Return-to-office mandates keep colliding with employee demand for flexibility. The argument circles the same drain: culture versus productivity, mentorship versus autonomy. The cleanest experiments available point to a better question. Use hybrid work as a stress test for asynchronous, output-focused management. If you can make two days at home work on the numbers, you are building the discipline that distributed organizations run on.
What the gold-standard trials actually say
In a randomized controlled trial inside Trip.com, roughly 1,600 university-trained professionals were assigned by birthday to work from home two days a week or remain fully in the office. Over two years of follow-up, the hybrid group’s voluntary quit rate fell from 7.2 percent to 4.8 percent, a one-third reduction that reached statistical significance, and there was no detectable difference in performance ratings or promotions between groups (Bloom, Han and Liang, 2024). This is unusually strong causal evidence. It is also specific: a single Chinese tech firm, a defined hybrid cadence, and employees with similar educational backgrounds. It tells us what a managed hybrid model can do, not what any remote policy will do by default.
A decade earlier, an experiment at Ctrip, another Chinese travel company, tested full-time work from home in call centers. The results were stark. Home workers increased performance by 13 percent, most of it from more minutes worked per shift and less from a quieter environment, reported higher satisfaction, and saw attrition drop sharply. Yet their promotion rate, conditional on performance, fell relative to office peers. When the firm later let everyone choose their preferred arrangement, the measured productivity gains rose to about 22 percent, combining selection and learning effects (Bloom, Liang, Roberts and Ying, 2015). The lesson is double-edged. Remote work can lift output, but visibility bias and one-size policies can undercut careers and waste potential.
Taken together, the two trials isolate what matters. Hybrid, run deliberately, protected retention without harming performance or promotion prospects. Fully remote, imposed rigidly, risked a penalty to advancement even while boosting output. In managerial terms, the real variable is the operating system: how you define, coordinate, and evaluate work.
Async is not a perk. It is an operating system.
The Trip.com study did not succeed by accident. Its authors later spelled out three conditions that underpinned the result: rigorous performance management, coordinated team schedules rather than ad hoc days, and visible support from senior leadership (Bloom, Han and Liang, 2024). Those are not window dressing. They are the scaffolding for asynchronous work, where progress depends on clarity of outputs and predictability of handoffs, not on constant real-time availability.
Look at how one large all-remote firm codifies the mechanism. GitLab runs a public, “handbook-first” system in which policies and processes are documented in a living manual that would exceed 2,000 printed pages. The company states plainly that asynchronous work is impossible without strong documentation because unanswered follow-up questions block colleagues for hours or days. The remedy is a searchable record that lets anyone self-serve the information needed to move forward (GitLab handbook, 2026). This is not an outcomes study. It is a practitioner case that names the dependency. Async scales only when the written record does the coordination work that meetings used to do.
Matt Mullenweg, who leads Automattic, offers a useful language for maturity. His “five levels of autonomy” place asynchronous, output-measured work at level four, with level five as an aspirational claim that distributed organizations can outperform any co-located peer. The taxonomy clarifies the managerial shift at stake, although the performance claim is opinion, not empirical evidence (Mullenweg, 2020). The point stands: in async systems, what matters is what ships.
Design the hybrid pilot that builds async muscle
Leaders who want distributed capability should stop asking whether remote is good in the abstract. Instead, run a hybrid pilot that forces the infrastructure of async into existence, then judge it by retention and performance, not anecdotes. The evidence offers a practical checklist.
-
Define outputs and measurement up front. Trip.com’s result rests on rigorous performance management. Make goals, service levels, and review criteria explicit, and test for equivalence in ratings across arrangements over multiple cycles (Bloom, Han and Liang, 2024).
-
Coordinate team schedules rather than allowing every manager to improvise. The randomized hybrid cadence in the Trip.com trial was scheduled, which reduced the friction of handoffs and made collaboration predictable (Bloom, Han and Liang, 2024).
-
Put leadership weight behind the model. Senior endorsement reduces the risk that hybrid becomes a second-class track and helps managers align on how to evaluate work fairly (Bloom, Han and Liang, 2024).
-
Build the documentation backbone. Treat the GitLab playbook as a design pattern. Move recurring decisions, processes, and FAQs into a maintained, searchable handbook. Require that decisions be recorded in writing so teams in different time zones or focus blocks can proceed without a meeting (GitLab handbook, 2026).
-
Guard against the visibility tax. The Ctrip experiment documented a promotion penalty for full-time remote workers, conditional on performance. Even in hybrid, make promotion packets rely on documented outputs and peer-reviewed evidence rather than presence signals, and audit promotion rates by arrangement to catch emerging bias early (Bloom, Liang, Roberts and Ying, 2015).
-
Quantify the retention economics. In Trip.com’s case, the hybrid-induced drop in quits was large enough that the company estimated millions of dollars saved in recruiting and training, an internal figure reported by the study’s authors (Bloom, Han and Liang, 2024). You do not need to guess. Track your own attrition, replacement costs, and time-to-productivity.
This is not a culture war. It is an engineering problem. The hybrid model that passed the Trip.com test wrapped clear metrics, coordinated schedules, and leadership support around two days at home. Layer on a handbook-first discipline and you have the core of async: work that moves through written interfaces, evaluated by output.
The practical read
If your organization cannot make a structured two-day hybrid model work on the numbers, you are not ready for real distributed autonomy. The remedy is not more meetings. It is better definitions of good work, fairer promotion mechanics, and documentation that lets people move without waiting on a ping. The experiments show hybrid can retain talent without sacrificing performance. The managerial task is to use that breathing room to build the asynchronous muscle memory that makes a distributed system reliable.
Sources
- Bloom, Han and Liang (2024). Hybrid working from home improves retention without damaging performance, Nature. A randomized controlled trial at Trip.com where two days at home cut quits from 7.2% to 4.8% with no loss in performance or promotions, establishing causal support for well-run hybrid. https://www.nature.com/articles/s41586-024-07500-2
- Bloom, Han and Liang (2024). One Company A/B Tested Hybrid Work. Here's What They Found, Harvard Business Review. Managerial summary of the Trip.com experiment that details three conditions for success and reports the firm's internal estimate of millions saved from lower attrition. https://hbr.org/2024/10/one-company-a-b-tested-hybrid-work-heres-what-they-found
- Bloom, Liang, Roberts and Ying (2015). Does Working from Home Work? Evidence from a Chinese Experiment, Quarterly Journal of Economics. A randomized trial at Ctrip where full-time home work raised performance 13% and cut attrition but reduced promotion odds, with larger gains when employees later chose their setup. https://academic.oup.com/qje/article-abstract/130/1/165/2337855
- GitLab handbook (2026). GitLab — handbook-first & asynchronous work, GitLab public handbook. Practitioner case describing a documentation-first system exceeding 2,000 printed pages and asserting that async is unworkable without strong documentation. https://handbook.gitlab.com/handbook/company/culture/all-remote/asynchronous/
- Mullenweg (2020). Five Levels of Autonomy, ma.tt. Practitioner taxonomy that defines async, output-measured work as level four, used here as a framing device rather than empirical proof. https://ma.tt/2020/04/five-levels-of-autonomy/