Every large organisation now has a sentence it likes to say out loud. It appears in town halls, in annual reports and on the slides that open the planning cycle. The sentence is: our people are our greatest asset, and we are committed to helping them thrive in the age of AI.
Then there is a number that rarely appears in the same room. In the spring of 2026, PwC fielded its Global Workforce Hopes and Fears Survey to 49,364 workers across 48 countries and regions and 29 sectors. One of the questions asked whether workers had access to the learning and development resources they needed. 51% said yes. A year earlier, 59% had said the same thing.
That is an eight-point fall in a single year. It happened in the same twelve months in which the share of workers using AI at work rose by ten points to 64%, and daily use of generative AI climbed from 14% to 22%. The tools arrived faster. The help with the tools went the other way.
This edition is about that gap. I call it the learning drought: the widening distance between how much organisations expect workers to learn and how much they are willing to carry of the cost of learning it. It is quiet. It does not appear in a layoff announcement or a restructuring memo. It appears in a calendar that has no free hour, a licence that comes with no lesson, and a weekend that has quietly become part of the job.
And it is the kind of problem that looks like a personal failing from the inside. If you are the worker who has not found time to learn the new tools, you do not think the system has failed you. You think you are behind.
The demand side is not in doubt
Nobody serious argues that the skills requirement is staying put. LinkedIn’s Work Change Report, published in January 2025, put it bluntly:
By 2030, we expect 70% of the skills used in most jobs will change, with AI emerging as a catalyst.
LinkedIn, Work Change Report, January 2025
The World Economic Forum’s Future of Jobs Report 2025, built on a survey of employers, uses a narrower definition and arrives at a smaller figure: employers expect 39% of workers’ core skills to change by 2030, down from 44% in the previous edition. The same report found that half of the workforce had completed training as part of long-term learning strategies, up from 41% in 2023. The yardsticks differ, and you can argue about which is better. The direction is the same, and it is not subtle.
So the demand side of the learning market is robust. Employers want people who can work with AI, adapt to changing tools and keep adapting. LinkedIn also found that professionals entering the workforce today are on pace to hold twice as many jobs over their careers as they would have fifteen years ago. More jobs, more change inside each job, more things to learn at every point.
Which makes the supply side the interesting question. Who is supplying the learning, and who is paying for it?
A short history of who pays
For much of the industrial era, the answer was obvious. When a firm installed a new machine, it taught people to run it. The loom came with an operator, the operator was shown how to use it, and the cost of that instruction was part of the cost of owning the loom. Nobody expected the weaver to arrive pre-trained on equipment that had not been invented. The training was part of the investment, because the machine was worthless without someone who knew how to use it.
That logic was never perfect, and it was never universal. Plenty of workers always learned on their own time. But it was the default assumption: if the organisation wants a capability, the organisation builds it. Apprenticeships, graduate schemes, in-house academies and years of sponsored study were the visible expressions of that assumption.
Then the assumption started to erode, slowly and for sensible-looking reasons. Employee tenure shortened. Firms noticed that trained people left, and that other firms collected the benefit. Why pay to train someone who may walk? Better, the thinking went, to hire people who already have the skills, and let the labour market do the training. Skills became something you bought, not something you built.
There is a close parallel in pensions, and it is worth dwelling on, because the pattern is the same. Employers once promised a defined retirement income and carried the investment risk themselves. Over time, they moved to defined contribution schemes: here is an account, here is a contribution, the outcome is now up to you. The employer still wants a loyal, well-prepared retiree at the end. It simply stopped carrying the risk of getting there.
Learning is following the same path. The employer still demands the outcome, a workforce fluent in AI. What has changed is who carries the cost, the time and the risk of failing. A defined-benefit model of learning, where the organisation builds the skill, is quietly turning into a defined-contribution one, where the worker is handed a tool and a deadline and told to take it from here.
The mechanism: how learning goes missing
The drought is not caused by one decision. It is caused by a chain of small, rational ones, each of which looks defensible in isolation.
First, the tool arrives before the lesson. A licence is cheap to buy and easy to measure. A lesson is expensive to design and hard to measure. So organisations buy licences, report adoption and treat adoption as capability. The dashboard says 80% of employees have access. It says nothing about how many of them know what good looks like.
Second, the work gets denser. ActivTrak analysed 443 million hours of activity across 1,111 organisations and 163,638 employees. About 80% of employees now use AI tools at work, up 52% in two years. Daily focus time for AI users fell by 23 minutes, a 9% drop. Collaboration rose 34%, multitasking 12% and weekend work by more than 40%. ActivTrak’s Gabriela Mauch summed up the pattern by saying that AI is not reducing work, it is increasing the speed and density of how work happens.
I want to be careful here. The data does not show that the extra weekend hours are spent learning. It may be catch-up, it may be overflow, it may be a bit of both. What it does show is that the working week has been packed tighter, with fewer uninterrupted stretches. Deep learning needs exactly the thing that is being squeezed out: sustained attention, a protected hour, a margin in which being slow is allowed.
Third, learning migrates to the edges. When work fills the week, the learning does not stop. It moves to wherever there is still room. That means evenings, weekends, commutes and the quiet hours before anyone else is awake. The learning still happens, but it happens on the worker’s own time, using the worker’s own energy, and often the worker’s own money.
Fourth, the support decays without anyone deciding to cut it. A budget is not slashed. A learning hour is simply never protected. A training programme is rolled out to the first cohort and never reaches the second. A manager says “do take the time to explore” and then schedules a review for the same slot. Nobody voted for the drought. Everybody contributed to it.
Put those four together, and you get the number PwC found. More than half the workforce is using AI, and just over half says it has the learning resources it needs. The second figure is falling while the first is rising.
What gets lost
The obvious cost is the skill itself. People who do not learn the tools properly use them poorly, and an organisation full of people using AI poorly is paying for the licence and not getting the benefit. That is a straightforward waste, and it is the one that finance teams will eventually notice.
But there are quieter losses, and they matter more.
The first is confidence. PwC found that daily AI users were 21 points more likely to feel confident about learning new skills, and 68% of them were confident in their job security against 57% of infrequent users. I would be careful about the direction of cause. Confident people may simply use AI more. But the pattern suggests a loop: use builds confidence, confidence encourages learning, learning builds more use. For people who are not in the loop, each turn of it makes the gap larger.
The second is shared language. When learning is formal, an organisation builds a common vocabulary. People know what a good prompt looks like, what a hallucination is, when to trust an output and when to check it. When learning is private, each person invents their own. The result is a workforce in which two colleagues use the same tool in incompatible ways, and neither knows the other is doing it differently.
The third is the middle. This is the loss I worry about most. PwC’s survey divides workers into segments, and two are worth naming here. Its front-runners, the 14% who combine strong AI use with scarce skills, are becoming what PwC’s Pete Brown called more confident and more mobile. The 56% that PwC calls the engine room are in a different place. Only two in five of them say they have the learning resources they need, and the steepest fall in access was among them.
Brown’s warning was that leaders should keep developing the front-runners, but not at the expense of the rest of the workforce, because more than half of workers are not yet benefiting from AI and skills in the same way. That is a diplomatic sentence. Translated, it means: the people who already have the most are getting the most, and the people who run the business day to day are getting the least.
Four people in the drought
It helps to put faces on a statistic. The following four types are my own shorthand, not PwC’s. Most workers are a mixture, but each of us leans towards one.
The Night-Class Learner. She pays for her own course, works through it on Sunday evenings and tells nobody. She is the success story the organisation will eventually point to as proof that its learning culture works. In truth, it is proof that she has the time, the money and the stubbornness to do the job twice: once at work and once on her own account. She is also the person with the most attractive CV in the building, and the most reasons to leave.
The Waiting Room. He has been told that training is coming. He has been told that for eight months. In the meantime, he has been given a licence and a deadline. He is not resisting change. He is waiting for the lesson that was promised, and each month of waiting makes the thing he has to learn bigger. He will be described, in due course, as slow to adapt.
The Credential Collector. She completes every online module the company offers. The completions are logged, the certificates are saved, and the dashboard shows her as fully trained. None of it has changed how she works, because none of it was connected to her work. She is the output of a system that measures learning by completion, and she knows it, which is why she does not mention it.
The Quiet Skipper. He is the most dangerous of the four, because he looks fine. He uses the tool every day in the way that he first worked out and has never revisited. He is productive enough, confident enough and unaware of what he does not know. Nobody has shown him a better way, so he has no reason to think there is one. He is what a drought produces when it is invisible: a well-fed plant on a diminishing water table.
Notice that none of these four has failed at anything. Each is the predictable product of the conditions they are working in.
The uncomfortable truths
One: upskilling has become an unpaid hobby. When learning is done in the evening, with your own money, because the working week has no room for it, it is no longer a development programme. It is a personal project that the employer benefits from. The organisation gets the capability and none of the cost.
Two: the people who can afford to learn are the people who least need to. Self-directed learning favours those with time, money and confidence. Those are exactly the people who already have the most options. A learning system that depends on individual initiative widens the gap it claims to close.
Three: the front-runners are the flight risk. PwC found that 29% of its front-runners say they are very or extremely likely to change employer within a year. An organisation that expects its most AI-capable people to stay while it spends little on helping them grow is making an assumption the data does not support. They learned on their own. They can leave on their own.
Four: a licence is not a lesson. Access to a tool is the cheapest thing an organisation can provide, and the easiest to count. It is also the least valuable. Anyone who has watched a colleague use an excellent tool badly knows that the difference is training, and training has to be designed, protected and paid for.
Five: the skills gap may be partly a self-inflicted one. Employers regularly report that they cannot find AI-ready people. Some of that is real. But some of it is the result of a long retreat from building skills in-house, now meeting a technology that changes faster than the labour market can supply new graduates. You cannot spend a decade deciding that training is somebody else’s problem and then complain that nobody is trained.
What to do, by role
If you are early in your career. You are the person for whom the drought is most dangerous and the one with the most to gain from handling it deliberately. Ask, before you accept a job, how many hours a month are protected for learning and who pays for the tools and courses. If the answer is vague, treat it as an answer. Once you are in, protect an hour a week as if it were a client meeting, and put it in the calendar before anyone else does. Build something visible each quarter: a workflow, a small tool, a written-up experiment. Evidence of learning travels further than the hours you spent on it. And share what you work out. Teaching a colleague is the cheapest way to turn private learning into shared language.
If you are hiring. Be suspicious of AI fluency as a stand-alone filter. A candidate who is fluent has learned on somebody’s time, and the question is whose. If you screen only for the result, you will select for people who could afford to teach themselves, which is a narrower and more privileged pool than the talent you actually need. Ask what they learned, how they learned it and what they would need to keep learning. And be honest in your own offer. Candidates are good at telling the difference between a learning culture and a learning slide.
If you are leading. Start by changing what you count. Licences issued and modules completed measure the supply of content, not the growth of capability. Count protected learning hours, and count how many people actually used them. Then look at where the money goes. If it goes mainly to the already confident, you are subsidising the people most likely to leave. Deliberately fund the engine room: the large middle that keeps the business running and is currently the least supported. Protect the time, because no budget can buy back an hour that a calendar has already sold. Ask your best AI users what it would take to keep them, before somebody else does. And build a shared practice, with common examples, common standards and common language, so that learning does not remain a private activity with private results.
The water table
A drought does not announce itself. For a long time, the fields look much the same. The plants that were already deep-rooted keep growing, the surface looks green and the reports are good. What is changing is the water table, a little lower each season, until one year the shallow roots find nothing there.
That is the shape of the learning drought. For now, the organisations that have been trading on goodwill and weekends are still seeing results. The Night-Class Learners are still learning, the front-runners have not yet left and the licences are all in use. But the support underneath is thinning, and the people who are drawing on it are drawing harder.
The question for any organisation is simple, and it is the one nobody has put on the slide: who pays for the lesson? If the honest answer is the worker, on a Sunday evening, with the evidence of the survey behind it, then the organisation does not have a learning culture. It has a learning subsidy, paid in reverse, by the people it says it values most.
Every company wants an AI-ready workforce. The unfinished business is deciding who gets to be responsible for making it ready, and then putting their name, their budget and a protected hour in the calendar beside it.