
Companies say it’s about collaboration. That’s garbage. The reality is a room full of people in headphones, doing alone what they could’ve done at home.

Alexa von Tobel, who runs an early-stage fund and has backed a long list of founders, posted a small, honest question. She hadn’t been pitched a single remote-only company in six months, and she wanted to know if anyone else was seeing it.
The replies were a mix. A few pushed back. One said she had landed a remote role just fine and that nobody had to pitch her anything. Another figured the remote crowd was mostly off building their own companies now.
The reply that stuck with me came from a CEO, who said being fully remote would be harder, and less fun. It resonated because he is half right. Something real does get lost when everyone is scattered. He is just wrong about what fixes it.
What I haven’t seen is anyone ask what should be the obvious question. If the office is winning, what exactly is it “winning” people back to do?
This matters more than it used to, because the thing we always said the office was for has left the building (pun intended). We told ourselves we showed up every day to get the work done together. The making, the producing, the actual output. You went to the office because that was where the work physically got made, and it got made faster with everyone in one place.
Thomas Allen measured the reason decades ago at MIT. The closer two people sit, the more they actually talk, and past about fifty feet the conversation falls off a cliff.
Proximity was the whole point. The commute bought you that.
The work doesn’t live there anymore. It runs on its own now, in the cloud, whether you are at your desk or asleep. Gartner expects forty percent of business software to have agents built into it by the end of 2026, up from under five percent a year earlier.
Offices may still exist, but perhaps we lost the reason for them.
My position is simple, even if the fight over it refuses to be. The office isn’t dead, but it shouldn’t be mandatory. It’s worth protecting for the handful of things people can only do well together, the ones AI can’t touch. Instead it’s being forced on everyone for the work that no longer needs a building at all.

The work left the building
The founders von Tobel is funding are increasingly building agents that do the production themselves. Y Combinator’s recent cohorts are full of companies whose entire pitch is autonomous software that writes, tests, and ships code with a human barely touching it, and the firm’s own partners say AI has collapsed the cost of producing software by a factor of ten to a hundred.
This is not a forecast. It already shows up in who has a job. Erik Brynjolfsson and his colleagues at Stanford’s Digital Economy Lab tracked real payroll data and found that employment for workers aged twenty-two to twenty-five in the most AI-exposed jobs sits about nineteen percent below where it should be, while their older colleagues are fine.
The gap is coming from companies not hiring junior team members, because AI is doing the checkable, process-heavy tasks that used to justify an entry-level seat. You know, the production work, the stuff we all started our careers doing.
It is not only software. Anthropic, looking at how its own models actually get used, found that coding is the single biggest category, and that while most use still has AI working alongside a person, a large and growing share, more than four in ten interactions, is now the model doing the task outright rather than assisting.
The line between the two is blurry, but it is clearly moving in one direction.
Designers feel it and say so. Francisco Barrera Aros points out that AI has swallowed the how of building an interface and is coming for the what, which is why job boards have started sprouting roles like AI Product Designer that nobody could have described two years ago.
AI took over the task of doing the work. The deciding, however, is still yours, and the deciding is where the money is.
Let’s look at what that leaves for the purpose of an office. For a century it was where the production happened, because production needed people in one spot with the files, the tools, and each other. Take that out, hand it to the agents, and the physical building becomes a very expensive room you commute to in order to do work that could happen anywhere.
Which should, by any reasonable prediction, empty the place out.
We all remember 2020, right? The pandemic sent nearly everyone home at once. Jose Maria Barrero, Nicholas Bloom, and Steven Davis called it a mass social experiment in working from home, with Americans doing roughly half of all paid hours from home through most of 2020, up from about five percent before.
The floors we were told could never go dark went dark, and the striking part, in their data, is how many people were surprised they could do the job from home at all. The mandates coming down treat remote work as an unproven indulgence, when we have years of evidence that most of the job travels. The question should be which parts don’t work and why not?

So everyone scattered, right?
Obviously, work is more portable than it has ever been, but the freer people are to work anywhere, the more they actually want to be together.
Y Combinator, which has funded a good chunk of the companies building these agents, spent years running its batches remotely, and in those years its founders were scattered all over the map. Then it brought the program back in person, in San Francisco, and the founders came right back with it. The trend is easy to see in YC’s own public directory, which one analysis tracked batch by batch.
When the program went remote, barely a quarter of companies listed San Francisco as home. Once it came back in person, that climbed past three quarters, and a chunk of those founders kept choosing the city even after the three months ended and nothing required them to stay. The same batches are stuffed with startups whose product is autonomous software.
The people automating the production hardest are also the people most determined to sit in the same zip code while they do it.
This is one contender to define whole AI era, at least for the workforce. Gensler’s global survey found that the workers most embedded in AI are also the most collaborative and the most drawn to the office, which is the reverse of the lonely-robot future everyone predicted.
As AI takes more of the solo production, the human collaboration it can’t do becomes a scarce, valuable thing, and the people who live in these tools feel the pull toward the room where that still happens.
If the office were really about the work, this is backwards. These founders have built, and are selling, the exact tools that make location irrelevant. They could run these companies from anywhere, and they tell you so with a straight face in the pitch deck. Then they sign a lease three blocks from each other.
The clustering is telling you something that a pitch deck won’t. Whatever they are gathering for, it is not the production, because they already shipped the production to the cloud. It is the other thing.
There is real research on what that other thing is. Daniel Carmody, Martina Mazzarello, Paolo Santi and their colleagues at MIT’s Senseable City Lab studied what happened to people’s networks when everyone went remote, and found that distance measurably starved the weak ties, the loose connections that only form when people share a space. Weak ties are the people you don’t work with closely, the person two teams over, the one you only ever talk to because you both happened to be refilling coffee.
Mark Granovetter named their importance decades ago, in the work that gave us the phrase. The loose connections, not the close ones, are where new information and unlikely ideas actually travel. They are also the first thing to die when everyone is a tile on a screen, because nobody schedules a video call with the person two teams over for no reason.
That is what the founders are re-clustering to get back. Not the work. The accidental stuff around the work. The collisions you cannot put on a calendar, which turn out to be the one thing the agents can’t do for you and the screen can’t give you either.

The office we actually built
Stroll into a lot of these offices and look at what is actually happening.
Rows of people who commuted in to sit at a desk and type into a chatbot. Four people at four screens running four separate conversations with four separate AIs, in the same room, not speaking to each other. The agents are collaborating with nobody. The humans are collaborating with nobody.
Everyone is doing, in a shared space, the exact solo work they could have done at home, now with an uncomfortable chair and a commute attached.
I know this office because I have been there. I’ve worked as a fractional CPO at a variety of startups over the past few years and some of them required everyone in office for collaboration and culture purposes, and what was actually happening was a floor of people at open desks and shared tables, everyone heads-down in their own tasks, earbuds in, which is not collaborative in any way whatsoever.
The only real interaction occurred when somebody took a call out in the open and derailed the eight people near them, who were now straining to hear their own thoughts over somebody else’s pitch meeting. The in-person mandate had not created the intended collaboration. All it did was take solo work that people were doing fine at home and made it harder.
This is the part of “Office Space” that could have been a documentary. Peter Gibbons didn’t hate his job because the work was hard. He hated the commute, the eight bosses who each wanted the same TPS report, and the open floor where a man could lean over the partition at any moment to ask why he hadn’t used the new cover sheet.
The building added nothing to the work and did create a great deal of misery. The movie played it as absurd, but, frankly, a lot of return-to-office plans have adopted that as an operating model.
This is not an anecdote the research argues with. Ethan Bernstein and Stephen Turban put sensors on employees before and after a company tore out its walls for an open floor plan, the layout that is supposed to spark all that spontaneous mingling. Face-to-face interaction dropped by about seventy percent.
People responded to being shoved together by putting their headphones on and retreating into email and chat. This is the equivalent of throwing a party, forcing everyone to come, and watching them text each other from opposite corners of the room.
So if being in office isn’t producing collaboration, what is the mandate for? Yuye Ding and Mark Ma looked at the S&P 500 companies that ordered everyone back and found no improvement in firm value and a measurable drop in how people felt about the job. Their read was blunt. The mandates lined up with managers using the office to reassert control and to have someone to blame, not to get better work.
Forget the productivity numbers. The real cost is the people you lose, and you lose the best ones first.
Nicholas Bloom and his colleagues ran a real randomized trial, published in Nature, with more than sixteen hundred employees split between five days in and a hybrid schedule. The hybrid group quit about a third less often, with no hit to performance or promotions.
Forcing everyone back five days a week does not buy more output. It buys more resignations, and the first ones gone are the people with somewhere else to go.
That a pretty big cost for the privilege of watching your agents do the work in a room full of other people watching theirs.
The thing is, the people deepest into AI are the ones who most want to be together, and not to sit in silence. Janet Pogue McLaurin and her team at Gensler surveyed more than sixteen thousand workers and found that the heaviest AI users, about a third of everyone, are also the most collaborative, the most connected to their teams, and the most likely to come into the office. They are coming in for each other.
These five-day in-office mandates are coming almost entirely from the biggest companies, and, to be fair, that is the one place the evidence says in-person helps. Arpit Gupta, Abhinav Gupta, and Elena Simintzi tracked remote work across thousands of firms and found it cut productivity at established companies by about nine percent, while raising it at startups by about twelve.
Big, old organizations genuinely struggle to coordinate when scattered. Small teams thrive, because remote lets them hire the best people anywhere instead of the nearest people around. So the companies swinging the hardest mandates are the ones with a real coordination problem, and they are solving it by forcing the fix on everyone, including the small teams it only slows down.

What the room is actually for
Full disclosure. I have run my company fully remote since 2019, before the pandemic made it “cool.” For most of our work, I still think remote is the right answer. It’s not my favorite thing to admit that I have found that the office turns out to be very good at a few specific things, and that those few things are suddenly the only ones that really matter.
Once the agents take the production work, the part of the human’s job that remains relevant is to decide what’s worth making, to catch the thing that’s wrong and to teach the person above and below you on the ladder how to see it too.
That happens to be the work the research ties to being in a physical place together.
Melanie Brucks and Jonathan Levav ran a study putting pairs together either in person or over video and asking them to generate ideas. The in-person pairs came up with meaningfully more, and more creative, ideas. The video pairs were just as good at the next step, picking which idea to run with.
Basically, generating creative ideas is helped by sharing a space, and choosing the ones to implement isn’t. This lines up exactly with the split the agents just created. The machine is a fine partner for narrowing and deciding against known options. It is a worse one for the wide-open, slightly stupid, what-if-we-tried part, and that part goes better with a human across the table.
There is one finding that gives me pause, because it’s the real cost of how I’ve chosen to work, and it is definitely one AI has made worse.
Someone has to learn to do the actual jobs. People learn skills by sitting near other people who already know how, absorbing the small corrections, asking the dumb question out loud instead of filing a formal ticket for it. Natalia Emanuel, Emma Harrington and Amanda Pallais measured this and found that junior engineers got dramatically more feedback when they sat near senior ones, and lost about a fifth of it when they went remote.
Proximity is how the craft gets handed down. Distance quietly starves it.
I feel this one personally, because it is the hardest part of running a company the way I do. One of my own designers, still growing in her career, put it this way. “Constructive feedback is how I know I’m improving, and that’s hard to get when you can’t just walk into your boss’s office and ask them to look at something with you. I’m lucky I’m on a small enough team that I feel comfortable reaching out. If I were part of anything larger, I think it would be much harder. Even on bigger client teams, the feedback I get from people outside my everyday group is thin, when I get it at all.”
She is right. The ambient coaching that used to just happen in a room, the over-the-shoulder correction, the pulled-aside-after-a-meeting, does not happen on its own when everyone is remote. We made it happen on purpose.
We hold a standing thirty-minute check-in every week, time that exists for nothing but the feedback and questions they’d otherwise have had to catch me in a hallway to ask. It is a deliberate, scheduled replacement for a thing that used to happen a bit more naturally. Remote turns the mentoring into something you have to build on purpose, instead of something you can assume.
This isn’t just happening at my company, of course. The KPMG American Worker Survey found AI adoption and office mandates rising together, and concluded that managers are now the linchpin for turning in-person time into real mentorship for the people earliest in their careers, the exact thing a crowded floor of people in headphones does not deliver on its own.
Now put that next to what the agents are doing. Remember the Stanford finding, that the entry-level jobs are the ones vanishing, because AI is doing the checkable, process-heavy tasks that used to be a junior’s whole first year. Those tasks were how you learned, by doing the boring version a hundred times until you could feel when it was wrong.
Anyone who watched Richie spend a week at that fine-dining restaurant on The Bear understands this. They hand him a mountain of forks to polish and nothing else, and he spends days furious about it, a grown man doing the most menial job in the building. Then somewhere in the repetition it clicks, and he comes out the other side understanding the whole craft in a way no shortcut could have taught him.
That was the art of apprenticeship. In our world, the agents are polishing the forks now, and the mentor who would have stood over you while you learned lives on a screen, if they’re there at all.
The on-ramp to every one of these careers is being pulled up from both ends at once, by the agents from below and by the distance from above, and almost nobody is talking about the two as the same problem. They are the same problem.
This is where the office earns its keep now, if it earns it anywhere. It’s not necessarily a place to produce deliverables, the agents clearly have that, but as the one place a person can still learn the judgment the agents can’t, from another person.
You cannot apprentice over a screen share nearly as well as you can over a shoulder.
Designers have already championed this work-style. Patrick Neeman has written that the next decade of design work is about owning the systems and the trade-offs underneath the screens, the part you can’t hand to a model. Vlad Derdeicea pointed out that most of a design lead’s day was never mired in pixels in the first place, but rather focusing on aligning, deciding, and convincing. Pushing pixels was a small piece of the day that a machine can now take, which gives leaders more time to do the teaching, which was always the actual job.
None of this adds up to an argument for the five-day in-office mandate. It adds up to something narrower. The office matters now for exactly the things it was never really designed to optimize, the idea you’d never have had alone, the greener team member who evolves by watching you work, the trust that only seems to set when people share a space. It matters more than ever, and for less than ever.

So what is the commute for
Every one of these mandates comes with the same justification. Andy Jassy wrote in his Amazon memo that collaborating, brainstorming, and inventing are simpler and more effective in person, and that teaching and learning from one another are more seamless. Jamie Dimon called the benefits of being together substantial and irreplaceable.
The reality is, these leaders are not wrong that those things matter, but it’s worth considering whether they are wrong that a mandate delivers them.
Chris Wardman describes this contradiction pretty bluntly, saying companies deploy AI that strips out the everyday interactions where knowledge actually transfers, then mandate attendance to protect the very thing they are automating away.
The Flex Index, which tracks the policies of roughly nine thousand companies, now puts about a third of American firms on full-time in-office, and the biggest companies are leading the charge. What that mostly creates, though, is a building full of people doing exactly the work that no longer needs a building. In the meantime, the collaboration they said they wanted is just sitting right there on the mandate, impossible to order up.
You cannot schedule serendipity for 9 a.m. Nobody mentors the new kid on command.
A good idea has never once shown up because a policy required it. Drag everyone in and chain them to the desks, and all you have manufactured is a more expensive version of the thing they were already doing at home, plus traffic.
A move that actually fits the current environment is narrower and a little strange. Treat the office the way you’d treat any specialized tool. You don’t run your whole life out of the one drawer with the good scissors, you open that drawer when you need the thing that’s in there.
Nicholas Bloom’s data keeps pointing at the same sweet spot, a couple of deliberate days together, the rest wherever the work actually gets done, which buys the retention and the output without the resentment.
Yuan Yuan and his colleagues found the same thing when looking at badge data from forty-three thousand people across a global company’s offices. Left alone, people coordinate their own in-person days around the work and the teammates that actually need to be there.
Any mandate that just counts heads on a Tuesday misses the point of being together. Come in for the ideation, the teaching, the arguing, the trust that only seems to set face to face. Then go home and let the agents run the production from a server farm that never needed to see your face anyway.
While the office technically never died, it shouldn’t be roaring back to what it was. AI merely changed what it’s for. For a hundred years it was where the work happened and now the work happens in the cloud, and the building is left holding the handful of human things we were always a little embarrassed to admit were the actual point.
Ryan Anderson, writing in Forbes about Atlassian’s internal findings, argues that as AI absorbs the routine work, value creation migrates toward the things that require human insight and collaboration, which makes the physical workplace more important, not less. Sure, AI doesn’t need an office, but the people using it do, mainly for the handful of things this quickly moving technology cannot do for them.
This brings us back to Alexa’s question. People are coming back to the office, but that was never really the question. The question is what they think they’re coming back to do. If the answer is to sit at a desk and quietly prompt a machine that works fine without them, next to other people doing the same, then they didn’t come back for the right reasons.
If people are merely driving in to be alone together, then, that, you really can do from home.
References and further reading
The work left the building
- Erik Brynjolfsson, Bharat Chandar and Ruyu Chen, “Canaries in the Coal Mine? Six Facts About the Recent Employment Effects of Artificial Intelligence” (Stanford Digital Economy Lab, 2026). On the widening employment gap for young workers in AI-exposed jobs, as AI takes the entry-level tasks that once justified a first job.
- Anthropic, “The Anthropic Economic Index.” Anthropic’s own analysis of how its models are used, with coding as the largest category and a large, growing share of use that automates tasks rather than assists.
- Francisco Barrera Aros, “The crisis of the what” (UX Collective, 2026). A designer on AI taking over the how of building interfaces and coming for the what, leaving judgment as the differentiator.
- Thomas Allen, “Managing the Flow of Technology” (MIT Press, 1977). The origin of the Allen Curve, the finding that communication between coworkers drops off sharply with physical distance.
- Gartner, “Gartner Predicts 40% of Enterprise Apps Will Feature Task-Specific AI Agents by 2026” (2025). The forecast that agents built into software will jump from under five percent of enterprise apps to forty percent in a year.
- Jose Maria Barrero, Nicholas Bloom and Steven J. Davis, “Why Working From Home Will Stick” (NBER, 2021). On the pandemic as a mass forced experiment in remote work, and how surprised workers were that they could do the job from home.
So everyone scattered, right?
- Y Combinator, Frequently Asked Questions. YC’s own statement that the batch runs in person in San Francisco, having moved back after a brief remote stretch, because it works better that way.
- Daniel Carmody, Martina Mazzarello, Paolo Santi and colleagues, “The effect of co-location on human communication networks” (MIT Senseable City Lab). On how remote work impedes the weak ties that form through physical proximity.
- Mark Granovetter, “The Strength of Weak Ties” (American Journal of Sociology, 1973). The foundational work on why loose connections, not close ones, carry new information and ideas.
- Janet Pogue McLaurin, Anita Grabowska and Katherine Engel, “Will AI Make the Workplace More Human?” (Gensler Global Workplace Survey 2026). A survey of more than sixteen thousand workers finding that the heaviest AI users are also the most collaborative, the most connected to their teams, and the most drawn to the office.
The office we actually built
- Ethan Bernstein and Stephen Turban, “The impact of the ‘open’ workspace on human collaboration” (Philosophical Transactions of the Royal Society B, 2018). Sensor data showing face-to-face interaction fell about seventy percent after a move to open-plan.
- Yuye Ding and Mark Ma, “Return-to-Office Mandates” (University of Pittsburgh). On RTO mandates at S&P 500 firms producing no gain in firm value, lower satisfaction, and higher turnover among skilled staff.
- Nicholas Bloom, Ruobing Han and James Liang, “Hybrid working from home improves retention without damaging performance” (Nature, 2024). A randomized trial of 1,612 employees where hybrid work cut quit rates by a third with no hit to performance.
- Arpit Gupta, Abhinav Gupta and Elena Simintzi, “Who Benefits from Remote Work?” (2025). On remote work raising productivity at startups while lowering it at established firms, based on data from thousands of companies.
What the room is actually for
- Melanie Brucks and Jonathan Levav, “Virtual communication curbs creative idea generation” (Nature, 2022). In-person pairs generate more creative ideas than video pairs, who are just as good at selecting among them.
- Natalia Emanuel, Emma Harrington and Amanda Pallais, “The Power of Proximity to Coworkers” (Quarterly Journal of Economics, 2026). Junior workers get far more feedback and mentorship when they sit near senior ones, and lose about a fifth of it when remote.
- Patrick Neeman, “From faster pencil to AI Experience Architect” (UX Collective, 2026). On design work moving from making the screens to owning the systems and trade-offs a model can’t handle.
- Vlad Derdeicea, “The 80% job: how design leads are using AI, and it’s not about mockups” (UX Collective, 2026). On most of a design lead’s work always being the human communication and judgment, with AI taking the production slice.
So what is the commute for
- Andy Jassy, message to Amazon employees on returning to the office (2024). The Amazon CEO’s own stated reasons for the five-day mandate, collaboration, invention, and learning from one another in person.
- Jamie Dimon, on JPMorgan’s return-to-office requirement. The CEO’s argument that the benefits of being together in person are substantial and irreplaceable.
- The Flex Index (Scoop Technologies). The tracker of work-location policies across roughly nine thousand U.S. companies, showing about a third now require full-time in-office attendance.
- Marie-Laure Charpignon, Yuan Yuan, and colleagues, “Navigating the new normal: Examining coattendance in a hybrid work environment” (PNAS, 2023). Badge data from forty-three thousand employees showing people coordinate their own in-person days around the work that needs it.
- Ryan Anderson, “AI Is Making the Physical Workplace More Important, Not Less” (Forbes, 2026). On how, as AI absorbs routine work, value shifts to human collaboration and judgment, making the office more valuable for those things, drawing on Atlassian’s internal research.
- Chris Wardman, “Companies Say Return-to-Office Mandates Are About Collaboration. Their AI Strategies Say Otherwise” (Built In, 2026). On the contradiction of mandating in-person work for collaboration while deploying AI that erodes the very interactions companies claim to want.
- The KPMG American Worker Survey (2025). On rising AI adoption alongside rising return-to-office, and why managers are the linchpin for turning in-person time into mentorship for early-career workers.
If AI works from anywhere, why are we commuting to watch it? was originally published in UX Collective on Medium, where people are continuing the conversation by highlighting and responding to this story.
