Why Current Output Can Stay Stable While the Pipeline Weakens
The subject here is the evidence for a moderate hybrid-work policy, and the specific question is what that evidence can tell us about established employees without pretending it has tested first-job training. A randomized experiment at Trip.com gives a rare chance to separate retention, measured performance, and development. Reading its design closely supports a precise judgment: two home days can make an existing workforce more likely to stay without lowering the outcomes the firm measured, while the experiment’s excluded newcomers remain the crucial boundary.
What the Trip.com experiment actually tested
The study took place at Trip.com, a large travel-technology company headquartered in Shanghai. In the summer of 2021, the company invited 1,612 employees in its Airfare and IT divisions to participate. They worked in engineering, marketing, accounting, finance, and related business services; all had university degrees, and the typical employee was in their mid-30s. There were 395 managers and 1,217 non-managers in the experiment.
The policy was specific. Employees eligible for the hybrid group could work from home on Wednesday and Friday and come to the office on the other three weekdays. Employees in the comparison group were expected to work in the office all five days. Working from home was an option, not an obligation. That detail matters: the experiment tested the effect of being offered eligibility for a moderate hybrid schedule, not the effect of forcing every participant to spend exactly two days at home.
The company used an odd-even birthday rule to create the comparison. People born on an odd day of the month were assigned to the hybrid-eligible group; those born on an even day were assigned to the office-only group. Birth date within the month is not a plausible reason for one employee to be better at their job than another, so the rule produced groups that were similar on average before the policy. This is the value of random assignment. If the groups later differ, the policy is a more credible explanation than the possibility that one group was already more motivated, more skilled, or more likely to quit.
There was an important complication. The company first asked employees to volunteer for the experiment, and only 518 volunteered. Management suspected that some employees feared that volunteering for home work would signal lower ambition or commitment. The remaining employees were then included, which made the final comparison broader than a self-selected group. Take-up was also incomplete: even among eligible employees, many worked from home on one of the two available days rather than both. The result should therefore be read as the effect of offering a workable hybrid option in an organizational setting, not as a pure comparison of two perfectly observed weekly schedules.
That distinction between eligibility and use is analytically useful. If only enthusiastic volunteers had been compared with people who stayed in the office, the result could mix the policy with pre-existing differences in commute, preferences, family obligations, or confidence about signaling. Randomizing eligibility after the volunteer and non-volunteer groups were included makes the comparison less dependent on those choices. It also captures a real management question: what happens when an organization makes a form of flexibility available, knowing that employees will use it unevenly? A policy can create value without every eligible employee using every permitted day, but its effect should not be described as the consequence of a uniform two-day absence from the office.
The researchers also excluded interns and rookies who were still in probation. The reason was explicit: on-site learning and mentoring were particularly important for those employees. This is not a minor footnote. The people most dependent on the feedback loop described earlier were not part of the population used to establish the trial’s main performance conclusion. The experiment tells us a great deal about what happens when an existing graduate workforce receives more flexibility. It tells us much less about how a person learns a first role when they do not yet know which questions to ask.
The design can be summarized without turning it into a verdict about remote work as a whole:
The experiment’s strength and its limitation are therefore the same fact: it isolates a clearly defined policy for a clearly defined group. A result can be strongly causal for that group while still being narrow in application.
A clean result with a sharp edge
The clearest benefit appeared in retention. During the six-month experimental period, attrition was 7.20% in the office-only comparison group and 4.80% among employees eligible for hybrid work. The researchers describe that as a one-third reduction in quitting. Job satisfaction also rose: the average score increased from 7.84 in the comparison group to 8.19 in the hybrid group. Participants reported saving commuting time and costs and gaining flexibility for occasional personal tasks.
The effect was not identical for every subgroup. Among non-managers, attrition fell from 8.59% to 5.33%, a reduction of roughly 40% relative to the comparison group. Among women, it fell from 9.2% to 4.2%, a reduction of about 54%. The trial also found a larger reduction for employees with longer commutes, although one of the long-commute comparisons did not reach conventional statistical significance. These patterns make practical sense: the value of removing a long commute or gaining control over the workday is not evenly distributed. They should still be treated as subgroup results from this company, not as a promise that every woman, manager, or long-distance commuter will respond in the same way.
Retention is not a consolation prize. Replacing an employee consumes recruiting time, institutional knowledge, and the attention of people who must train the replacement. A policy that reduces unwanted quitting can create value even if it leaves each remaining employee’s measured output unchanged. It can also change who is willing to remain in a job: the option to work at home may be particularly valuable to people balancing commuting, family responsibilities, or the physical strain of traveling to work every day.
The more controversial question is performance. The researchers followed employees through four six-month review periods covering two years after the experiment began. The review process at Trip.com was not a single manager’s impression. It combined assessments from managers, project leaders, peers, and sometimes direct reports or customers, and it was linked to pay and career progression. Across the review periods, the researchers found no meaningful difference between the hybrid-eligible and office-only groups.
That conclusion is more careful than saying, “The study found no statistically significant effect.” A conventional significance test asks whether the data provide enough evidence to reject a zero difference. A non-significant result alone can also occur when a study is too small or too noisy to detect a real effect. The authors therefore used equivalence tests. For performance grades, they defined half a letter grade as the smallest effect that would matter for the question at hand and tested whether the true effect could plausibly be larger than that in either direction. The tests supported interpreting the performance effect as practically null within those bounds.
In ordinary language, the test asks whether the data are compatible with a difference large enough to change a manager’s practical conclusion. It does not ask whether every individual received exactly the same evaluation, nor does it rule out effects smaller than the chosen boundary. A hybrid employee could perform slightly better or worse in a way that would not amount to half a letter grade, and the study would still support its stated equivalence conclusion. The value is that the researchers did not turn an inconclusive “no difference detected” into a confident “nothing happened”; they defined what size of difference the evidence could meaningfully exclude.
The same caution applies to the other measures. Promotions are relatively infrequent events, so a similar average promotion rate does not describe every person’s career path. Lines of code are an output measure for the computer-engineering subgroup, not a complete measure of design quality, collaboration, learning, or long-term capability. The study’s several measures point in the same direction for current performance, but they are still proxies observed after employees have entered the organization. None directly records how much explanation a beginner needed before producing work that could be reviewed on those terms.
The distinction is worth keeping. The evidence does not prove that hybrid work produces exactly the same performance for every employee. It supports the narrower claim that, in this setting and over the measured review periods, any effect on performance grades was smaller than the study’s meaningful threshold. That is evidence of a bounded absence of practically important difference, not a metaphysical proof of zero effect.
The same pattern held across other measures. The study found no evidence of a promotion difference overall or across major employee subgroups. It also examined nine performance subcategories, including development, innovation, leadership, and execution, without finding a material treatment difference. For 653 computer engineers, the researchers compared daily lines of code submitted and found no difference under their equivalence analysis. These measures do not all mean the same thing, but their convergence makes it harder to sustain the simple claim that two home days automatically damage current execution.
Managers’ beliefs moved as well. Before the experiment, managers expected hybrid work to reduce productivity by an average of 2.6%, while non-managers expected a small improvement. After employees had experienced the policy, managers’ average assessment shifted to a 1.0% improvement and was no longer distinguishable from non-managers’ view. Experience changed the organization’s judgment about the policy. That does not prove that every organization would update in the same direction, but it illustrates why a real test can be more informative than a debate conducted entirely through assumptions.
What, then, should we infer about development? First, the trial weakens a broad productivity objection. A three-days-in-office, two-days-eligible-at-home schedule improved retention and satisfaction while leaving measured reviews, promotions, and code output practically unchanged for the participating workforce. Second, it does not answer whether a new graduate can acquire standards and judgment as efficiently under the same arrangement. Interns and rookies were excluded precisely because their learning needs were different.
This difference is not semantic. Performance reviews observe the work of people who have already crossed an important threshold: they know the organization’s language, tools, expectations, and social routes for getting help. A review can tell us whether those employees are meeting established standards. It is a much less direct measure of how quickly a beginner discovers the standards, how often they make avoidable mistakes before anyone sees them, or whether managers decide not to hire them because teaching them feels too costly.
The tested schedule also preserves substantial shared time. Employees were expected to be in the office three days a week, and the company’s teams worked together in modern open-plan offices. The authors explicitly say they are uncertain whether the results would extend to more remote settings, such as one day a week or less in the office, because training, innovation, and culture may become more difficult. That uncertainty is not an argument against hybrid work. It is a reminder that remote intensity is part of the treatment.
Nor is the trial a perfect model of every hybrid arrangement. It took place at one Chinese technology company in 2021–22, involved university graduates in particular divisions, and measured outcomes in a period shaped by the pandemic and its aftermath. The option was offered rather than mandated, and take-up varied. Those facts do not cancel the causal result; they specify where it is strongest. To apply it elsewhere, a manager would need to ask whether their team has similar review systems, task visibility, shared-office rhythm, employee composition, and ability to make early learning observable.
The disciplined conclusion is therefore two-sided. Hybrid work can be a genuine retention benefit and need not lower the current output of established employees. At the same time, the evidence that supports that conclusion is least informative about the people at the beginning of the learning curve. The next question is not whether the trial was positive or negative. It is how remote work became a varied, persistent job design—and how the distribution of those arrangements determines who receives flexibility, who needs proximity, and where the development risk is most likely to appear.