Arguments about remote work tend to be conducted at the level of the whole company. Bring everyone back, or let everyone stay home. The strongest evidence does not support that framing. It shows that the gains from distributed work are real, but that they land unevenly across roles and across people. The right question is rarely whether a firm should be remote. It is which jobs travel home well, and for whom.
The gains are concentrated, not uniform
The randomized trial at Trip.com is usually summarized in one line: two days at home cut resignations by about a third with no loss in performance (Bloom, Han and Liang, 2024). That summary hides the more interesting structure underneath. The retention effect was not spread evenly across the workforce. It was larger among non-managers, among women, and among employees with long commutes (Bloom, Han and Liang, 2024). The sample itself was mixed by design, roughly 1,200 non-managers alongside close to 400 managers in the company's Shanghai operation (Stanford, 2024). Averaging across such different groups produces a headline number that describes almost no individual accurately.
That pattern has a plain reading. Hybrid work is most valuable to the people for whom the office imposes the heaviest tax, a long commute, caregiving load, or a role whose output does not depend on constant informal contact. For a manager whose day is coordination, the calculus differs. Designing one policy for all of them guarantees it fits some badly.
The job description matters as much as the person
The earlier Ctrip experiment sharpens the point because it studied a very specific kind of work. The subjects were call-center employees, whose output is close to fully measurable in minutes worked and calls handled. In that setting, working fully from home raised productivity by about 13 percent, roughly nine points from more minutes per shift and four from a quieter environment (Bloom, Liang, Roberts and Ying, 2015). A quiet room and no commute help most when the task is individual, measurable, and interruption-sensitive. It would be a mistake to read that number as a general law of remote work. It is a result about a job that happens to suit remote work almost perfectly.
Different tasks sit elsewhere on that spectrum. GitLab, which runs an all-remote software workforce, has built its entire operating model around documentation precisely because knowledge work generates constant follow-up questions that would otherwise block colleagues for hours (GitLab handbook, 2026). That is a self-declared case rather than a measured comparison, and worth treating as one. But it illustrates the same principle from the other side. Where a job is dense with interdependence, remote work does not fail, it simply requires more infrastructure to make the coordination explicit. The engineering task and the call-center task are both doable from home. They are not doable from home in the same way.
Autonomy where measurement allows it
The practitioner literature reaches a compatible conclusion by a different route. Matt Mullenweg, who runs the distributed company Automattic, describes a ladder of distributed-work maturity in which the higher rungs are defined by judging people on output rather than hours or process (Mullenweg, 2020). That is an opinion from a working founder, not a study, and his stronger claim that a fully distributed organization beats any in-person one is offered without evidence. Still, the operative insight is sound and matches the trials. You can extend autonomy exactly as far as you can measure results. Roles where output is legible can be pushed furthest. Roles where it is not need either better measurement or more overlap before they go the same distance.
A caveat holds all of this together. The trials ran inside single Chinese firms, one in call centers and one among graduate professionals, so the subgroup effects are strong evidence about those settings rather than precise coefficients for every company. What generalizes is not the size of any one number. It is the shape of the finding: gains that vary by role and by person.
The practical read
A company that wants the retention benefit the evidence describes should stop looking for a single answer to a single question. The useful exercise is a segmentation. Which roles have output the firm can measure without watching? Which workers carry the heaviest office tax? Where is the interdependence dense enough to demand real documentation before it goes remote? The Trip.com result is a mandate to disaggregate, not to standardize. A blanket policy is attractive because it is simple to announce and simple to enforce, but simplicity is exactly what the evidence argues against. The retention that the trial documented was concentrated, and a company-wide switch spreads a benefit thinly that a targeted design could deliver where it counts (Bloom, Han and Liang, 2024). The firms that get the most from distributed work will be the ones that match the arrangement to the job, because the evidence was never really about where people sit. It was about who they are and what they do.
Sources
- Bloom, Han and Liang (2024). Hybrid working from home improves retention without damaging performance (Nature 630). The retention gain was larger among non-managers, women, and long-commute workers, the core evidence that remote benefits are uneven by group. https://www.nature.com/articles/s41586-024-07500-2
- Bloom, Liang, Roberts and Ying (2015). Does Working from Home Work? Evidence from a Chinese Experiment (QJE 130). The +13% result in a highly measurable call-center role shows how task type determines how well a job travels home. https://academic.oup.com/qje/article-abstract/130/1/165/2337855
- Stanford (2024). Hybrid work is a win-win-win for companies and workers. Documents the sample composition (about 1,217 non-managers and 395 managers in Shanghai), supporting the argument that averages hide role-level differences. https://news.stanford.edu/stories/2024/06/hybrid-work-is-a-win-win-win-for-companies-workers
- Mullenweg (2020). Five Levels of Autonomy. A practitioner taxonomy in which higher autonomy tracks measurable output, aligning with the article's point that roles differ in how far they can go remote. https://ma.tt/2020/04/five-levels-of-autonomy/