One in three. That is roughly how many workers who used AI at work in the past week say it made a specific task one to two hours faster. The other two in three used the same generation of tools, sat through the same rollout, and came away with something closer to nothing worth mentioning.
That split is not a rounding error. It is the actual shape of what happens when a whole team adopts AI at once, and it looks nothing like the tidy average number that ends up on the town hall slide. An average of moderate time savings across everyone hides a much sharper reality underneath it: a smaller group getting a real win, and a larger group getting very little.
That matters for time because most teams are still budgeting training, support and follow up as if the payoff will land evenly across every role. It will not. Spending equally on the third of people who are already winning and the two thirds who are not is a quiet way of guaranteeing the average stays exactly where it is.
------------- Context -------------
Most organisations rolled AI out the way they roll out any other new piece of software. One licence, one training session, one dashboard tracking who logged in this month, and the same access for a marketing coordinator, a finance analyst and a customer service rep.
The logic behind that approach is reasonable on the surface. If a tool is genuinely useful, more use should eventually mean more time saved for more people, spread out roughly the way usage is spread out. But that assumption quietly treats every task as similar enough to benefit by a similar amount, which is exactly the part turning out to be wrong.
This is where it helps to separate adoption from concentration. Adoption tells us how many people opened the tool this week. Concentration tells us how much of the real time saved is sitting inside a small number of tasks and a small number of people, while everyone else is doing work that AI simply does not speed up nearly as much.
Once concentration becomes visible, the flat rollout looks like the wrong shape entirely. A task that is mostly structured information transfer, drafting a summary, restructuring a spreadsheet, turning notes into an update, hands over cleanly to AI and the time saved shows up immediately. A task that is mostly judgement, negotiation or unique context hands over badly, and the person doing it often spends the saved minutes rechecking or rewriting anyway.
That matters for time because a team's real AI dividend is not sitting evenly across the org chart. It is sitting in a specific, findable set of tasks, and the fastest way to grow it is to go looking for those tasks on purpose rather than waiting for the average to rise on its own.
------------- The Confident Average Number Is Measuring the Wrong Thing -------------
One of the quietest costs in AI rollout reporting is the confident average headline, the "our team saves a set number of hours a week" figure that ends up in the board deck. It sounds precise. It is usually built from a wide spread of individual results, and averaging across that spread erases the exact pattern that matters most.
This is where the one-in-three figure becomes useful. It tells us the real distribution is not a bell curve clustered near the average. It is closer to a small cluster genuinely winning big sitting right next to a much larger cluster winning very little, both getting folded into one smooth number.
Consider a customer support team where the reported average saving is forty minutes a day. Dig into the actual tickets and one group, usually the reps handling standard, well-documented requests, are saving closer to ninety minutes. Another group, handling the ambiguous or emotionally loaded cases, are saving closer to five.
That is a direct time cost hiding inside a time win. The forty-minute average tells leadership the rollout is working. It quietly tells the second group that either they are doing something wrong or the tool simply does not work for them, when the honest answer is that their tasks were never going to behave like the first group's.
The pattern repeats anywhere a team reports one number for a genuinely uneven benefit, from sales proposals to financial reporting to content review.
------------- The Lucky Third Are Usually Doing a Specific Kind of Task -------------
What most rollout plans miss is that the workers landing in that fortunate third are not necessarily the most enthusiastic users, or the ones who sat through the most training. They are disproportionately the ones whose regular work already looks like structured information transfer, turning a call into notes, notes into a summary, a spreadsheet into a cleaner one.
That matters because it means the real lever is not usage frequency, it is task shape. Two people can open the same tool the same number of times a day and get completely different payoffs, because one of them is handing over a task that compresses cleanly and the other is handing over a task that does not.
A finance analyst reconciling a recurring monthly report and an account manager renegotiating a contract renewal might use the exact same AI assistant the same number of times this week. The analyst's task hands over almost entirely. The account manager's task keeps its hardest part, reading the client's hesitation, squarely with the human, no matter how good a draft AI produces.
That is a real time distinction, not a motivation one. The analyst can reasonably expect the time saved to hold up week over week. The account manager should expect a smaller, different kind of saving, mostly in drafting language faster, while the actual negotiation still takes the time it takes.
Knowing which of those two situations a given role sits closer to is more useful than another reminder to use AI more.
------------- Training Spent Evenly Is Training Spent on the Wrong Third -------------
The natural response to a patchy result is to add more training, spread evenly across everyone who has not yet found their win. That instinct is understandable and, based on where the actual time is sitting, usually backwards.
This is where it helps to separate two very different training investments. Training aimed at the third already getting real time back tends to compound, because it reinforces a pattern that already works and can be pushed further. Training aimed evenly at everyone, regardless of task shape, spends the same hour on someone whose task will barely move.
A support team that identifies its highest-leverage ticket types and trains reps specifically on handling those with AI will see its average rise for a defensible reason. A support team that runs the same generic AI workshop for every rep, standard tickets and ambiguous ones alike, will see its average barely move, and will not know why.
That is a time cost that compounds in the wrong direction. Every hour spent training someone on a task that structurally will not speed up is an hour not spent finding the next task that will, and across a team of any size that adds up to a genuinely slower path to real time saved.
The honest fix is not more AI training. It is better targeting of the training that already exists.
------------- Practical Moves -------------
First, pull the actual task-level data behind your team's average time-saved number before repeating it in a meeting, because the average by itself will not tell you where the real win is sitting.
Second, sort your team's regular tasks into the ones that are mostly structured information transfer and the ones that are mostly judgement or negotiation, since that split predicts the AI payoff better than usage frequency does.
Third, target your next round of AI training at the specific tasks most likely to compress cleanly, rather than running one generic session for every role.
Fourth, ask the people getting little from AI what their week actually looks like, because they are often not behind, they are doing work that was never going to behave like the lucky third's.
Fifth, set a realistic, task-specific time-saved expectation for each role instead of quoting one company-wide average, so nobody spends this month chasing a number that was never going to be theirs.
------------- Reflection -------------
The lesson underneath the one-in-three number is not that AI's benefits are overstated. It is that the benefit is real, and it is unevenly distributed in a way that a single average number is built to hide.
That is why the time question worth asking is not how much time our team is saving on average, but which specific tasks are actually producing that time, and who is doing them. Once that is visible, training, support and even role design can follow the real pattern instead of the flattering headline.
In the end, that is the more useful way to read any AI adoption number that sounds too clean. Somewhere behind it is a smaller group doing well and a larger group doing something else entirely, and the real time win only grows once both groups are seen clearly.
Where in your own week does AI genuinely compress a task, and where does it just change what kind of work you are doing?
If you looked at your team's real task-level data, do you think the time saved would turn out to be even, or concentrated in a specific few tasks?
What would you train differently this month if you knew exactly which tasks were producing your team's actual time savings?