Latest / In The Loop / Why the companies cutting junior headcount are making a decade-long mistake

Transcript

Why the companies cutting junior headcount are making a decade-long mistake · In The Loop

0:00
The current state of the employment market is quite
0:03
interesting. The data on junior employment
0:07
companies that are adopting AI and really drinking the kool-aid
0:10
of AI is is really only moving in One Direction.
0:15
And right now across the S&P 500 headcount fell in 2025 for the
0:19
first time since 2016. That's a net reduction of around
0:23
400,000 jobs ending 8 consecutive years of just
0:28
constant workforce growth. This discussion is going to be
0:31
one of the most important discussions to have as a society
0:35
over the next few years, and there's some really compelling
0:38
arguments against worrying about it.
0:41
Some of the people that I'm going to mention in today's
0:42
episode have discussed the fact that there's been many waves of
0:46
technology that's reshaped, for example, programming and often
0:50
thinks the doomsday narrative mysteries, history.
0:54
But both things can be true at the same time.
0:57
We can be an optimistic and pessimistic view of the future.
1:01
Today's episode's going to explore those views, the data
1:04
points in the history behind them and why the companies that
1:07
are cutting junior headcount are going to make a decade long
1:11
mistake. This episode brings together
1:14
many forms of the latest research and thinking in this
1:17
field to try and present the most up to date.
1:20
View this in the loop with Jack Horton.
1:23
Hope you enjoy the show. So I guess let's start off with
1:38
some of the, the context and the data behind, you know, where
1:41
we're at today. And, and I think a really
1:44
interesting place to start here is, is a fantastic comparison
1:47
that I found online. And I, I fell in love with it.
1:49
So in January 1971, a billboard got put up on the, I was going
1:54
to say motorway, but in the US they're called highways around
1:57
Seattle. And it read with the last person
1:59
leaving Seattle, turn off the lights.
2:02
And, you know, Boeing, the context behind this is Boeing.
2:04
It had started 1969 with 134,000 employees and ended in 1971 with
2:11
about less than 80,000. So, you know, about 4445
2:16
thousand employees have been let go in 18 months.
2:21
They wound on the space program. They had a big supersonic, you
2:25
know, transport project that had been completely cancelled.
2:28
And Boeing claimed they didn't need engineers anymore.
2:31
But you look at today, and the aerospace industry is projecting
2:35
a shortage of more than a million engineers by 20.
2:38
Thirty, 25% of the current workforce is over 55 years old.
2:43
And you know that the cohort that would now say be the senior
2:49
experienced employees never ever got hired.
2:53
And that really sets the scene for today's discussion.
2:56
You know why not hiring junior employees is going to really
3:00
hurt the industry in 1015 or 20 years.
3:04
And this episode isn't about just whether or not AI is just
3:07
destroying jobs, which is often the version of this, I guess,
3:10
conversation that generates a lot of big headlines and noise.
3:15
It's actually just talking about the realities of talent
3:17
pipelines and specifically what happens when the entry point of
3:22
a career is closed before people have had the opportunity to
3:26
start on their journey. As I said, the S&P 500 across as
3:30
a whole in headcount fell in 2025, which was the first since
3:35
2016, a net reduction of about 400,000 jobs, which is ending
3:40
eight years of basically uninterrupted growth.
3:42
And and that's a massive signal. We've had the multiple research
3:45
projects released in the last year all pointing broadly in the
3:48
same direction, which is a companies that have adopted AI
3:52
junior employment drop of between about 7 to 10% within
3:57
about a year and a half, while senior employment at those same
4:01
places keeps rising. Doesn't take a genius to figure
4:04
out that that isn't sustainable. And this decline by the way, in
4:08
most places has happened through slower hiring rather than just
4:11
redundancies as a whole. So people have stopped bringing
4:15
people in at the bottom and and we've had another study for
4:19
example that drilled into the specifics of how that's
4:21
happening and they found that workers aged 22 to 25 in an AI
4:26
exposed job and we can discuss what that is defined as shortly,
4:30
are basically down 13% relative to their peers in less exposed
4:34
jobs since 2022-2023. For example, 22 to 25 year old,
4:40
let's say software engineers, it's around 20%.
4:43
So that's like nearly 25% of the developers that would have been
4:47
hired and no longer being hired anymore.
4:50
That's a huge cohort of people. And a really good telltale here
4:54
is that the gap between where employment should be on, let's
4:56
say pre ChatGPT employment trends is 500,000 less jobs than
5:03
was expected. So there are 500,000 less
5:06
developers than there was predicted based on the current
5:09
trajectory before ChatGPT released.
5:12
And it's a huge mistake for the industry.
5:14
Like yes, maybe some job roles are going to be compressed
5:16
unless you maybe there are new job roles that will emerge.
5:19
But not hiring junior employees and and training people is
5:24
obviously unsustainable. So let's explore why this is
5:27
happening to junior employees specifically because it's less
5:31
obvious than you think. Let's take Luis Garciano, who
5:35
who presented a really interesting argument here.
5:38
And Garciano is a Spanish economist at the London School
5:41
of Economics, and his main work has been exploring how
5:44
organization structure knowledge.
5:46
So who knows what who does what? You know, how information might
5:49
flow through a group of people. And he said a job isn't a list
5:55
of tasks that can be chopped up and automated arbitrarily.
5:59
It's a bundle, So a bundle of tasks that are entangled into
6:04
each other. The question AI pose isn't can
6:07
this task be automated? It's can this task be separated
6:11
from the rest of the job without unraveling the whole bundle of
6:14
tasks. And basically what he's saying
6:16
here is that a job isn't just a set checklist.
6:19
You know, some parts are completely independent and
6:21
freestanding, so you can lift them out, hand them to someone
6:24
else and nothing like breaks down.
6:27
Whereas some parts are woven into everything else and the
6:29
relationships, the judgement calls the accountability.
6:32
And if you try to separate those tasks, you make the entire job
6:36
role collapse. And junior work is typically
6:39
those who have the the most, let's say freestanding
6:42
independent tasks that can be automated.
6:45
But the problem here is that the only way to become a senior
6:47
advanced experienced employee is to spend years doing the junior
6:51
work. You know, the the first year
6:54
legal associates document review is separable, so you can
6:58
automate it. Same with the junior analysts,
7:00
you know, financial model or the graduates, I don't know,
7:04
research deck who they produce for, you know, leaders of the
7:08
company. These tasks can be just lifted
7:10
out and done by AI really quickly and easily.
7:13
And then the tasks that resist that kind of separation and
7:16
automation are those that are senior jobs reading, for
7:20
example, a witness in a disposition and deciding whether
7:24
to settle on or pricing, let's say a deal and managing the
7:27
board and the stakeholders through that deal.
7:31
Choosing which pitch to take, for example, into a chief
7:33
marketing officer and the person that reads it in the room once
7:36
you're there. These are these are very
7:37
entangled tasks that are heavily dependent and have a massive
7:42
impact on the relationships or contexts or consequences and
7:46
accountabilities that hold the whole job role together.
7:49
So your junior work doesn't have that protection.
7:52
And the pattern isn't unique to just software engineers.
7:56
It's happening in journalism, in architecture, basically
7:59
anywhere. The entry level work is
8:01
primarily about turning information into more structured
8:04
output versus having to exercise, you know, really
8:09
difficult judgement calls with real consequences attached.
8:12
But the problem is this creates a massive pipeline problem.
8:15
You know, if AI does the junior work, there's no economic reason
8:18
to ever hire a junior employee. And if nobody hires them, how do
8:22
they become experienced senior employees of a company?
8:26
You know, the pipeline didn't just slow down, it just
8:27
completely stops. So I guess that's a very doom
8:46
and gloom view. And, and there are arguments for
8:48
or against that point of view here.
8:51
And and let's say let's start on the optimistic side because it's
8:54
it's useful and really important to know.
8:56
So Tim O'Reilly, you might recognize his name O'Reilly
9:00
publications, but you got his books at school.
9:03
But he has some really interesting essays.
9:07
Now, he's been a developer in his entire life.
9:09
He's a very, very unique and fantastic point of view here.
9:12
And his argument is, is that he's been watching programming,
9:15
say, transition for, you know, 40 odd years.
9:18
And a lot of the time these types of panics misread history.
9:21
You know, every new programming wave from assembly to Fortran to
9:27
C to consumer BASIC to Windows to cloud look from when you
9:33
were, I guess, standing in the middle of it like the end of
9:36
programming. Every single one of them turned
9:38
out to be just an expansion of programming.
9:41
But the expansion wasn't just that the bar got lower, some
9:43
more of the same people could get involved.
9:45
Each way brought with it an entirely new skill set and
9:49
population that previously hadn't been involved.
9:53
So Fortran didn't make, for example, assembly programmers
9:56
pointless. It brought scientists and
9:58
engineers into computing. People would have never got
10:01
involved with machine code. Consumer Basic didn't, let's
10:05
say, replace C program as it brought in teenagers in the
10:08
bedrooms, A demographic who couldn't have ever been involved
10:11
in the industry. Operating systems like Windows
10:14
didn't stop people building applications.
10:17
It meant that an entire industry of applications is born as a
10:20
result. So the pyramid got wider every
10:23
single level, and the people from the previous way didn't
10:25
disappear, they just moved up. Their skills became foundations
10:29
that basically others just would build on, rather than suddenly
10:33
this new ceiling that meant that they were screwed and they had
10:35
to leave the job. Now personally, I used to buy
10:39
into this belief and recently I've been wondering whether that
10:42
is still the case. Now, one argument is the Javon's
10:45
paradox. So I've talked about this a
10:47
number of times. And this is when something
10:49
becomes so readily available, whether it's affordable or easy
10:52
to do, whatever it might be efficient, suddenly there's an
10:55
explosion of that thing. At first it looks like the
10:58
efficiency gain from the technology that enabled it would
11:02
mean that people would lose their jobs.
11:04
But just as the as a result of the sheer adoption of a thing,
11:08
there is so many more applications and jobs as a
11:11
result. So, you know, steam didn't
11:14
remove the need for coal, it actually increased it because
11:17
efficiency made steam power economically viable in
11:20
applications where previously it would have been impossible to
11:23
justify the cost. You can apply that to
11:25
engineering, which is, you know, making people 1020 more 20 times
11:28
more productive might seem like you may need less of them.
11:31
But over time the argument is that something there's loads of
11:34
new tools that you would build that would never have ever been
11:40
justified or or even considered. And as a result, more engineers
11:44
needed. Now where I have an issue with
11:46
this, to be honest, is the time scale.
11:48
This argument is basically about where the industry could be in
11:51
2035 or 2040, but the pipeline problem I'm really discussing
11:55
here is about 2028 to 2033. You know, if there's no genius
12:00
in 20252026, who were the seniors by 20-30?
12:06
That requires a specific cohort to have been in the job for
12:08
five, 6-7 years, and those people aren't being hired.
12:12
You can't suddenly have more software getting built in 2035
12:16
and needing more employees and and hires but having less people
12:19
to hire because there are just less careers there.
12:23
And the Jevons paradox is essentially an aggregate outcome
12:27
across the entire economy over a longer period of time.
12:31
So it's the culmination of all these things happening over a
12:34
10/15/20 year period. But that pipeline problem that
12:38
I'm actually discussing is about a very specific set period of
12:42
time. To have seniors in the job for
12:44
7-8 years, they need to have been hired 7-8 years ago.
12:47
Another argument against that I guess positive outlook right
12:51
now, and they're doing just seeing somewhat negative here is
12:54
that every previous say technology wave, so Fortran
12:58
BASIC cloud essentially arrived after the experience of the
13:03
previous layer had already kind of been established.
13:05
Fortran programmers had spent years working in Fortran before
13:09
it then got abstracted away. C programmers broad our
13:12
understanding of same memory management into JavaScript and
13:16
each abstraction essentially is being built on the last group of
13:19
people who had actually done the thing.
13:22
The wave would come after the experience form, not before it.
13:26
But this current wave is just happening so much faster.
13:29
You know something really is happening to entry level work
13:32
faster than any previous transition in history.
13:35
Although the Chevron's paradox in the long run, 1520 years,
13:38
period, period of time might be true, there might be just such
13:42
an explosion of things over this next long period that eventually
13:45
there's more, there's more jobs, but over the shorter window,
13:50
companies are making a massive mistake by not thinking about
13:52
their junior pipelines. So the Gevons paradox argument,
13:56
that's very optimistic and the pipeline problem that I'm really
14:00
discussing here aren't competing.
14:02
It's kind of like a sequence of events.
14:04
There needs to be an expansion of demand, but the expertise
14:07
just isn't there anymore to meet it.
14:10
And you eventually get that shortage that I mentioned about
14:12
Boeing facing in the aerospace industry.
14:14
And that will happen in 2035 rather than in 50 years later,
14:18
but it will happen. So I guess to conclude here that
14:37
the arguments to say that each transition of new technology and
14:40
say programmers and software development really did expand
14:44
the surface area the the amount of potential in the space.
14:48
And that is a historical analogy that will probably play out to
14:52
be true. But the problem right now, and
14:55
This is why I think companies are making a massive mistake, is
14:57
that the cohort that will become the leaders of a company in 2032
15:02
and 2035 aren't being hired and therefore won't be trained.
15:07
And there are companies that are doing the right thing here.
15:11
IBM tripled its entry level intake in 2026, one of the very
15:15
few companies that had done done this.
15:17
Publicist, the marketing group tried to restrict its graduate
15:21
training program around AI fluency rather than just getting
15:24
rid of it. But these are a drop in the
15:26
ocean in terms of the general statistics across the the market
15:29
right now. And that that billboard that
15:32
people put up in Seattle, it's a joke, but it's a reality.
15:36
Now, 50 years later, the aerospace companies who made
15:40
those cuts are now probably, looking back, really regretting
15:44
some stupid decisions. And many of these big, big
15:47
companies especially are going to do the same and are going to
15:50
make the same mistake in the future.
15:53
Anyway, thank you for listening and I'll see you next week.