AI does two things to your team's work. It makes some activities faster, and it lets people do things they could not do before. Hours-saved figures count only the first. Both change the job, and in most teams nobody has decided what the job becomes. This note looks into the data on both, and shows how to find out what is happening in your own unit, in two learning actions.
Ask a team leader where the day a week went that AI was supposed to give back. The honest answer is a list: twelve minutes on a summary, half an hour on a draft, an hour checking what the tool produced, and a meeting that filled the rest. The time was saved. It never arrived as a day. And the list leaves something out: the analysis the team leader now runs herself, the client proposal an engineer drafted on his own. That is not time saved. It is different work in the same time.
What AI does to time plays out at three levels, and it does not travel up on its own. AI changes activities: it does some faster and makes others possible. Whether that becomes a better job depends on how the job is designed. Whether it shows up for the organization depends on the business model. The leader's decision sits in the middle, and it comes with a second one: what counts as good enough for work nobody wrote alone.
The headline figures measure faster activities. In BCG's 2026 survey, about half of regular AI users say they save a workday a week or more (BCG, AI at Work 2026). Studies that follow the time further find far less: around 3% of working hours in Danish payroll data, with no effect on earnings (Humlum and Vestergaard), and 5.4% for US users (St. Louis Fed). Both studies measured 2023 and 2024, when most people used early chat assistants. The tools have moved a long way since, so the gap may well have narrowed. What has not changed is that time saved on a task still has to find its way into the job.
The second effect is harder to count, because nobody asks about it. In the Danish data, 8.4% of workers said the tool had brought them new tasks. In a field experiment with 776 professionals at Procter & Gamble, people working alone with AI did as well as two-person teams without it, and R&D and commercial specialists stopped proposing only from their own corner: with AI, both came up with ideas that balanced the technical and the commercial side (Dell'Acqua and colleagues).
Between activity and job, the minutes leak. In Workday's survey of 3,200 employees, nearly 40% of the time saved went into rework (Workday, 2026; vendor research). And one person's saving becomes the next person's cost: 40% of US desk workers had received "workslop" in the previous month, work that looks finished and is not, at about two hours per incident (Harvard Business Review). The risk is highest in the new work, where people produce output in a field they were never trained to judge. The deeper risk sits in the job itself. Work motivates when it offers a whole piece of work and feedback from the work (Hackman and Oldham). If the tool takes the first draft and the first analysis, it may take exactly those parts. What is left is checking, and checking without feedback is how quality drops quietly.
For the firm, it depends on what it sells. In software, revenue has pulled away from headcount; in banks and professional services, the two still move together (a recent chart by the FT's John Burn-Murdoch). Most of that predates the current tools. But where output still grows with people, the saving will not show up as fewer hours on the books.
Our position: as long as the organization around it stays the same, the unit of AI's payoff is the job, not the activity. In many places, that condition has started to change: teams are cut differently, roles are merged, and whole steps between people disappear. Where that happens, the payoff sits in how work is divided between people, not in the single job.
Either way, if nobody decides what the work becomes, the minutes go to volume, rework and the next meeting, the new work goes unchecked, and the parts of the work that taught people their craft disappear. Our last note made the same point about resistance: people push back on being redesigned without a say. A tool can redesign a job without anyone saying anything.
Working on this in your own organization? See how the MK Practice works →
Not by asking the team to "use the time strategically". It is a question of organizational design. For now it sits at the level of the role; how work is divided between roles and units is the next question, and it is already on the horizon in many organizations. Both start with learning actions on your own unit, and the first two are the same for every team.
For a team whose saved time had gone mostly into checking, the path went on like this:
A team or role whose time went into volume needs a different path: a decision, by whoever owns the targets, on what the extra capacity is for. A team that now does new work needs another: a decision on whether that work belongs in the job, and who can tell good from bad in it. In our work with leadership teams, the second action surprises people most. Leaders expect to find idle time. They find that the tool took the part of the job their best people learned from.
This is pragmatic work at team level: a closer look at the activities, and adjustments to one role at a time. It is not a full AI-powered redesign of the organization, and it does not need to wait for one. In the MK Practice, these actions sit in the working topic on activity analysis in our Organizational Structure kit, and go onto your work agenda for review. For the feedback side of the job, see our post on why over a third of feedback backfires.
A decision rule: if your team says the tool saves time and nobody can say where it went, it went nowhere. If a quarter passes and the job looks the same, the saving was absorbed, not invested.
So before the next all-hands celebrates the hours AI saved, ask three people where theirs went, and what they now do that they did not do before. It takes 45 minutes and tells you more about your rollout than the licence count. The same holds for how leaders learn: if AI takes the generic part of a curriculum off the table, the hours left are for what only your own unit can teach you. The next note is about what your rollout plan is missing: a story.
You don't read a practice. You run it.
The MK Practice is a structured process you run on your own organization: a diagnostic that reads where your unit needs work first, working topics like the ones in this series, a work agenda that keeps the moves coming, and a practitioner who looks at your problem within 48 hours. Founding Member: USD 990 a year, 30-day money-back guarantee.
Want to talk it through first? Book 20 minutes with a Management Kits team member. Running it with a whole leadership team? See the MK Practice for organizations.