Six months on, the data is harder and one part of the argument was incomplete.
In March I wrote about my first day as a graduate at KPMG, when someone handed me a 300-page inventory register and told me to add it up. On a 10-key, with a tape. Every page twice, both tapes stapled to the page, then all the page totals added together twice, then a run to the FedEx office to make the 7pm cutoff.
My argument was that this was not just entry-level work. It was the training ground, and AI is quietly deleting it.
Six months later I still think that is right. But two things have changed, and one part of what I wrote was incomplete in a way that has been bothering me.
The data got harder
In March I leaned on research from the Federal Reserve Bank of Dallas showing that AI substitutes for work built on codified, textbook knowledge while augmenting work that depends on tacit knowledge. That was the mechanism, and it still holds.
What I did not have was a number.
In August, the Stanford University‘s Digital Economy Lab published a revised study using ADP payroll records covering roughly one in six American workers across more than seven hundred occupations. Not a survey. Administrative reality.
Workers aged 22 to 25 in the most AI-exposed occupations now sit around 19% below their peers in less-exposed occupations. A year earlier that gap was 13%.
It is widening.
Stanford is careful, and so should we be. Their own headline reads “No Widespread Displacement,” and the authors are explicit that these are descriptive patterns rather than proof that AI caused them. If you tell your board that AI has destroyed entry-level work, someone who has read the paper will correct you.
But the direction is not in dispute, and the mechanism they identify is the one I described in March. Employment is falling among young workers in codified-knowledge occupations and rising among experienced workers in tacit-knowledge ones.
The ladder is not being cut. The bottom of it is being quietly unbolted.
The part I understated
Here is what I did not pay enough attention to in March.
Stanford found that the adjustment is happening through reduced hiring, not layoffs.
Nobody is being let go. The door is simply not opening.
I skated past that in the original piece, and it turns out to be the most important thing in the whole finding, because it explains why almost no board has this on its agenda.
Layoffs are events. They have a date, a memo, a communication plan and a set of questions at the next board meeting. Somebody owns the decision and has to defend it.
Not hiring is none of those things. It is a graduate intake deferred by a year, then quietly deferred again. A vacancy absorbed rather than backfilled. A headcount line that stays flat while revenue grows, which reads as efficiency and gets applauded in the same meeting.
So the most consequential workforce decision many organizations will make this decade is being made without anyone making it. There is no memo to point at, no meeting to review, and nobody to hold accountable, because in a real sense nobody did it.
That is a different problem from the one I described in March. In March this was something happening to organizations. It is actually something they are doing, invisibly, and applauding themselves for.
The part I missed entirely
The second thing has been nagging at me since I read the Stanford paper, and it is not in the March piece at all.
In March I wrote about what the junior learns. Reading a room. Recovering from mistakes. Building credibility from scratch. All true, and all about knowledge flowing in one direction.
But entry-level work was doing something else at the same time, and I only saw it recently.
In 1994, after five years and several jobs of being a fairly poor accountant, I rang my friend Deb Bennett to say I needed to work out what I wanted to do when I grew up. She said, “I’ve waited four years to have this discussion with you. You want to become a recruiter.”
Four years. She had been watching me for four years.
I have always told that as a story about Deb, and she deserves it. But look at what made it possible. She could tell me what I was good at because she had four years of watching me do unglamorous work. Not presentations. Not strategy. The grind, which is where you actually discover how somebody thinks, what they notice, what they avoid, and what they are like when the work is dull.
The bottom rung was not only where I learned the business. It was where the people around me learned me.
That is the second loss, and I have not heard anyone name it. When the entry-level years compress, you do not just lose the technical apprenticeship. You lose the observation period. You lose the accumulated, unstructured watching that lets an experienced person say to a younger one, with authority, here is what you are actually for.
You cannot do that from a performance dashboard. It requires time in proximity, on real work, over years. And it is exactly the thing efficiency removes first.
What it costs you when it goes
There is a number that makes this concrete.
A recent Bamboo HR survey found a three-to-one ratio of senior-level hires to entry-level hires in accounting and finance, alongside a finding that a third of new hires in those functions leave within their first year.
That second number almost always gets filed as a retention problem, which sends the money to the wrong place. Another engagement survey. Another wellbeing initiative. Another round of stay interviews.
None of it will work, because the issue is not how those people feel about the organization.
It is that the job got easier and emptier in the same year. The work that used to teach them something went to the machine, and what was left did not. And nobody is watching them long enough to tell them what they are becoming good at, so there is nothing to stay for.
You cannot retain someone into a role that has stopped developing them. They will be polite about it, and then they will go, and their exit interview will say something vague about wanting a new challenge.
A fourth element for the Blueprint
In March I described what I called the Bench-Building Blueprint, which the organizations getting this right are using: deliberate exposure to judgment moments, real apprenticeship rather than passive shadowing, and better measures of readiness than tenure and performance ratings.
I would add a fourth now.
Protect proximity. Whatever else changes about how work gets done, somebody experienced needs enough sustained contact with somebody junior to form a genuine view of them. Not a rating. A view. The kind that lets you say, with four years of evidence behind you, here is what you should actually be doing with your life.
That does not survive being distributed, automated, or compressed into a quarterly form. It has to be deliberately protected, and unlike the other three it costs almost nothing except time that nobody is currently accounting for.
The question, six months on
I asked in March where your next senior leader would come from. Not in three years. In ten.
I would sharpen it now.
Who in your organization is currently in a position to look at a twenty-three-year-old and tell them, credibly, what they are good at?
If nobody is, you do not have a leadership pipeline. You have a hiring plan, and in eight years so will everyone else, all at once, for the same scarce people.
Somebody watched me long enough to change the direction of my working life. That is not a nice story about mentoring. It is infrastructure, and it is quietly being decommissioned.
If you’re leading an organization through this: the number to ask for isn’t your attrition rate. It’s how many people you hired into their first job this year, and what those people actually spend their days doing.
If you’re planning a conference or leadership program: this is one of the conversations I have on stage, alongside live AI demonstrations built for the room in front of me.
Either way, send me a message. I read them all.
