. The cost arrives years later, in a promotion round, as a cohort of senior associates who can approve a draft but cannot say why a claim will fail. By the time it is visible it has been compounding for years, and an apprenticeship that took years to erode cannot be rebuilt in a quarter.
The other trap is borrowing the answer. The wider evidence on AI and skill is genuinely mixed across professions and study designs, and this trial is one well-built study in one practice area, with 91 completers and a sponsor that has an interest in the tool. A firm that reads it as proof that AI deskills juniors will over-correct, and keep the tool from the people who stand to learn most from it used well. A firm that reads only the headline gain in drafting quality will under-correct, and hollow out its bench while its dashboards improve. Neither reading is a strategy. What the study does settle is the shape of the risk: output quality and expertise can diverge, the divergence is invisible unless unaided work is deliberately measured, and it lands unevenly inside a single cohort. That makes it an organisational responsibility rather than an individual one. A junior cannot see their own cushion. Only the firm, deciding where the tool is pointed and what gets measured, can see it and act on it.
Work with CLRT
CLRT works on exactly this question: not whether to use AI, but where to point it inside the way your firm makes its people. We map which tasks in your practice build judgment and which only consume time, decide with you where the assistant belongs and where it must stay out of the apprenticeship, and build the measurement that shows whether your juniors are rising or resting on the tool. If your outputs improved after rollout and nobody can yet say what happened to your bench, the CLRT Ascent diagnostic at ascent.clrtstudio.com is where that conversation starts.

Vishal Sachar is the Co-Founder and CEO of CLRT in Dubai. He helps businesses put agentic AI to work, leads practical AI workshops for leadership teams and operators, and writes on agentic systems, AI governance and the economics of automation. Reach him at vishal@clrtstudio.com or on LinkedIn.


