AI is saving time. Few companies capture it.
BCG’s fourth annual AI at Work survey, published on 3 June 2026, records a full workday a week of self-reported savings among regular frontline users. Independent measurement and Swiss official statistics suggest most of that time is being reabsorbed inside the same working week.
Boston Consulting Group published the fourth edition of its global AI at Work survey on 3 June 2026: 11,749 workers across 14 markets. One finding travelled further than the rest. Among frontline employees — staff without managerial responsibilities — who use AI regularly, 42% report saving eight hours a week, the equivalent of a full working day.
A second finding travelled far less, and matters more. Sixty-six percent of those employees receive limited or no guidance on what to do with the time, and more than half are not reinvesting it in more strategic work. BCG’s own summary is direct: the time individuals save “leaks out of the organization” unless it is tracked and deliberately reinvested.
That turns a technology question into an accounting one: what matters is not how many hours a tool saves, but how many arrive somewhere the company can see.
Adoption has stopped being the binding constraint
Seventy-four percent of frontline employees now say they use AI daily or several times a week, 23 percentage points more than in 2025. For the past few years it had sat at around 50%, the plateau BCG calls a silicon ceiling; the break came mainly from older workers, operational roles and lagging markets. The share reporting the full day saved is higher in support functions: 60% in marketing, 53% in IT, 50% in human resources. Agents follow one step behind: 30% of respondents say agents are already integrated into workflows, against 13% a year earlier, and half say their company lacks clear governance for teams of people and software.
For a Swiss SME, the licensing decision has largely been taken, often informally and from the bottom up. What remains undecided is organisational.
Reported hours and measured hours are not the same quantity
The eight hours are self-reported, and studies that measure rather than ask return smaller numbers.
Anders Humlum of Chicago Booth and Emilie Vestergaard of the University of Copenhagen ran two adoption surveys in Denmark — about 25,000 workers in 7,000 workplaces across eleven exposed occupations, from accountants to teachers — linked to matched employer–employee administrative records. Users reported average time savings of 2.8% of work hours. Difference-in-differences estimates for earnings and recorded hours were precise zeros; the version revised in March 2026 rules out effects larger than 2% two years after ChatGPT’s launch, and only 3% to 7% of workers’ reported time savings translate into earnings.
Alexander Bick, Adam Blandin and David Deming found, in nationally representative US surveys, that between 1% and 5% of all work hours are assisted by generative AI, with reported savings equivalent to 1.4% of the total.
The gap between perception and measurement is the uncomfortable part. In a randomised trial at the February–June 2025 tool frontier, METR had sixteen experienced open-source developers work through 246 tasks in their own repositories. They took 19% longer when allowed to use AI tools, and estimated afterwards that AI had made them 20% faster. METR now labels that result out of date. Its February 2026 follow-up, across 57 developers and 800-plus tasks, points the other way: estimated time reductions of 18% and 4% for the two cohorts, with confidence intervals spanning zero, which METR calls very weak evidence and a likely lower bound, since developers increasingly declined to submit tasks they did not want to attempt without AI.
Populations and dates differ: BCG surveyed regular users in 2026, the Danish rounds ran in late 2023 and 2024, METR covers one profession with tools that keep changing. What these studies establish is narrower than a verdict on the technology, and more durable: asking people how much time they saved is a generous instrument, and eight hours against 2.8% is not a rounding difference.
Swiss aggregate hours have not moved, which is what an uncaptured saving looks like
Switzerland offers a clean scoreboard. On 21 May 2026 the Federal Statistical Office reported that 8.114 billion hours were worked in the country in 2025 — unchanged on 2024, with jobs up 0.3% and actual annual hours per job down 0.3%. Actual weekly hours of full-time employees stood at 40 hours and 3 minutes, still below the 40 hours and 54 minutes of 2019; on the internationally comparable basis they worked 42 hours and 24 minutes, the highest figure in the EU/EFTA area against an EU average of 37 hours and 54 minutes (across all employed persons, 35 hours 14 minutes against an EU average of 35 hours 24 minutes).
Set the survey figure against that denominator: eight hours is a fifth of a Swiss full-time week. Swiss labour productivity grew by about 1% a year on average between 2000 and 2022, close to the 1.1% in the EU and the OECD area, and the aggregates show no break.
The reconciliation is not mysterious: time saved on a task is reallocated inside the same week rather than released, and BCG’s respondents describe the mechanism from the inside. Sixty-seven percent say AI has taken over simpler tasks and left them with more complex ones, 60% say the bar for work that counts as good enough has risen, and 47% say they now spend more time managing and directing AI than doing the work itself. The gross saving is real; the net saving is smaller, and it lands where no one is counting.
The firms reporting impact changed a process, not a licence
BCG’s differentiator is strategic clarity rather than tooling: a clear strategy lifts measurable business impact by 25 percentage points, better tools without one by roughly 5. Respondents in companies pursuing end-to-end workflow redesign are 24 percentage points more likely to report measurable business improvement, and 22 points more likely to report saving a full day a week. The share of organisations using AI to reshape workflows end to end or build new business models nearly doubled year on year, from 22% to 42%.
These are associations in a self-reported survey, not causal estimates: firms that redesign workflows differ in many other ways. The Danish study, quasi-experimental in design, reaches a compatible conclusion: where employers encouraged use, deployed in-house models and trained staff, take-up nearly doubled from 47% to 83% and reported benefits, time savings included, were 10% to 40% greater.
The complement to the tool is management, and it is not being supplied. Seventy-two percent of BCG’s respondents say skills expectations in their role have shifted, only 36% feel adequately reskilled, and only a third of frontline employees find leadership’s communication on AI clear.
What a management team can decide this month
- Pick one process rather than one tool: quotation to order, invoice to payment, service request to resolution. Measure it end to end — elapsed days, handovers, rework — before anything changes.
- Name the destination of the freed hours in advance: faster customer responses, more quotations, quality checks, structured training. “More capacity” is not a destination.
- Net the savings instead of grossing them: ask what now takes longer — reviewing output, correcting it, deciding — and subtract it.
- Fund the complements, not only the subscriptions: the Danish evidence points to encouragement, in-house deployment and training, all within reach of a fifty-person firm.
- Convert freed hours into something the accounts recognise. Swiss employers reported 98,200 vacancies in the first quarter of 2026, 5.0% more than a year earlier, with 34.3% reporting recruitment difficulties, down two points on the quarter. Freed hours are worth most where the alternative is a hire that cannot be made.
- Write the governance down before agents arrive: who reviews output, who signs it off, what is logged.
The test is unglamorous, and it has dates
Two calendars matter. The company’s own comes first: one redesigned process, reviewed quarterly in elapsed days and orders handled per person, not seats licensed. The public one is the check — the Federal Statistical Office publishes hours worked for 2026 in May 2027, and a fifth BCG round would fall around mid-2027, where the guidance figure, 66% today, is the number to watch.
The first wave of this technology was distributed to individuals and measured by adoption. The second is organisational, and the honest test at year end is narrow: can the company point to a process that now takes fewer days, a customer answered sooner, or work delivered without a hire it could not make? If not, the hours were saved. They were simply never collected.