Bristol Myers Squibb is exploring a "well-being quotient" to decide how much of the productivity its workers gain from AI should return to them rather than the company, executive vice president and chief people officer Amanda Poole said at the Fortune Most Powerful Women conference. "Let's say we're going to get 20% productivity with AI: What part are we going to take back as a company, and what part are we going to give back?" Poole asked, calling the figure hypothetical and adding, "We haven't solved it, so more to come." The question, raised alongside peers from PepsiCo, Indeed and Dropbox, marks a shift in the corporate AI conversation from whether the tools raise output to who keeps the output they create.
The gains being divided are real and, at some firms, large. Indeed says AI coding assistants help its software engineers write 50% more code and save them more than six hours a week, with overall engineering productivity measured as more than 50% higher. But that output has not translated cleanly into time returned to workers. At Dropbox, 85% of the time saved with AI went straight back into work activities, and only 49% of employees said they directed saved time toward higher-value work. "Time saved does not necessarily equal value created," Dropbox chief people officer Melanie Rosenwasser said.
The strain is showing up in the people doing the work. US employees who frequently use AI report higher burnout than those who do not, according to 2026 data from Quantum Workplace, and across survey waves from 2023 to early 2026 daily AI users were more than twice as likely as infrequent users to say automation was very likely to eliminate their jobs within five years. Dropbox is also examining what Rosenwasser called "cognitive atrophy" as workers outsource thinking to AI, and has introduced memory exercises, attention-span training and device-free offsites in response.
A sharper counterpoint comes from Anthropic's own data on how work has changed inside the company. It reports code commits per engineer rising fivefold in six months, alongside more contracts processed by lawyers and more deals closed per salesperson. Yet median meeting time per employee rose 20% year on year, and "meta-work," the effort of building visibility tools and monitoring agent pipelines, climbed from 10% of interactive sessions in autumn 2025 to 22% by spring 2026. One employee described facing "an army of Claudes" each demanding attention. The productivity dividend, in other words, arrives with a coordination tax attached.
That reframing is spreading among the firms that sell to employers. At Workday Rising, the company introduced Work Intelligence, which breaks jobs into tasks and maps which should be automated, augmented or left human; its talent-management VP Nikki Helmer said "the basic building block of work is shifting from the job to the task." Analyst Josh Bersin's new Workday partnership, branded HR 2030, argues that agentic work needs new operating models, not just software. Both land on the same unresolved question Poole posed: whether freed capacity is reinvested in better work and development, or simply absorbed as a higher baseline for the same headcount.
None of this is settled, and the executives were candid that it is not. What has changed is that the debate is now explicit. For two years the pitch was that AI would automate the dull work and hand the time back; the evidence this month suggests most of that time is being reclaimed by the employer, and that managing a fleet of agents is itself becoming a demanding kind of work. The firms treating the split as a deliberate decision, rather than a default, are the ones their own staff will be watching.