How Many People Work from Home?

Nicholas Decker

Oct 16, 2026

15

3

2

Share

I was in a seminar, listening to a presentation on remote work. One of the claims was that 27% of work days in the United States today are done remotely. I thought that was too high. I was correct. But in doing so, I’ve had to realize – there really are a lot of people working from home, many more than I thought, and we don’t know much about what the effects are.

The figure comes from Barrero, Bloom, Buckman, and Davis (2025) (BBBD), who synthesize several data sources. Going over the details of the various surveys may be boring, but the details are critically important to the point I’m making.

This post is brought to you by Mechanize, Inc. They are hiring for a variety of positions, including software engineers. I encourage you to apply here.

The main one is a survey of working arrangements and attitudes (SWAA henceforth) which they have conducted every month since May 2020. It is an online survey, sampling up to 10,000 U.S. residents who are between 20 and 64 years of age and (for purposes of analysis) earned more than $10,000 in a prior year. They also have another online survey, the U.S. Census Bureau’s Household Pulse Survey (CHPS). They were able to sample people who had not opted into the survey platform, but had low response rates. Meanwhile, the Current Population Survey (CPS) and the American Time Use Survey (ATUS) are a mix of in-person visits and surveys over the phone.

These surveys give very different results. The online surveys report higher shares of people working from home: 28.7% from the SWAA, and 29.1% from the CHPS. (Notably CHPS is measuring anyone in the household, not just workers). The ATUS gives 6+ hour days (excluding any incidental work from home, such as, for instance, a teacher grading some essays) at 19.8%, while the CPS gives 14.3% of hours worked. These are estimates not made compatible by random error.

So one possible explanation is that the CPS sample is biased toward people regarding their work as not remote, because it is preceded with a few sentences about the pandemic. This could cause some people to think that their work isn’t really remote, and thus not report that. However, the American Time Use Survey covers the same sample of people, and has them report everything which they have done in the past 24 hours, as accurately as possible. They report 19% of work days being remote. Presumably, the fraction of workers misreporting cannot be greater than five percent.

A better explanation is that the SWAA sample is online, and even conditional on controls, they must differ systematically in how much they work online. This is the same sort of bias that occurs in regular surveys of politics, when the people who actually answer the surveys have different opinions than those who don’t, even when observably identical. I have several arguments in favor of this point besides it being obvious, though.

We can compare the implied number of jobs done remotely to the number of jobs which actually can be done remotely. Dingel and Neiman (2020) go through O*Net, and classify each job by the feasibility of working from home. They find that 37% of jobs can be done fully remotely. The SWAA numbers imply that at least 70% of the jobs which can be done remotely, are done remotely, which seems very high.

The Dingel and Neiman coding of whether jobs can be done remotely was done during the pandemic. It was measuring whether jobs could be done remotely in an emergency, not whether employers would want to keep it that way. I had Claude go through each occupation, compare it to O*Net as they did, and recode it, then had a separate agent go through and argue about which ones they had wrongly coded as not doable remotely.

This resulted in a considerably lower share of jobs being remotable than Dingel and Neiman, the biggest category being education. One can indeed conduct school over Zoom, but nobody does that now. Any reasonable measure of which jobs are done remotely now should not consider them as a serious possibility. Stripping out people related to education alone pushes the share of remotable jobs down to 30%; fully going through and considering what people actually can do remotely on a regular basis pushes us down to 25%. Clearly, we cannot have more remote workers than there are remote jobs to be done.

The problem becomes worse, though, when we consider what the responses were in the industries which we know cannot be done remotely. Look at the last page of BBBD, which reports the employment shares and rate of working from home.

The row of particular interest is food/accomodation. How could it possibly be that 20% of workers in food and accommodation are working remotely? How can we believe that 20% of baristas, cooks, and hotel maids are working remotely?

Everything points to the ATUS number being the correct one. Somewhere around 20% of work days are remote. That’s still really quite a lot, though.

Unfortunately we know very little about the effects. We know only a few things with confidence. Mas and Pallais (2017) find that workers are willing to accept a cut of 8% in wages to get a job which allows for remote work, at least among call-center applicants. There are a number of randomized controlled trials, but they all involve contexts where productivity and output is exactly measurable. Thus, software engineers and call-center employees. (Listing some out, a positive result in Bloom, Liang, Roberts, and Ying (2015), a null result from Bloom, Han, and Liang (2024), a negative result from Atkin, Schoar, and Shinde (2023)). Notably, that last one found that most of the losses from working from home showed up in the long run, as employees learned less on the job. Emanuel, Harrington, and Pallais (2026) tracks junior employees learning from senior employees using the exact seating arrangements of teams within office buildings, and finds that even small increases in friction lead to substantial reductions in learning. This reduction in connections shows up in innovation from Microsoft, where Yang et al (2021) shows that there were fewer connections across separated teams.

This is really quite weak, though. I think it is perfectly plausible that remote work allows for there to be an increase in surplus, although it will reduce wages and GDP. (This is because we are mismeasuring labor input). It seems plausible that, in the long run, it leads to less innovation, and less learning, but I would not expect it to be so large as to not be drowned out by other macro trends.