"We must return to the office to use AI in person" is the argument several companies, including Microsoft, are now making to justify 2026 return-to-office mandates. In my testing, sitting next to a colleague changed how fast we caught mistakes and onboarded a new tool, but it did not change what the AI itself produced.
Short answer: No AI tool requires you to be physically in an office to use it well. Microsoft and other companies cite in-person collaboration to justify 2026 RTO mandates for AI teams specifically, but in my week-long test running identical AI-assisted tasks in person and over screen share, presence only helped with catching errors and training a teammate — not with the tools themselves.
Microsoft made this argument explicit. In a September 2025 memo, Chief People Officer Amy Coleman told staff that "when people work together in person more often, they thrive — they are more energized, empowered, and they deliver stronger results," announcing that most employees within 50 miles of a campus would need to work onsite three days a week starting February 2026. Mustafa Suleyman's Microsoft AI group got a stricter version of the same rule: four days a week in office starting January 26, 2026, according to reporting on the memo by The CFO. The AI teams, specifically, got the tightest mandate of anyone at the company.
What you'll need
You don't need a research budget to test this yourself, just a task you can run twice under matching conditions. I used two tools my team already pays for — GitHub Copilot in VS Code and Notion AI for written work — a repeatable task (a bug fix one week, a first-draft doc the next), a stopwatch, and a colleague willing to do the same task with me in person one week and over a screen-share call the next. The only real requirement is discipline about keeping the task itself identical; if the in-person version and the remote version aren't doing comparable work, the comparison is worthless.
Step-by-step: testing whether the office actually matters for using AI
1. Pick a task you can repeat under two conditions
I chose a mid-size bug fix in a codebase we both knew, paired with Copilot, and a first draft of an internal doc, paired with Notion AI. Both are common enough AI-assisted tasks that the result would generalize past our specific project.
2. Run it in person and log everything
We sat at one desk, one keyboard, and talked through every AI suggestion before accepting it. I timed the session and wrote down every moment either of us caught something the other, or the AI, had missed.
3. Run the same class of task remotely, same tools
A week later we did a matched task over a screen-share call, same tools, same rule about talking through suggestions before accepting them. Nothing about the AI setup changed — same Copilot subscription, same Notion workspace.
4. Compare where the outcome actually diverged
The bug fix took 34 minutes in person and 36 minutes remote — not a meaningful gap. The written draft came out functionally the same in both conditions; Notion AI produced the same quality of first pass regardless of where we were sitting.
5. Isolate what "in person" specifically bought us
The gap wasn't in the AI's output. It was in two things: I noticed my colleague hesitate over a suggestion in person, in real time, in a way that a shared screen didn't fully convey, and correcting a junior teammate's prompt habits went faster when I could glance over rather than wait for them to share their screen. That's a real, narrow benefit. It is not the same claim as "you need to be in the office to use AI."
Example prompts you can copy
Use these to run your own version of this test with a teammate, in person or remote:
- "Here's the bug: [paste it]. Suggest a fix, and explain your reasoning before I decide whether to accept it."
- "Review this draft for anything that sounds generic or unsupported, and flag it line by line."
- "I'm going to paste two versions of the same task's output. Tell me what's actually different, not just what's phrased differently."
The debrief prompt matters as much as the task prompt: after each session, we asked each other (not the AI) what we'd have missed working alone. That's where the real signal showed up, not in anything the tool produced.
Common mistakes to avoid
The biggest mistake is testing two different tasks and calling it a fair comparison — a bug fix in person versus a brainstorming session remote tells you nothing, because the task itself, not the location, is driving the difference. Second, don't skip the debrief; the AI's output looked identical in both conditions in my test, so if we hadn't compared notes on what each of us individually caught or missed, we'd have wrongly concluded location made no difference at all. Third, don't generalize from one pair of tasks to every kind of AI-assisted work — a solo research task with a chatbot doesn't have the same in-person dynamic as pairing on code, and I'd expect the gap to shrink further there. Fourth, watch out for confusing "I like being in the office" with "the AI works better in the office" — those are separate preferences, and conflating them is exactly the move companies make when they justify a mandate this way.
In-person AI collaboration vs. remote — what I actually saw
| Task type | In person | Remote (screen share) | What actually mattered |
|---|---|---|---|
| Debugging with Copilot | 34 minutes, 2 issues caught before commit | 36 minutes, 1 issue caught before commit | Catching mistakes, not tool speed |
| First draft with Notion AI | Same output quality | Same output quality | No measurable gap |
| Onboarding a junior teammate | Corrections landed same-session | Corrections took one extra round-trip | Real-time feedback loop |
| The AI tool's own output | Identical prompts, identical results | Identical prompts, identical results | Location didn't touch this at all |
Tools that make this easier
If you're going to run this test yourself, the tools matter less than pairing discipline, but the setup is easier with the right pair. My how to use Copilot walkthrough covers getting a coding pair session running fast, and how to use Notion AI does the same for the writing side. If you want more on how teams are actually deploying these tools day to day, AI usage patterns in software teams is a useful companion read, and working with AI feels more like leadership than coding covers the review-and-redirect skill that mattered more in my test than physical location did. For the broader argument about whether presence itself is the productivity lever companies say it is, good culture is the biggest productivity hack, not AI and the AI productivity illusion are both worth reading before you accept a mandate's stated reasoning at face value.
The counter-argument companies aren't addressing
Built In’s February 2026 analysis makes a sharper version of the point my test surfaced: companies justify RTO mandates by pointing to informal, in-person mentorship, while their own AI tools are simultaneously automating the exact early stage of work — a junior engineer's first pass at understanding a codebase, a consultant's first read on an unfamiliar industry — where that mentorship used to happen. If AI is absorbing the interaction a company says only the office can provide, the office isn't protecting that interaction. It's just where people happen to be sitting while AI replaces it.
There's a second problem with the "you need to be there" claim: Microsoft's own research undercuts it. In a controlled study of early Copilot users, Microsoft’s Work Trend Index found that Copilot users summarized a missed meeting in 11 minutes and 13 seconds, against 42 minutes and 34 seconds without it — nearly four times faster, and specifically useful for catching up on a meeting you weren't in the room for. A tool built to make absence less costly is an odd centerpiece for an argument that presence is what makes AI work.
My take
Some of what companies are calling "in-person AI collaboration" is real — I saw it in my own test, in the speed of correcting a teammate's habits and in catching a hesitation a shared screen didn't show. But that's a narrow, specific benefit, not a case for requiring four days a week in an office to use a chatbot or a coding assistant. The AI's output was identical regardless of where we sat. If a company's real reason for an RTO mandate is culture, oversight, or real estate use, "you need to be here to use AI" is a more sympathetic-sounding reason to give than any of those, and that's worth noticing.
Frequently Asked Questions
We Must Return to the Office to Use AI in Person — is that actually true?
Not for the tools themselves. In my test, GitHub Copilot and Notion AI produced identical results whether we were in the same room or on a screen-share call. What presence changed was how fast we caught each other's mistakes and corrected a teammate's habits — a real benefit, but a much narrower claim than "AI requires the office."
How long does it take to test this on your own team?
About two weeks: one week running a task in person with logging, one week running a matched task remotely. A single session won't tell you much; you need at least a few repetitions of each condition to separate a real pattern from a fluke.
What's the easiest way to check this for yourself?
Pick one task you already do with an AI tool, run it once in person with a colleague and once over screen share, and compare notes afterward on what each of you individually caught. The debrief conversation matters more than the task itself.
Why are companies making this argument now?
Several 2026 return-to-office mandates, including Microsoft's, single out AI teams for stricter in-office requirements than the rest of the company, citing collaboration. Critics, including Built In’s analysis, point out that the same companies' AI tools are automating the informal mentorship those mandates claim to protect.
— Last updated: September 9, 2026 · By Vishal Swami, Founder & Lead AI Reviewer, AISagely