Eight points. That is how far access to learning resources fell in a single year, from 59 percent of workers to 51 percent, according to a new global survey of nearly 50,000 people.
What makes that number strange is the one sitting next to it. Over the same twelve months, daily use of generative AI at work rose from 14 percent to 22 percent. More of us are using these tools every single day, and fewer of us are being given the courses, the guidance or the working hours to learn how to use them well. Those two lines are moving in opposite directions, and somebody has to close the gap between them.
That somebody is usually the individual, in their own time. It is a time issue, and an unusually hidden one, because those hours never appear on a timesheet, a project plan or a training budget. They get spent on Tuesday evenings and Sunday mornings, which is why so many of us feel that AI saves time at work while quietly costing time outside it.
------------- Context -------------
Most of us were told that AI would be easy to pick up. Open a chat window, type a sentence, get something useful back. For the first week that is largely true, and the speed of that first result feels like proof that nobody needs any training at all.
The trap arrives after the first week. A useful first answer is easy. A reliable answer, on a real task, with real context, in a way that fits how a team already works, is a different skill altogether. Nobody calls it training because it mostly happens as tinkering: a tutorial video at 9pm, a prompt copied from a post, three attempts at the same task until one comes out right.
This is where we think the more meaningful shift sits. Learning to work with AI has become part of the job without becoming part of the schedule. Organisations are counting the time the tools save while ignoring the time people spend becoming good enough for the tools to save anything. Both sides of that ledger are real, but only one side is being recorded.
That matters because a time saving that depends on unpaid evening study is not really a saving. It is a transfer. The hours move out of the working day and into the margins of our lives, and the margins are where we have the least left to give. Counting the learning hours is the first step towards making the time gain real.
------------- The Learning Is Real Work, It Just Happens Off the Clock -------------
One of the quietest costs in AI adoption is the time people spend teaching themselves. It rarely gets named, so it rarely gets managed. It sits in the gap between "we have rolled out the tool" and "people actually know how to use it", and it is paid for entirely in personal time.
Consider an operations coordinator who is asked to start using AI for the weekly reports. There is no course and no time set aside. She spends about forty-five minutes on each of three evenings watching walkthroughs, testing prompts on old reports, and rewriting them when the output comes back wrong. By the time the new process works, she has become the unofficial expert, and nobody has noticed what it cost.
That is roughly two and a quarter hours a week, or well over a hundred hours across a working year, spent on a skill her employer is already counting on. Even if guided practice during the day only halved it, that is more than fifty hours a year handed back to her evenings.
This is a direct time loss that never shows up in the numbers. Until the learning hours are visible, the headline claim that AI saves hours each week is only half of the sum.
------------- Learning Alone Means Learning the Same Lesson Five Times -------------
The second cost comes from how the learning is organised, or rather how it is not. When each person works things out privately, the team pays for the same discovery again and again. Nothing is wrong with any individual effort. The waste lives in the lack of a shared starting point.
Picture a team of five that each works out their own way of turning a meeting recording into a usable summary. Each person spends about forty minutes getting to a version they trust, testing how much context to paste in and which parts of the output need checking. Together, that is three hours and twenty minutes of effort, spent five times over on one routine.
Now imagine one person does the forty minutes and records the result as a ten-minute walkthrough for the others. The team's total drops to about ninety minutes, which is a saving of nearly two hours on a single workflow. Multiply that across the handful of routines a team relies on and the pattern becomes hard to ignore.
The time connection is plain. Shared learning is not a nice cultural extra. It is the cheapest source of reclaimed hours in the whole adoption process, and it costs very little to start.
------------- Hesitation Is the Most Expensive Symptom of Thin Training -------------
The third cost is harder to see because it looks like normal working. When we learn in fragments, we pick up which tasks seem to work but not why they work. That leaves us unsure how much to trust our own judgement, and unsure is slow.
Take a marketing manager who knows AI can draft a client update but is not certain how much background to give it or what to check afterwards. A task that should take ten minutes takes twenty-five, because fifteen of those minutes go on rewording the request and second-guessing the answer. She does this about four times a week, which adds up to an hour of pure hesitation.
None of that hour is spent on the work itself. It is spent on the uncertainty around the work, and that is precisely the part a short worked example from a colleague removes. Seeing one real request, one real output and one real round of checking turns a vague worry into a repeatable habit.
This is time-to-first-useful-draft in action. Training does not only teach new skills. It shortens the pause before we trust ourselves to start, and that pause is multiplied across every task we hand to AI.
------------- Protected Learning Time Pays Back Fastest When It Is Attached to Real Work -------------
If the learning hours are real, the obvious question is where they should live. Our instinct is often to reach for the training day, the course or the workshop. Those have their place, but they tend to teach in general terms, and general terms are hard to turn into Monday morning.
A better pattern is smaller and closer to the work. Imagine a team that sets aside thirty minutes every Friday to improve one recurring task with AI, using a real example from that week. After a month, they have four improved routines and have invested two hours in total. If each routine saves just twenty minutes a week, that is eighty minutes back every week afterwards.
The investment is repaid in under two weeks after the fourth session, and it continues paying every week from then on. More importantly, the time sits inside the working day, where it can be seen, protected and counted.
That is the shift worth making. Learning time that is scheduled and attached to a real task turns the hidden evening hours into visible, repaid ones.
------------- Practical Moves -------------
First, estimate how many hours you personally spent learning AI over the past month, evenings and weekends included, and write the number down. That figure is the real cost of the time you believe AI is saving, and it cannot be managed while it stays invisible.
Second, ask your team which AI task took longer to learn than the old way took to do. Those tasks are where a shared walkthrough will repay the effort first, because everyone is probably still paying for it separately.
Third, whenever someone finds a prompt or routine that works, spend ten minutes recording it as a short written or video example. The next person then skips the forty minutes of trial and error and starts from a working version.
Fourth, protect thirty minutes a week inside the working day for improving one recurring task with AI. Put it in the calendar like any meeting, because learning that depends on spare evening energy tends to be the first thing to disappear.
Fifth, after a month, compare the minutes saved on those tasks with the minutes spent in the sessions. A simple tally shows whether the learning is paying back, and it gives you the evidence to keep protecting the time.
------------- Reflection -------------
The gap in that survey, with daily use climbing while access to learning falls, is a signal that adoption has outrun support. The tools arrived quickly and the teaching did not. What we are seeing is that the difference gets absorbed by individuals, quietly, in their own hours, and then described as productivity gained.
That is why counting the learning hours matters so much. It ties AI adoption directly to time, in both directions. The hours saved are real, but so are the hours spent earning them, and a fair account of the first can only be made once the second is on the table. When the learning moves into the working day and gets shared, the net gain grows and stops depending on our evenings.
In the end, that is the kind of AI shift worth paying attention to. Not simply more people using more tools, but how much of the cost of getting good at them is being carried by the people least able to see it. Make that cost visible, and a surprising amount of time comes back.
Roughly how many hours of your own time have gone into learning AI that no one at work ever counted?
Which task did you teach yourself that a colleague is probably still working out alone?
If thirty minutes a week of protected learning time appeared in your calendar tomorrow, which recurring task would you point it at first?