This is Part 1 of a series on work, capability and meaning in an AI-enabled economy.
An AI business case usually includes a number on labour saved. I want it to ask what people will lose the opportunity to learn.
When calculating ROI, organizations should include the costs of reducing talent development channels. Internal recruitment to more senior positions is a normal career development path that AI adoption can disrupt. Early investment in junior employees pays organizational dividends, and the long-term costs of a poorly trained or inexperienced workforce are hard to plan for and absorb.
Let's explore. As things stand today, the work most immediately exposed to AI is often work done by professional near the beginning of their careers: answering first-line customer questions, gathering evidence, preparing summaries, writing routine code, testing, documenting and producing first drafts. Basically, jobs where you learn the ropes.
These tasks seem expendable because they can be repetitive, structured and supervised. But they are also how many people learn. Remove enough of them through automation and AI without replacing that learning paths and an organization does more than reduce its headcount. It weakens its ability to grow the capability it will need later.
Investing in automation, either through robotic processes or AI, is, in essence, a workforce decision. When we make a decision to automate, we make an implicit decision about who gets to learn and grow.
Canaries in the recruitment coal mine
The first warning signal is that fewer people appear to be getting their start.
A recent Stanford study, using payroll data for millions of American workers, found that employment among workers aged 22 to 25 in AI-exposed occupations was 19 per cent below where it would have been had it kept pace with peers in other age groups. The gap appeared primarily through reduced hiring, not layoffs. Experienced workers showed no comparable gap in the same study.
For leaders thinking about succession, that pattern deserves some attention. An organization can retain its experienced workforce and meet today's commitments while bringing in fewer people to learn the work. The immediate results may look healthy even as the pathway into the profession narrows.
The researchers describe these findings as early indicators, not irrefutable proof that AI caused the divergence. Nor do they do not establish a future shortage of expertise - that was not the scope of their analysis. But the trend raises a question leaders should be asking now: if fewer people get the opportunity to begin, where will tomorrow's experienced workforce come from? How will job marketplace react if there are not enough experienced employees to promote? (Hint: Before you say we’ll cover the gap with more automation, read on.)
The job and the learning go together
I have spent much of my career in technology and service operations. Like most other disciplines, it's common practice to consider entry-level work as a waiting room for more important work. It is clearly not. It’s also about learning.
Learning a job from the ground up gives people perspective and context, and is often the foundation for how a person will operate for most of their career. Time spent on the front line shows a junior team member how a decision affects another part of the organization, why a rule exists, where a process fails and what customers or citizens actually experience. Learning on the job also creates consistent feedback loops: make a choice, see the result, receive correction and try again. That is how critical judgement develops.
Here are some examples of how someone junior learns. A customer-support agent learns more than how to close a ticket. They learn how systems really behave, how policy affects a frustrated client, when an exception matters. Each interaction adds context that cannot be acquired from a procedure manual alone. A student researcher learns by searching, comparing sources, following weak leads and having an early conclusion challenged. Feedback from someone more experienced teaches them why one source is credible and why another conclusion is overstated. The final briefing may be the visible product, but the discipline is built through the work that precedes it. A junior developer learns by writing imperfect code, debugging it, seeing it fail in production and having someone explain why. Code review, testing and incidents connect a line of code to architecture, security and users. Senior developers were not born senior. They accumulated judgement through thousands of encounters with real systems and real consequences.
How about automating learning?
It is also true that AI can make each of these learning paths better but it can also remove them. In one well-known field study involving over 5,000 customer-support agents, generative AI increased productivity by about 15 per cent on average. The largest gains went to less experienced and lower-skilled workers, and the researchers found evidence that the tool helped people learn.
That is an encouraging result and may justify an AI investment. In this example, AI is shown as a capability accelerator, making expertise available to people who have not yet had time to acquire it.
Now imagine two executives reading the same study. One sees an opportunity to help new employees become capable faster or quickly close a talent gap. Another might see an opportunity to employ fewer new people. Both make make similar investment choices, but different follow-up actions. One invests in both AI and humans, the other on automation alone. The workforce outcome reflects what each leader decides the productivity gain is for.
How we are accumulating capability debt
Technical leaders understand technical debt. A technical workaround produces an immediate benefit while leaving behind a cost that someone will eventually have to pay.
Organizations can accumulate capability debt in much the same way. It is the growing shortfall between the capabilities an organization will need and its capacity to develop them. The debt grows when organizations remove today's learning work without creating another way for people to acquire the experience, feedback and judgement that work once provided.
To put a fine point on it, capability debt accumulates when junior positions disappear without a replacement pathway. It also grows when AI completes the task while employees receive no insight into how or why, or when productivity gains are harvested entirely as savings rather than reinvested in development. It’s exacerbated when experienced staff become reviewers of machine output and stop being mentors of people. It’s evidenced when organizations start to buy expertise increasingly from the market while contributing less to creating it.
The debt may remain invisible for years, until experienced employees leave or retire, or a vendor relationship ends. In those situations, an organization discovers that while it has people who can operate the process, it has precious few who understand it well enough to challenge, repair or reinvent it.
If enough organizations make the same choice, the consequences extend across the labour market. Retirements and smaller entry-level cohorts leave organizations competing for a shrinking pool of proven talent. Those organizations with money pay a premium to hire from a shrinking talent pool. Others become dependent on vendors, accept vacancies or operate without the expertise they need. Savings taken from development today return as higher labour costs and greater risk tomorrow.
Automating drudgery is therefore not the problem. Much entry-level work is frustrating, poorly designed or unnecessary. We clearly should not preserve bad jobs in the name of character building. Our obligation is to preserve the learning, not the drudgery.
The employment bargain needs a new clause
There is a human side to that obligation. People take jobs for a paycheque, but many also want somewhere to grow, make a contribution and become someone others can count on. Being trusted with more difficult work can mean more than a step towards promotion. It tells a person that their judgement matters.
Most of us recognize an unwritten bargain at work. Call it the Deal. The organization receives effort, judgement and results. The employee receives compensation, but also expects some combination of challenging work, growth, safety and accomplishment. The expectations vary, and they are rarely spelled out. Still, people notice when the organization asks more of them while offering less in return.
I think the Deal remains relevant. For leaders adopting AI, the question is how to keep our side of that bargain as the work changes.
Organizations increasingly tell employees that they must be adaptable, continuously learn, demonstrate resilience and take responsibility for remaining relevant. Employees do have responsibility for their development. But that expectation needs somewhere to land: time to learn, people to learn from and real work through which to put that learning into practice.
Without those opportunities, we are asking people to keep rescuing their careers from changes they did not choose and cannot control. Continuous learning cannot become continuous self-rescue.
Resilience does not make the Deal fair. It is neither inexhaustible nor equally available. Mental health, caregiving responsibilities, financial insecurity and other circumstances affect how much disruption a person can absorb at a given time. Leaders need to take those circumstances seriously when deciding what support people will need.
An organization cannot reasonably eliminate the work through which people develop, withdraw the time and opportunity required to learn, and then treat an employee's lack of readiness as a personal failure. A modern employment Deal should include a developmental obligation: when technology removes a learning pathway, the organization must consciously build another one.
Rational decisions can still produce a bad outcome
The AI Layoff Trap, a recent economic paper by Brett Hemenway Falk and Gerry Tsoukalas, makes a related argument at a macroeconomic level. They model how organizations acting rationally in competition can automate beyond what is collectively beneficial when displaced workers are not reabsorbed quickly enough into the economy. Each organization captures the advantage of reducing its own labour costs, while the broader loss of worker income and consumer demand is shared across the market.
The paper's caution is that collective awareness alone does not solve the problem. And neither does generic upskilling. In their model, retraining does not eliminate the problem while displaced workers' income remains unreplaced.
I would extend the authors' argument to capability. An organization can reduce the cost of developing junior employees and then recruit experienced people trained somewhere else. One or two companies can conceivably conduct themselves that way sustainably. But that strategy depends on someone else continuing to invest. If fewer organizations replenish the talent pipeline, some of the future cost falls on workers, other employers and society: higher recruitment costs, longer vacancies, greater dependence on outside expertise.
The risk could grow as AI becomes more reliable. Today the pressure may be greatest on customer support, junior research and routine coding. Tomorrow it may move into intermediate analysis, solution design and more consequential professional work. That progression is not inevitable, but it is plausible enough that leaders should design for it now.
Place capability development into the investment decision
In an earlier post, I argued that AI governance is ultimately about making purposeful decisions about what and why we automate. This is the workforce version of the same argument.
CIOs and other leaders should consider a capability impact assessment whenever an AI investment materially changes work. It does not need to become another hundred-page gated artefact, but it does need to force honest answers to a few questions. Which tasks will disappear, and what did people learn by doing them? What new source of practice, feedback and increasing responsibility will replace that learning? Which capabilities will the organization need in three to five years, and where will those people come from? Who is accountable for maintaining the pathway from novice to expert? How much of the productivity gain will be reinvested in developing people? Are we measuring progression and judgement, or only hours and positions saved?
Our answers may lead to redesigned entry-level roles rather than fewer of them. Referencing the junior resource examples above, a support agent might use AI for routine interactions while rotating through escalations and service improvement. A junior researcher might let AI create the first map of an issue, then be accountable for verifying the sources, finding what was missed and defending the conclusion. A developer might generate more code but spend more time in architecture discussions, peer reviews and incident response.
In each example, AI removes some work while the organization deliberately preserves the conditions for growth.
Employees have agency and accountability here too - it’s not all the employer’s obligation. As workers, should all use AI to expose ourselves to better questions, faster feedback and work that stretches us, not to conceal what we do not understand. We need to practise verification, seek difficult assignments, build domain knowledge and remain accountable for the conclusions we advance.
But let me be clear that this responsibility cannot be downloaded entirely onto the individual. No online course can be a substitute for the experience of being trusted with real work, receiving informed feedback and gradually carrying greater consequences. Institutions still control access to many of those opportunities. They should provide them fairly, and employees should make sure they are ready for them.
So, who gets to become an expert?
Work has never been the only source of meaning, and employers should not pretend they can provide an entire life's purpose with a job title. But the workplace has offered many people a place to become useful, exercise their judgement, contribute to something larger and grow into a stronger version of themselves.
Those opportunities matter. Feeling valued should not become an accidental casualty of an efficiency program, even though a person's worth does not depend on their employment.
The debate about AI and work usually focuses on how many jobs will remain. That is an important question, but it is not enough. We also need to ask who will still be able to enter a field, learn its craft, earn greater responsibility and eventually become expert.
If we automate the bottom rung of the ladder, we should not be surprised when fewer people reach the top.
References and further reading
Brynjolfsson, Erik, Bharat Chandar and Ruyu Chen. Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence. Stanford Digital Economy Lab, revised August 2026.
https://digitaleconomy.stanford.edu/publication/canaries-in-the-coal-mine-six-facts-about-the-recent-employment-effects-of-artificial-intelligence/
Brynjolfsson, Erik, Danielle Li and Lindsey R. Raymond. Generative AI at Work. Revised November 2024.
https://arxiv.org/abs/2304.11771
Falk, Brett Hemenway and Gerry Tsoukalas. The AI Layoff Trap. Revised June 2026.
https://arxiv.org/abs/2603.20617
St-Jean, Tim. AI Governance Isn’t About Control. It’s About Purpose.
https://www.timstjean.ca/focus/ai-governance-isnt-about-control-its-about-purpose