This spring, for the first time, most of the frontline said yes to AI. BCG’s June survey of 11,749 employees across 14 markets found 74 percent of frontline workers using AI regularly, a 23-point jump in one year, and among those regular users, four in ten report getting roughly a workday back every week. Then comes the number that explains why so few income statements have noticed: 66 percent of them receive little or no guidance about what to do with the time, and more than half do not reinvest it in anything more valuable. The hours arrive; nobody budgeted them. The constraint on AI returns in 2026 is not the model’s capability. It is whether anyone manages what the capability frees up.
Different sellers, same finding
Treat any single survey with suspicion, starting with its publisher’s incentives. BCG sells transformation, so of course its data says transformation is the gap. What makes this spring’s evidence unusual is that instruments with different owners and different motives converge. Microsoft, which would happily sell you the tools alone, surveyed 20,000 AI-using knowledge workers this winter and found organizational factors (culture, manager support, talent practice) explain about twice the reported AI impact that individual factors do, 67 percent against 32. The same survey caught the incentive gap in one pair of numbers: 65 percent of AI users fear falling behind if they do not adapt, while 13 percent say their employer rewards them for reinventing how work gets done. Fear without reward is a recipe for private, unshared, unaudited use.
Gallup, which sells neither models nor transformation, asked 23,717 US employees in February what separates frequent AI users from holdouts. The strongest correlates were not age, role, or enthusiasm. They were organizational: where people strongly agree AI is integrated into their systems and workflows, 88 percent use it frequently; where managers actively support it, 78 percent do. Employees with supportive managers were nine times as likely to say AI has transformed how work gets done in their organization. Half the US workforce now touches AI at work at least occasionally, up from a fifth in 2023, so the population being mismanaged is no longer a pilot group. It is the payroll.
The Danish control group
The cleanest evidence comes from the one place with a control group and a paper trail. Two economists surveyed roughly 25,000 Danish workers across 7,000 workplaces, twice, and linked the answers to administrative earnings records. Employer initiatives (encouragement, enterprise tools, training) nearly doubled adoption, exactly as the consultancies promise. Realized average time savings still came to 2.8 percent of work hours, against gains above 15 percent for the same occupations in randomized trials, and the measured effects on earnings and hours were zeros precise enough to rule out anything above about one percent. The detail worth pinning to the wall: workplaces that deployed training or enterprise tools in isolation reported lower savings than those that did nothing loud at all, while the full stack together worked. A partial change program is not a partial success. It is overhead.
Perception makes this worse, because the feeling of speed arrives before the speed does. METR ran a randomized trial on experienced open-source developers last year and found AI-assisted tasks took 19 percent longer, while the same developers believed they had been about 20 percent faster (a gap the lab’s early-2026 follow-up shows shrinking, on its own admission as weak evidence). At the economy’s scale the honest number is modest: the St. Louis Fed puts aggregate reported time savings at 1.6 percent of all US work hours, maybe 1.3 points of cumulative productivity since late 2022. Real, and a long way from the demo.
Undisciplined rollouts send you the bill anyway
Skipping the management work does not hold the spend at zero return; it goes negative. The BetterUp and Stanford research on “workslop” measured what happens when tools spread without norms: 40 percent of US desk workers received AI-generated work in the prior month that looked polished and carried no substance, and each instance cost about two hours to untangle, roughly $186 per employee per month. The social ledger runs deeper: in the authors’ summaries, about half of recipients rated the sender less capable and less trustworthy afterward. That is the 13-percent reward gap coming home. People told to be faster, and rewarded for nothing else, produce fast-looking work and mail the cost to a colleague.
Training follows the same arithmetic. BCG’s 2025 wave found regular use jumps sharply after about five hours of training with coaching, and that only a third of employees had been properly trained; by this year 72 percent say the skills expected of them have changed while 36 percent feel adequately upskilled. Five hours is not a moonshot. It is a Tuesday.
What the disciplined minority manages
Who actually gets return on AI is a process story: pick one workflow, rebuild it, measure the before. The people side has its own short list, and none of it is glamorous. Name what the freed hour is for, out loud, per team: the follow-up call, the backlog, the review no one had time to do (the alternative, BCG’s data says, is that the hour quietly dissolves). Make managers the unit of adoption, because the 9.3x multiplier in Gallup’s data does not live in the tool, it lives in the person who assigns work. Pay the five hours of training before judging the tool. And reward visible reinvention rather than raw usage, so the energy behind that 65 percent fear number compounds in the open instead of leaking out as workslop. BCG’s executive study put the durable ratio on it years before the frontline broke through: 10 percent of the effort is algorithms, 20 percent is technology and data, 70 percent is people and process. The 70 was never the model vendor’s job.
The honest counterweight
Read the other side before concluding the sky is falling. Wharton’s survey of about 800 US enterprise decision-makers found 82 percent using generative AI weekly, 72 percent formally tracking ROI, and roughly three in four reporting it positive, though executives rate the same programs rosier than the mid-level managers running them, 81 percent to 69, which is the perception gap wearing a tie. Narrow, capability-led wins are real too: randomized studies keep finding double-digit gains in bounded work like customer support, heaviest for novices, no reorganization required. And the tools are improving fast enough that METR’s slowdown may not survive the year. None of this argues for waiting. It argues that the ceiling is rising on its own schedule, while the floor, the part your company actually stands on, moves only when someone manages it.
So take the question to your own floor. Somebody on your team got Friday afternoon back this quarter. Who decided what it was for?