Does working from home improve performance?

Working from home does not carry one demonstrated universal productivity effect. The strongest field experiments produced opposite estimates in different jobs, populations, homes and schedules; they identify effects within those settings, not which difference caused the reversal.

In one Chinese call centre, 134 order-takers within a 249-person randomized sample handled 13% more calls per week when assigned mainly to home. In a Chennai data-entry experiment, people randomly assigned home produced 18% less than those assigned an office. A later hybrid trial at Trip.com preserved measured performance while reducing attrition. And in June 2026, researchers reported that just one coordinated office day per month improved calls per hour in an otherwise fully remote team.

The useful question is therefore not whether remote work wins. It is which parts of a job need quiet, autonomy, feedback, learning or shared presence—and whether the chosen schedule supplies them.

Productivity is not one number

Search results often flatten productivity into a single percentage. Workplace studies measure at least six different things:

• total output, such as calls handled;
• output per active hour or minute;
• time spent working;
• quality or error rates;
• manager-assigned performance grades;
• organizational outcomes such as retention.

Those measures can move in different directions. Someone can complete more work because they stay active for longer without becoming faster. A team can retain more experienced employees without changing each person’s immediate output. A manager can rate performance as stable while one narrow activity, such as submitted lines of code, changes.

Remote work also describes different interventions. Fully remote work is not a two-day hybrid schedule. Two home days are not one shared office day per month. An emergency pandemic move is not the same as a planned policy with stable equipment, managers and routines.

Any claim that merges those designs loses the information a decision-maker actually needs.

The randomized results run from minus 18% to plus 13%

The best-known positive experiment began at Ctrip, a Chinese travel company. Of 994 call-centre employees, 503 volunteered to work from home and 249 met the eligibility rules, including having broadband and a private home workspace. Birth-date parity assigned those eligible volunteers to work mainly from home or remain in the office for nine months.

Across all 249 eligible volunteers, home assignment raised a standardized performance index by 0.232 standard deviations. The familiar 13% figure comes from the 134 order-takers: the estimated change in log weekly calls was 0.120 (SE 0.025). Its decomposition matters. About 9% came from being logged in for more minutes during scheduled shifts, with fewer breaks and sick days; about 4% came from more calls per active minute. The two recorded quality measures for this subgroup did not decline.

The broader trade-off was not uniformly positive. Attrition fell by about half and job satisfaction rose. Promotion was not significantly lower overall; after controlling for measured performance, the estimated penalty was about half and statistically weak. When the company later let people choose their location again, many switched. The resulting larger performance gap combined the original treatment with worker sorting; it was not a second randomized effect.

A 2023 Chennai experiment produced the opposite headline. Of 892 applicants, roughly half were invited to training, 280 began and 45 dropped out before work started. The resulting 235-person sample entered an eight-week data-entry operation and was randomly assigned home or office with the same tasks, laptops and compensation structure. The principal estimates use later waves after a ₹2,000 first-week completion bonus removed the earlier differential attrition.

Home-assigned workers produced 18% less in log net typing speed (SE 0.050). About two-thirds of the gap appeared on the first day; the rest came from faster learning in the office. The result also appeared in typing speed and accuracy.

Preference did not solve the matching problem. Applicants who preferred home were about 12% faster at baseline. The estimated home penalty was roughly 27% for them and 13% for those preferring the office, but this interaction was weak: it reached the 10% significance level only after controlling for baseline performance and was not conventionally significant without that control. Family care, children and other measured constraints explained only part of the selection pattern.

The two experiments do not cancel each other out. Ctrip selected experienced volunteers with a private room for measurable call work. The Chennai project recruited people into a short data-entry operation, and many worked amid the demands of an urban household. Those observed differences limit generalization; the studies did not isolate which difference caused the reversal.

Hybrid work protected performance and improved retention

In 2021, Trip.com randomized 1,612 graduate employees in engineering, marketing, accounting and finance. Employees with odd-numbered birthdays could work from home on Wednesday and Friday and attended the office on the other three days. The control group attended all five days.

Over six months, attrition was 4.8% in the hybrid group and 7.2% in the office group—a 2.40 percentage-point difference (SE 1.18; 95% CI 0.075–4.72) and a one-third relative reduction. Job satisfaction also improved.

The performance result is more restrained than many summaries suggest. Formal tests placed the difference in performance grades inside a prespecified half-grade equivalence bound across four review periods, although the review samples declined from 1,507 to 1,254 as employees left. There was no evidence of an overall promotion effect, but promotion did not meet the equivalence bound in every period. Among 653 coders over 95,494 employee-days, raw lines per day met the prespecified ±29 lines equivalence bound—10% of the control mean (p=0.003). Separate conventional zero-null sensitivity tests differed: log2(lines), which excluded zero-line days and used 27,605 days, was null (p=0.750), while log2(1 + lines) favored treatment (p=0.0103).

That is not a failed intervention. If a schedule preserves measured performance while retaining more people and improving satisfaction, it can still create substantial organizational value. But the causal claim is no performance damage in this design, not two home days made everyone more productive.

Managers changed their minds. Before the trial, they expected hybrid work to reduce productivity by 2.6% on average. Afterward, they expected a 1.0% increase. That shift records belief after experience; it is not an objective one-percent performance gain.

One office day a month changed a fully remote team

The freshest controlled evidence tests an unusually light intervention. In a June 2026 NBER working paper, 248 eligible volunteers from a fully remote customer-service team in Turkey were split evenly by company-ID parity. One group stayed fully remote. The other worked together in the office on one fixed day per month for nine months, then returned to fully remote work.

Nothing else about production was meant to change. Pay, shifts, laptops, call routing and performance rules stayed the same. The company covered transport and meals. About 95% of assigned office days were attended.

The effect built slowly rather than appearing as an immediate novelty boost. During the five months after the final required visit, the treatment group averaged 12.4 calls per hour and controls 11.5—a raw gap of 7.8%. A model without employee fixed effects estimated 0.87 additional calls per hour (SE 0.23); comparing change within employees estimated 0.63 (SE 0.20), about 5.8% of the pre-treatment mean. The difference between those estimates was consistent with better retention changing who remained in each group.

Quality did not show a significant average decline in customer ratings or manager audits. Cumulative attrition reached 13.7% in treatment and 21.0% in control.

The mechanism evidence is consistent with increased contact. Among the 223 endline-survey respondents, treated employees reported 36 more minutes of communication with their three most frequent colleagues in the previous week. Within the treatment group, randomly assigned desk neighbours were 11 percentage points more likely to communicate the following week. The study cannot tell how much of the performance effect came from peer learning, manager feedback or team attachment, and it did not separately test visibility or coordinated versus uncoordinated attendance.

The result does not validate any return-to-office mandate. Participants had volunteered and lived within commuting distance; they were less likely than non-volunteers to be married or have children. The job consisted of measurable, standardized customer-service work in one firm. And the paper is a registered working paper, not yet peer-reviewed.

It tests a precise idea: an otherwise remote team may gain from sparse, coordinated contact. It does not show that one day every week—or an arbitrary day when teammates are absent—would do the same.

The newest meta-analysis finds a small average, not a universal rule

An open-access meta-analysis published online in July 2026 assembled 3,574 estimates from 82 studies. After statistical adjustments for publication bias and reported design differences, its average relationship between working from home and productivity was small and positive.

These are modeled conditional fitted values, not simple pooled averages. The authors set the standard error to zero, assumed a continuous home-work measure, used randomized or quasi-experimental identification with the selected controls, and set the pre-pandemic year to 2019. Under those assumptions, the adjusted pre-pandemic overall estimate was a partial correlation of 0.024, with a 90% interval from −0.012 to 0.061. The pandemic and post-pandemic estimate was 0.019, with a 90% interval from 0.008 to 0.031. Pre-pandemic hybrid evidence was positive in the reported conditional analysis; the fully remote interval included zero. In the later period, both remote and hybrid conditional intervals included zero when separated.

That synthesis expands the map, but it cannot make unlike measures identical. Its source studies included task output, earnings, hours and self-rated productivity. A wage is not a completed task. More hours are not necessarily more output per hour. And statistical adjustment cannot turn observational associations into one multi-firm randomized trial.

The authors, Colin F. Mang of McMaster University and Amar Anwar of Cape Breton University, reach the practical evidence gap: the field still lacks a randomized study spanning multiple firms and occupations.

The accessible publisher page displays no separate funding or conflict declaration. That absence is not evidence that no interest exists; it records only what the page discloses.

US industry data offer another warning against one-number answers. A Bureau of Labor Statistics analysis used the American Community Survey’s usual or majority commute mode—not a direct measure of every remote day—and found that industries with larger increases in working from home also had stronger total-factor-productivity growth from 2019 to 2022 after accounting for earlier trends. A San Francisco Fed analysis across 43 industries used occupation-level teleworkability rather than observed remote work rates and found little relationship with pandemic- era labor-productivity growth. Different constructs, periods and industry mixes can move an aggregate result. Neither analysis randomly assigned remote work.

Primary sources

• Bloom et al., Ctrip randomized experiment
• Atkin, Schoar and Shinde, Chennai randomized experiment
• Bloom, Han and Liang, Trip.com hybrid experiment
• Aksoy et al., one office day a month randomized trial
• Mang and Anwar, July 2026 meta-analysis
• US Bureau of Labor Statistics industry analysis
• Federal Reserve Bank of San Francisco industry analysis

Independent editorial summary. The authors and organizations are not affiliated with LifeScore.