The capability debt point is the one I'd underline for any leadership team. Cutting entry level hiring looks like a clean efficiency win this year, but the bill shows up five or six years out when you go looking for mid level people who can actually make the calls and theres nobody in the pipeline. And every company seems to be assuming someone else will do the training. The Brainlabs example is the more interesting signal to me, not keeping juniors around out of charity but redesigning the first job around checking and deciding instead of drafting. My guess is the firms that figure that out early end up with a real talent advantage later, while everyone else competes for the same small pool of experienced people and pays for it.
Junior roles are becoming apprenticeship traps because we stripped away the grunt work that forced people to learn the mechanics of a threat. When you automate the triage process, you remove the friction that builds intuition. We need to replace that lost repetition with intentional simulation exercises. If we do not create artificial pressure, we end up with a generation of operators who can read a dashboard but cannot interpret a packet capture. My own experience with incident reporting suggests that the value lies in the slow, manual connection of disparate signals rather than the speed of the initial alert. We have to build environments where the struggle is part of the design.
The screening side has the same problem one step earlier. Once the question moves from "what did you do" to "how did you figure it out", the resume can't answer it. A line like "led a cross-functional Salesforce transformation" gives you the logo and the verb, not what the person decided, how big it was, or what the result was measured against.
What seems to work is picking the one claim that matters most for the role and asking about the decision behind it before the interview loop: what was your part, what did you assume, what changed your mind. Your point about polished artifacts applies to applications too. Greenhouse's 2026 benchmark has applications per job going from 116 in 2022 to 244 in 2025, and AI makes every one of them read better without adding evidence.
The line that should stick is that the junior job was the factory floor where judgment got made. It is worth naming what on that floor actually did the teaching, though, because it was not the grunt work itself. It was the cheap correction loop wrapped around it: a first-year was wrong in a dozen small, low-stakes ways a week and got corrected before any of it mattered, and that is where calibrated judgment comes from, not from producing the draft but from watching it get taken apart when being wrong cost nothing. AI deletes exactly those low-stakes reps, so the first time many people exercise real judgment is now on expensive, unsupervised work. That is why capability debt cannot be documented away: judgment is not transferable information, it is compressed experience of having been wrong safely, and your redesigned first job (less drafting, more checking) only rebuilds the factory if someone senior is still correcting the checking. Otherwise you have just moved the ungoverned guessing up one rung.
The gap you name is real: judgment is a rep count, and AI quietly removed the junior reps that used to build it. The fix is manufacturing those reps deliberately, make calls, get feedback, repeat. Full disclosure I work on allthingspm.app, and the mock interviews exist to force exactly that practice when the job stops supplying it.
The apprenticeship problem is the uncomfortable bit. Companies want judgement sooner while cutting the junior work that used to build it.
Your next senior hire has to learn somewhere too 🔗 https://millennialmasters.net/p/ai-junior-work-senior-hire
The capability debt point is the one I'd underline for any leadership team. Cutting entry level hiring looks like a clean efficiency win this year, but the bill shows up five or six years out when you go looking for mid level people who can actually make the calls and theres nobody in the pipeline. And every company seems to be assuming someone else will do the training. The Brainlabs example is the more interesting signal to me, not keeping juniors around out of charity but redesigning the first job around checking and deciding instead of drafting. My guess is the firms that figure that out early end up with a real talent advantage later, while everyone else competes for the same small pool of experienced people and pays for it.
Junior roles are becoming apprenticeship traps because we stripped away the grunt work that forced people to learn the mechanics of a threat. When you automate the triage process, you remove the friction that builds intuition. We need to replace that lost repetition with intentional simulation exercises. If we do not create artificial pressure, we end up with a generation of operators who can read a dashboard but cannot interpret a packet capture. My own experience with incident reporting suggests that the value lies in the slow, manual connection of disparate signals rather than the speed of the initial alert. We have to build environments where the struggle is part of the design.
https://cyrilsimonnet.substack.com/p/reported-in-24-hours-shared-with?utm_source=substor&utm_medium=substack&utm_campaign=comment
The screening side has the same problem one step earlier. Once the question moves from "what did you do" to "how did you figure it out", the resume can't answer it. A line like "led a cross-functional Salesforce transformation" gives you the logo and the verb, not what the person decided, how big it was, or what the result was measured against.
What seems to work is picking the one claim that matters most for the role and asking about the decision behind it before the interview loop: what was your part, what did you assume, what changed your mind. Your point about polished artifacts applies to applications too. Greenhouse's 2026 benchmark has applications per job going from 116 in 2022 to 244 in 2025, and AI makes every one of them read better without adding evidence.
The line that should stick is that the junior job was the factory floor where judgment got made. It is worth naming what on that floor actually did the teaching, though, because it was not the grunt work itself. It was the cheap correction loop wrapped around it: a first-year was wrong in a dozen small, low-stakes ways a week and got corrected before any of it mattered, and that is where calibrated judgment comes from, not from producing the draft but from watching it get taken apart when being wrong cost nothing. AI deletes exactly those low-stakes reps, so the first time many people exercise real judgment is now on expensive, unsupervised work. That is why capability debt cannot be documented away: judgment is not transferable information, it is compressed experience of having been wrong safely, and your redesigned first job (less drafting, more checking) only rebuilds the factory if someone senior is still correcting the checking. Otherwise you have just moved the ungoverned guessing up one rung.
The gap you name is real: judgment is a rep count, and AI quietly removed the junior reps that used to build it. The fix is manufacturing those reps deliberately, make calls, get feedback, repeat. Full disclosure I work on allthingspm.app, and the mock interviews exist to force exactly that practice when the job stops supplying it.
Validating model output is a strange first job. You need the judgment to spot what's wrong before you've had the reps that build it.