What RTO Means for AI

Disclosure: These views are my own and do not represent my current or any former employers.

Executive summary #

The COVID-19 pandemic forced a large part of the North American knowledge workforce to work from home. The change was introduced as an emergency measure, not as a planned experiment in organisational design.

It still revealed something important.

Remote work did more than move office tasks into spare rooms and kitchen tables. It expanded the range of lives that could coexist with an engineering career. People could work without relocating to a small set of expensive cities. Disabled and neurodivergent workers could shape their environments around access needs. Parents and other caregivers could recover time previously lost to commuting. People whose responsibilities, communities, or ways of living did not fit the conventional corporate schedule could remain in demanding technical work.

I experienced that change on my own farm. Working remotely allowed me to care for animals when they required attention while continuing to meet the responsibilities of engineering work. The farm was not a distraction from professional life. It was part of a life that made sustained, high-quality work possible.

This paper calls the broader value lived experience engineering: the contribution made when engineers can retain meaningful contact with the bodies, families, places, responsibilities, and communities that shape how people encounter technology.

Return-to-office policy is therefore more than a decision about location. It is a workforce-selection rule. A requirement may be written equally for everyone while imposing very different costs. The employee with a short commute and few care obligations receives one policy. The employee managing a disability, a farm, childcare, elder care, or a long journey receives another in practice.

The available evidence does not show that offices are useless or that remote work is always superior:

• A randomised trial of 1,612 professional employees found that two work-from-home days per week reduced resignations by one-third without harming performance, promotions, or software engineers' code output.[4]
• United States studies of return-to-office mandates have found lower employee satisfaction, higher turnover, loss of senior employees, and increased hiring difficulty, without clear evidence of improved firm performance. Several of these studies remain working papers and should be treated accordingly.[5][6][7]
• Peer-reviewed United States research found that the post-pandemic increase in working from home explained 68 to 85 percent of the rise in full-time employment among people with physical disabilities.[9]
• Canadian data show that the number of employed people with disabilities receiving work-from-home accommodation more than doubled between 2017 and 2022.[10]
• In-person proximity can improve feedback, mentoring, weak ties, and some forms of creative idea generation. Remote work can make collaboration networks more siloed if organisations do not deliberately compensate.[15][16][17]

Most quantitative workforce evidence in this paper comes from the United States. Canadian public data and international studies are identified where they are used.

There is no direct causal evidence that return-to-office mandates produce worse AI systems. This paper does not claim otherwise.

The argument is narrower. AI systems inherit choices about users, risks, data, evaluation, acceptable failure, and the meaning of harm. Those choices are made by people. Policies that change who can enter, remain in, and advance through engineering organisations may also change the range of experience available when those choices are made.

The proposal is not to close offices. It is to stop treating routine attendance as evidence of contribution:

• Use offices for work that benefits from physical proximity.
• Make attendance role-based and task-based rather than universal.
• Preserve remote access for work that can be performed well remotely.
• Build mentoring and cross-team connection deliberately.
• Make information, promotion, and influence available regardless of location.
• Measure who leaves, who stops applying, and whose career progression changes after a policy shift.
• Evaluate engineering outcomes rather than badge activity.

The important question is not how many people returned to a building.

It is what experience, judgement, and capability the policy caused the organisation to lose.

1. A life that fits the work #

My working life does not end at the edge of a corporate calendar. I live on a farm. Animals sometimes require attention at times that are inconvenient to a conventional office routine.

Remote work gave me a practical way to meet those responsibilities without lowering the standard of my engineering work. I could handle what needed to be handled, return to the problem in front of me, and work in an environment that supported concentration and judgement.

This was not an arrangement where personal life replaced professional responsibility. Work still had deadlines, incidents, meetings, decisions, and people depending on the result. The difference was that the structure of the day did not require me to pretend the farm and the engineer were separate people.

That integration made my working life more sustainable. It also made my work better.

The experience is personal, but the principle is wider. People bring many responsibilities and conditions to work:

• children and other family members who need care;
• disabilities and chronic illnesses;
• sensory and communication needs;
• long or unreliable commutes;
• rural and remote communities;
• religious and cultural obligations;
• volunteer work and public service;
• homes, land, animals, and local ties;
• lives built outside the corporate centres where large employers tend to locate.

The office model often treats these facts as private exceptions. The worker is expected to arrive as a standardised professional unit, available at the same place and time, able to tolerate the same environment, and prepared to keep the rest of life outside.

Some people prefer that separation. Offices can provide focus, social contact, reliable facilities, and a boundary between home and work. The problem begins when one arrangement is treated as the definition of commitment.

Remote work showed that a broader range of lives could support serious engineering careers. Return-to-office mandates risk narrowing that range again.

2. COVID-19 became an accidental inclusion experiment #

Before 2020, remote work was often treated as a special arrangement. The pandemic turned it into infrastructure.

The change happened unevenly and under terrible conditions. People worked through illness, grief, closed schools, inadequate space, and social isolation. Emergency remote work should not be romanticised as a clean test of what well-designed distributed work could achieve.

It nevertheless demonstrated that many activities previously described as office-dependent could continue elsewhere.

In the United States, full days worked at home accounted for 28 percent of paid workdays among people aged 20 to 64 by mid-2023, around four times the 2019 rate.[1] In Canada, the share of workers performing most of their hours from home rose from about 7 percent in 2016 to 40 percent in April 2020, then settled near 20 percent in November 2023.[2]

This persistence matters. The emergency ended. A substantial change in how work is organised remained.

Remote-capable work is not universal

The benefits described in this paper are not available to everyone.

Dingel and Neiman estimated that 37 percent of United States jobs could be performed entirely from home, representing 46 percent of wages.[3] The difference between those figures reflects an important inequality: remote-capable work is more common in professional and higher-paid occupations.

Nurses, builders, drivers, retail staff, hospitality workers, laboratory workers, and many others cannot move the essential functions of their jobs into a home office. A remote-work argument that ignores them can become another account of professional privilege.

Even within remote-capable work, access to quiet space, reliable broadband, suitable equipment, and a safe home environment is unequal. Flexibility should include access to an office, not replace one compulsory location with another.

The appropriate conclusion is not that remote-capable employees should lose flexibility because other jobs require presence. It is that workplace policy should begin with the work actually being performed.

The presence requirements of one role should not be imported into another merely in the name of consistency. Equality of scheduling is not equality of work.

The experiment revealed hidden feasibility

Before the pandemic, organisations could reasonably say they did not know whether large-scale remote work was possible.

They know now.

The relevant question has changed from whether remote-capable work can happen to where particular tasks happen best, what controls are required, and how organisations preserve learning and connection.

A return to universal assumptions is not a return to evidence. It is a choice to disregard what the experiment revealed.

3. The worker the office assumes #

Every workplace is designed around an implied worker.

The conventional office assumes a person who can:

• live within commuting distance or relocate;
• travel at fixed times;
• remain in a shared physical environment for most of the day;
• organise care outside working hours;
• tolerate the office's noise, lighting, temperature, and social demands;
• be visible when leaders are visible;
• separate professional identity from the rest of life.

No single worker matches this template perfectly. Some are much closer to it than others.

The template becomes powerful when organisations mistake compatibility with the office for professional merit.

Conformity consumes attention

Kenji Yoshino and Christie Smith described covering as the pressure to downplay aspects of identity in order to fit mainstream expectations. In a survey of 3,129 employees across seven industries, 61 percent reported pressure to cover some part of themselves at work.[14]

Covering did not begin with the office, and remote work does not eliminate it. Digital workplaces can reproduce the same norms through cameras, schedules, language, and access to leadership.

Physical offices can make conformity especially visible. Clothing, food, mobility, speech, family interruptions, sensory needs, and social habits are continuously observable. A person may spend part of the working day managing how their life appears rather than applying the knowledge that life has given them.

The cost is not limited to comfort. Yoshino and Smith argued that people were included on terms that required them to manage identity in ways dominant groups did not. That management can weaken commitment and conceal useful context.

Flexibility can carry a career penalty

Remote work can also create a different form of pressure.

A 2024 United States survey experiment involving 518 employed parents found that respondents expected remote workers to receive fewer job rewards, including pay and promotion, even while expecting greater family satisfaction.[13]

This is the flexibility stigma. A worker may be permitted to work remotely while being treated as less serious for doing so.

The result is a false choice:

• return to the office and absorb the personal cost; or
• retain flexibility and accept reduced visibility, influence, or progression.

A credible remote or hybrid policy must address both location and power. Access to a laptop is not inclusion if decisions, sponsorship, and promotion remain office-centred.

4. Return to office is workforce selection #

Return-to-office debates are often presented as a contest between employee preference and managerial authority.

That framing is too narrow.

An attendance policy changes the cost of holding a job. When the cost changes unevenly, the composition of the workforce changes too.

Equal rules can impose unequal costs

Consider a three-day office requirement.

For one employee, it may add a short commute and welcome social contact.

For another, it may require:

• several hours of weekly travel;
• paid care;
• relocation;
• disclosure of a disability;
• loss of a carefully controlled sensory environment;
• reduced ability to manage an unpredictable health condition;
• withdrawal from a community or way of life.

The written policy is the same. The price is not.

When an organisation says that people who cannot pay that price are choosing to leave, it avoids its own role in setting the price.

The evidence does not show a general performance gain

Ding and Ma examined return-to-office mandates among S&P 500 firms using employee ratings and firm performance data. Their working paper found reduced employee satisfaction without evidence of improved financial performance or firm value.[5]

This does not prove that no individual office policy can improve a particular organisation. Firm-level financial outcomes are blunt measures, and companies adopting mandates may differ from those that do not.

It does weaken the claim that compulsory attendance is a generally established route to better performance.

The people who leave are not random

A separate working paper studied 54 large United States technology and financial firms using more than three million employee profiles. The authors reported a 13 to 14 percent increase in abnormal turnover after return-to-office announcements.[6]

The increase was larger among women, managers, and highly skilled employees. Vacancy duration rose and hiring rates fell. Employees leaving mandate firms became more likely to accept lateral or lower-ranked positions in exchange for flexibility.

These findings use online employment records and inferred worker characteristics. The paper is not yet peer-reviewed, and its estimates should not be treated as settled fact.

The pattern is still strategically important. If the employees most able to leave are experienced, skilled, or underrepresented, a mandate can remove knowledge that is costly to rebuild.

Research examining three large United States technology companies found a shift towards shorter tenure and below-senior positions after return-to-office mandates. The authors attributed this to longer-tenured and senior workers leaving for larger competitors.[7] This study is also a preprint, covers only three firms, and does not directly measure product quality.

Gartner survey data point in a similar direction. Among 2,080 knowledge workers, intent to stay was lower under strict return-to-office mandates, with larger differences among high performers, women, and millennials.[8] Intent is not actual turnover, but it identifies where risk may concentrate.

Selection also happens before hiring

The workforce affected by a location policy includes people who never join.

A job tied to a particular office excludes applicants who cannot relocate, cannot commute, or have built lives elsewhere. The exclusion is usually invisible because those candidates never enter the hiring data as employees.

Remote work turns geography from an employment requirement into a coordination problem. That is not always easy. Time zones, tax rules, employment law, and team cohesion still require design.

It expands the available labour market and the set of lived contexts represented inside it.

5. Who remote work made room for #

Remote work did not produce equal benefit for every worker. It did create measurable access for groups whose needs were previously treated as exceptions.

People with disabilities

A 2026 peer-reviewed study used United States labour data to examine whether working from home contributed to the post-pandemic rise in disability employment. The authors found that a one percentage point increase in working from home within an occupation increased full-time employment among people with physical disabilities by about one percent. They estimated that working from home explained 68 to 85 percent of the post-pandemic rise in full-time employment for that group.[9]

The result concerns physical disability, not every disability. It does not mean remote work is sufficient for accessibility.

It provides unusually strong evidence that work location can change who participates in the labour market.

Canadian data show a related shift. In 2022, 11 percent of employed Canadians with disabilities needed working from home as an accommodation, up from 8 percent in 2017. The number who received that accommodation rose from about 120,000 to 245,000.[10]

The United States Equal Employment Opportunity Commission states that telework may be a reasonable accommodation under the Americans with Disabilities Act when essential job functions can be performed from home.[22] Accessibility Standards Canada likewise documents practical measures for accessible remote meetings, including captioning, interpretation, accessible materials, and alternatives for weak internet connections.[23]

Remote work is not merely a preference in these cases. It can be the difference between performing the job and being excluded from it.

Neurodivergent professionals

A qualitative study of 36 United States neurodivergent professionals found that working from home allowed participants to create accessible physical and digital spaces, negotiate communication practices, and manage tensions between productivity and wellbeing.[11]

The study also reported difficulty. Some participants struggled with structure, boundaries, and isolation.

This is exactly why universal claims are weak. The same environment can remove barriers for one person and create them for another.

Choice and adaptation matter more than declaring one location inherently inclusive.

Parents and caregivers

Care responsibility remains unevenly distributed, especially by gender.

Harrington and Kahn found that greater work-from-home feasibility in United States occupations narrowed employment gaps between mothers and other women. Women able to work from home before childbirth were less likely to leave employment afterwards.[12] The study is a working paper and focuses on employment rather than promotion or leadership advancement.

The finding supports a practical mechanism. Removing commuting and adding control over location can make demanding careers compatible with care.

It does not make care free. Remote workers can experience longer hours, blurred boundaries, and simultaneous demands. Caregiving during paid work can become unsustainable when organisations assume flexibility means constant availability.

The value is not that employees can perform two full-time roles at once. It is that they can recover time and respond to life without every interruption becoming an absence from work.

Rural and place-bound lives

Remote work also made engineering careers more compatible with place.

Corporate engineering has concentrated opportunity around a limited set of cities and campuses. The requirement to relocate selects for people able and willing to leave family, land, community, housing, and local obligations.

My farm is one example. Other people remain tied to a place through culture, cost, care, environmental preference, or community contribution.

Geographic flexibility does not guarantee geographic diversity. Companies can still hire from the same networks and promote the same profiles.

It removes one automatic exclusion: the idea that professional ambition requires everyone to organise life around corporate real estate.

6. Lived experience is engineering input #

Engineering is often described as the application of technical knowledge to defined problems.

The definition hides an earlier decision: which problem is defined, for whom, and what counts as failure.

Those decisions depend on experience.

Defining lived experience engineering

Lived experience engineering is the practice of treating knowledge gained through a person's body, responsibilities, geography, identity, community, and daily life as relevant input to technical work.

It does not mean:

• every personal opinion is technically correct;
• identity substitutes for expertise;
• one person can represent an entire group;
• diverse teams automatically produce good systems.

It means that technical judgement is not formed by technical training alone.

An engineer who depends on captions may notice when audio is treated as the default. A caregiver may recognise that a workflow assumes uninterrupted time. A rural user may understand what unreliable connectivity does to a cloud-only design. A person with a mobility impairment may question an identity check that assumes everyone can reach a particular location. Someone familiar with livestock, weather, public services, education, or shift work may see assumptions that an office-centred team does not.

These observations still require evidence, design, and engineering discipline. They change which questions enter the room.

Diversity changes available information

A qualitative study of some of Google's most gender and racially diverse software engineering teams found that homogeneous perspectives affect how, why, and for whom software is designed. The researchers identified recruiting, hiring, and inclusive team practices that reinforced diversity over time.[18]

A study of more than 23,000 GitHub projects found small but positive associations between gender and tenure diversity and team productivity after controlling for several project characteristics.[19] The effects were modest, observational, and specific to open-source software.

These studies do not prove that any single demographic mix guarantees quality.

They support a less ambitious conclusion: changing who participates can change the information, experience, and questions available to a team.

Inclusion determines whether experience is used

Representation alone is insufficient.

If a person must hide care obligations, avoid discussing disability, imitate the dominant career path, or remain silent to preserve credibility, the organisation may employ them without receiving the benefit of their experience.

Lived experience engineering requires:

• psychological safety to raise an unfamiliar risk;
• decision processes that record dissent;
• evaluation that includes affected users;
• career systems that do not punish flexibility;
• managers willing to distinguish different methods from lower standards.

Remote work can support these conditions. It can also undermine them if remote employees are absent from informal decisions.

The issue is not where diversity sits. It is whether it can influence the work.

7. Why AI raises the value of lived experience #

AI systems operate across contexts their builders cannot fully anticipate.

They classify, summarise, recommend, generate, and increasingly act. Small assumptions can be repeated across millions of interactions.

Data does not choose the right question

AI development often places great weight on data scale. More data can improve coverage. It cannot decide by itself:

• which groups require separate evaluation;
• what an acceptable error rate is;
• whether a proxy is harmful;
• which omissions matter;
• what recourse a user needs;
• when automation should not be used.

People make those choices.

The Gender Shades study showed why the evaluation question matters. Three commercial gender-classification systems had error rates of up to 34.7 percent for darker-skinned women, compared with a maximum of 0.8 percent for lighter-skinned men.[20]

The study used binary gender labels and a particular skin-type classification, which limits what it can represent. It does not show that a remote team would have prevented the disparity. It shows that aggregate accuracy can conceal severe differences until someone tests the system across relevant lived groups.

AI increases the cost of narrow problem framing

Traditional software can embody narrow assumptions. AI makes some of those assumptions harder to inspect because behaviour is probabilistic and context-dependent.

Teams need people able to ask:

• Who is missing from the test set?
• Whose language is treated as abnormal?
• Which user cannot follow the expected recovery path?
• What happens when the household, body, connection, or schedule differs from the assumed norm?
• Which failure looks minor to the builder and consequential to the user?

These are technical questions. They are also questions of experience.

The direct RTO-to-AI evidence does not exist

I found no causal study showing that return-to-office mandates reduce AI safety, fairness, accessibility, or product quality.

The relationship proposed here is a mechanism, not an established result:

• Location policy changes the cost of employment.
• Different workers experience that cost differently.
• The policy affects who applies, stays, and advances.
• Workforce composition affects available knowledge and problem framing.
• Problem framing can affect which AI risks are identified and tested.

Each step has supporting evidence or a clear rationale. The complete causal chain has not been demonstrated.

That distinction matters. A serious argument should identify what is known and what remains to be tested.

8. What offices do well #

An argument for flexible work becomes weak if it denies the benefits of physical proximity.

Those benefits are real.

Mentoring and feedback

Emanuel, Harrington, and Pallais studied software engineers at a Fortune 500 firm. Engineers working in the same building as their teammates received about 22 percent more online feedback than engineers whose teammates were elsewhere.[15]

The benefits were concentrated among less-tenured workers. Proximity supported learning, but it also carried an output cost for senior engineers who spent more time providing support.

This finding supports offices as places for apprenticeship.

It does not establish that every engineer needs the same number of office days indefinitely.

Cross-team connection

A study of 61,182 employees at a large United States technology company found that the pandemic shift to firm-wide remote work made collaboration networks more siloed. Cross-group collaboration fell by about 25 percent relative to the pre-pandemic level, and communication became more asynchronous.[16]

The authors used pre-pandemic remote workers to separate remote-work effects from wider pandemic disruption. The setting still included an extraordinary crisis, and collaboration traces do not directly measure customer value or product quality.

The risk is credible. Distributed organisations need deliberate ways to create weak ties and move information between groups.

Creative idea generation

Brucks and Levav found that pairs using videoconferencing generated fewer creative ideas than pairs meeting in person. The effect appeared in laboratory experiments and a field study across five countries. Virtual pairs were not worse at selecting which idea to pursue.[17]

This suggests that location can be matched to task:

• physical rooms for divergent ideation;
• remote work for concentration and execution;
• either environment for evaluation, depending on the work.

Connection can be episodic

Atlassian reports that intentional team gatherings produced a 27 percent increase in feelings of connection, with effects lasting four to five months. It also reports that most employees visited an office at least once per quarter despite having no attendance mandate.[21]

This is internal corporate evidence, not an independent trial.

It illustrates a useful design alternative: concentrate in-person time around a purpose rather than spreading it across arbitrary weekly quotas.

9. The problem with universal mandates #

A universal mandate is administratively simple. Simplicity is not the same as sound design.

Attendance is an input measure

Badge swipes show where a person was. They do not show:

• whether the right decision was made;
• whether code was reliable;
• whether an incident recovered faster;
• whether a junior engineer learned;
• whether a customer problem was understood;
• whether an AI system was evaluated across relevant users.

An organisation that cannot measure engineering outcomes may substitute attendance because it is easy to count.

That creates the appearance of control without proving performance.

Universal rules shift exceptions onto individuals

When remote work is available only through exception, the employee must justify why their life deserves deviation from the norm.

That may require disclosure of:

• disability;
• illness;
• family circumstances;
• reproductive decisions;
• financial limits;
• responsibilities that colleagues are not required to explain.

The process turns a work-design question into a personal petition.

Some roles require presence. Some tasks benefit from it. The organisation should establish that need before requiring the individual to establish why they cannot comply.

Hybrid evidence supports a middle ground

The strongest randomised evidence in professional work comes from a six-month Trip.com trial involving 1,612 employees in engineering, marketing, finance, and accounting. Employees assigned to two work-from-home days per week had one-third lower resignation rates, with no detectable harm to performance reviews, promotions, or code output.[4]

The study took place in China, not North America. It tested a managed hybrid schedule, not fully remote work.

Its lesson is not that every organisation should copy the exact schedule. It is that a substantial reduction in office attendance can preserve measured performance while improving retention.

The burden of proof should therefore sit with a mandate's claimed purpose.

10. A better operating model #

The alternative to a universal mandate is not an absence of management.

Flexible work requires clearer management.

1. Start with the work

Classify activities by what they require:

Work	Likely best setting	Reason
Focused implementation and analysis	Employee choice or quiet environment	Concentration and control
Architecture disagreement	Synchronous, sometimes in person	High-bandwidth discussion
Early-career mentoring	Structured proximity plus documented follow-up	Feedback and observation
Incident response	Distributed-ready	Incidents do not respect office days
Brainstorming and relationship building	Intentional in-person gathering	Idea generation and trust
Review and decision record	Remote-accessible by default	Participation and auditability
Customer or site-specific work	Where the customer or system requires	Direct operational need

The table is a starting point. Teams should test it against their work.

2. Replace weekly quotas with purposeful presence

If an organisation wants people together, it should be able to state:

• who needs to attend;
• what work will be done;
• why physical presence improves it;
• how often the benefit is required;
• what remote participants need;
• how the result will be measured.

"Three days" is a schedule. It is not a purpose.

3. Make remote access the information default

Distributed work fails when the real decision happens in a room and the document appears later.

Teams should:

• write proposals before consequential meetings;
• record decisions and dissent;
• publish notes promptly;
• use captions and accessible materials;
• avoid making hallway access a prerequisite for influence;
• give remote participants equal ways to speak and review.

These practices also improve office-based work across buildings and time zones.

4. Build mentoring as a system

Offices often provide mentoring accidentally through proximity. Remote work removes some of that accident.

The answer is not to leave development to chance again.

Use:

• assigned mentors and sponsors;
• regular code and design review;
• paired work;
• office hours;
• rotation through incidents and architecture reviews;
• intentional in-person onboarding;
• explicit expectations for senior engineers to teach.

Mentoring should be measured as work, not treated as invisible generosity.

5. Make progression location-neutral

Promotion systems should examine:

• whether remote workers receive fewer stretch assignments;
• whether office attendance predicts ratings after controlling for outcomes;
• who receives informal sponsorship;
• whether important meetings include distributed staff;
• whether leadership roles require relocation without a task-based reason.

A hybrid policy without location-neutral advancement becomes an office mandate through career pressure.

6. Preserve individual adaptation

Teams need common operating agreements. Individuals also need room to manage energy, disability, care, and environment.

The standard should be reliable contribution, not identical routines.

11. Measure what the policy removes #

Most return-to-office dashboards measure compliance.

An accountable policy should measure consequences.

Area	Questions
Retention	Who left after the policy changed? What were their level, tenure, performance, and role?
Hiring	Did applicant volume, geographic reach, offer acceptance, or time-to-fill change?
Career	Did promotion, ratings, sponsorship, or high-impact assignments differ by work location?
Accessibility	Did accommodation requests, denials, leave, or exits change?
Care	Did employees report new care costs or reduced ability to remain in role?
Engineering	Did delivery, defects, incidents, review latency, or reliability change?
Learning	Did feedback, mentoring, and new-hire ramp time improve?
Collaboration	Did cross-team ties, decision speed, or duplicated work change?
Inclusion	Did belonging, covering pressure, or willingness to raise risk change?
AI and product risk	Did evaluation coverage, accessibility findings, user complaints, or escaped harms change?
Office value	What work occurred during visits, and did it require physical presence?

These measures should be aggregated and privacy-preserving. Organisations should not turn inclusion analysis into surveillance of disability, family status, or private life.

Examine the missing population

Employee surveys describe people who stayed.

Policy review should also examine:

• candidates who withdrew after learning the location requirement;
• employees who declined relocation;
• former employees who accepted less senior roles for flexibility;
• people who never applied because the job was office-bound;
• workers who obtained exceptions but lost career access.

The most consequential effect may sit outside the current employee dataset.

Compare against the stated goal

If the goal is mentoring, measure mentoring.

If the goal is innovation, measure idea generation and product outcomes.

If the goal is culture, define the behaviours that constitute culture.

If the goal is performance, measure performance.

Do not claim success because attendance increased.

12. What this paper does not claim #

This paper does not claim that every job can be remote. Most cannot.

It does not claim that every home is a good workplace. Some people need the equipment, safety, quiet, or separation an office provides.

It does not claim that remote work removes bias, covering, isolation, or career inequality. Poorly designed distributed organisations reproduce all of them.

It does not claim that care responsibilities should routinely interrupt work without coordination. Flexibility does not remove accountability to colleagues or customers.

It does not claim that offices have no value. Evidence supports their use for mentoring, weak ties, relationship building, and some forms of creative collaboration.

It does not claim that diverse teams automatically build fair or reliable products. Inclusion, authority, testing, and engineering practice determine whether different experience affects the result.

It does not claim that return-to-office mandates have been proven to damage AI systems. No direct study establishes that causal link.

The claim is this:

Remote work expanded the set of lives compatible with remote-capable engineering. Return-to-office mandates can reverse part of that expansion by changing who can afford to join, remain, and advance. In an era when AI systems need broader human judgement, organisations should understand that workforce change as a technical and strategic cost.

The public-source review for this paper was completed on 24 August 2026.

Conclusion #

The pandemic did not invent remote work. It made organisations confront how much of office life was habit rather than necessity.

One unexpected result was a wider definition of who could sustain an engineering career and what that career could look like.

For me, it meant that caring for animals on a farm and doing serious engineering work did not have to be competing identities. For others, it meant access, care, geography, health, or community no longer had to be negotiated around a daily commute.

That expansion did not solve workplace inequality. It did reveal that some exclusions were designed into the location of work.

Offices remain useful. The right response is to use them deliberately:

• for the work that benefits from proximity;
• at a frequency supported by purpose;
• with equal information and career access for people elsewhere;
• without treating one lifestyle as the proof of professionalism.

AI raises the stakes because technical systems increasingly make decisions across lives their builders may never have lived. Organisations need engineers who can recognise different users, constraints, and harms.

The question for return-to-office policy is therefore larger than productivity.

It is whether the organisation is willing to narrow the lives represented in its engineering judgement, and whether it has measured what it loses when it does.

Appendix A: Return-to-office policy review #

Policy:
Effective date:
Roles in scope:
Named owner:
Review date:

## Stated objective

[State the outcome the policy is intended to improve.]

## Work analysis

| Activity | Presence required | Evidence | Frequency |
| --- | --- | --- | --- |
| ... | Yes/No/Sometimes | ... | ... |

## Baseline

- Attrition:
- Applicant volume:
- Offer acceptance:
- Time-to-fill:
- Promotion rates:
- Engineering outcomes:
- Mentoring and feedback:
- Collaboration:
- Accessibility requests:
- Employee sentiment:

## Distributional impact

[Assess effects by role, tenure, level, geography, commute, and privacy-safe demographic groups.]

## Alternatives considered

- Intentional gatherings:
- Role-based attendance:
- Team-determined schedules:
- Structured mentoring:
- Remote-first documentation:
- Accommodation by default:

## Results

[Compare outcomes with the stated objective. Attendance alone is not an outcome.]

## Decision

Continue / revise / stop:
Evidence:
Residual risk:
Next review:

Appendix B: Questions for engineering leaders #

• What work specifically improves when people are physically together?
• How was the required frequency selected?
• Which employees bear the largest cost?
• Who left or stopped applying after the policy changed?
• Are senior and high-performing employees leaving at different rates?
• Do remote employees receive equal promotion and sponsorship?
• Can employees obtain flexibility without disclosing unnecessary personal information?
• Are accessibility and captioning defaults available to everyone?
• Is mentoring structured or assumed to happen through presence?
• Are office gatherings designed around a clear outcome?
• Does attendance predict performance after controlling for engineering results?
• Which lived experiences may be disappearing from the team?
• What AI or product risks might those experiences have helped identify?
• What evidence would cause the organisation to revise the policy?

About the author #

Jason Doyle writes about reliable software, observability, applied AI, and practical controls for systems that influence human decisions. He publishes at jasondoyle.ie and can be contacted at [email protected].

References #

• Jose Maria Barrero, Nicholas Bloom, and Steven J. Davis, The Evolution of Work from Home, Journal of Economic Perspectives, 2023,https://www.aeaweb.org/articles?id=10.1257/jep.37.4.23. - Statistics Canada, Working from home in Canada, 18 January 2024,https://www150.statcan.gc.ca/n1/daily-quotidien/240118/dq240118c-eng.htm. - Jonathan I. Dingel and Brent Neiman, How Many Jobs Can be Done at Home?, Journal of Public Economics, 2020,https://www.nber.org/papers/w26948. - Nicholas Bloom, Ruobing Han, and James Liang, Hybrid working from home improves retention without damaging performance, Nature, 2024,https://doi.org/10.1038/s41586-024-07500-2. - Yuye Ding and Mark (Shuai) Ma, Return-to-Office Mandates, working paper, 2024,https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4675401. - Yuye Ding, Mark (Shuai) Ma, Betty (Bin) Xing, Yucheng (John) Yang, and Zhao Jin, Return to Office Mandates, Brain Drain and Gender Difference, working paper, 2024, revised 2025,https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5031481. - David Van Dijcke, Florian Gunsilius, and Austin L. Wright, Return to Office and the Tenure Distribution, preprint, 2024,https://arxiv.org/abs/2405.04352. - Gartner, High-Performers, Women, Millennials Are Greatest Flight Risks When Strict Return to Office Mandates Are Implemented, 30 January 2024,https://www.gartner.com/en/newsroom/press-releases/2024-01-30-gartner-hr-research-finds-high-performers-women-millennials-are-greatest-flight-risks. - Nicholas Bloom, Gordon B. Dahl, and Dan-Olof Rooth, Work from Home and Disability Employment, American Economic Review: Insights, 2026,https://www.aeaweb.org/articles?id=10.1257/aeri.20240538. - Christoph Schimmele, Sung-Hee Jeon, and Rubab Arim, Changes in workplace accommodations among employed Canadians with disabilities, 2017 to 2022, Statistics Canada, August 2024,https://www150.statcan.gc.ca/n1/pub/36-28-0001/2024008/article/00004-eng.htm. - Maitraye Das, John Tang, Kathryn E. Ringland, and Anne Marie Piper, Towards Accessible Remote Work: Understanding Work-from-Home Practices of Neurodivergent Professionals, Proceedings of the ACM on Human-Computer Interaction, 2021,https://doi.org/10.1145/3449282. - Emma Harrington and Matthew E. Kahn, Has the Rise of Work from Home Reduced the Motherhood Penalty in the Labor Market?, NBER Working Paper 34147, 2025,https://www.nber.org/papers/w34147. - Stephanie Moller, Leah Ruppanner, and Jill Yavorsky, Do working parents in the United States expect work location to impact job and family satisfaction for mothers and fathers differently?, Frontiers in Sociology, 2024,https://doi.org/10.3389/fsoc.2024.1368594. - Kenji Yoshino and Christie Smith, Uncovering Talent: A New Model of Inclusion, Deloitte University Leadership Center for Inclusion, 2013; summary athttps://www.law.nyu.edu/news/kenji-yoshino-explores-the-cost-of-conformity-at-work. - Natalia Emanuel, Emma Harrington, and Amanda Pallais, The Power of Proximity to Coworkers: Training for Tomorrow or Productivity Today?, NBER Working Paper 31880, 2023,https://www.nber.org/papers/w31880. - Longqi Yang et al., The effects of remote work on collaboration among information workers, Nature Human Behaviour, 2022,https://www.microsoft.com/en-us/research/publication/the-effects-of-remote-work-on-collaboration-among-information-workers/. - Melanie S. Brucks and Jonathan Levav, Virtual communication curbs creative idea generation, Nature, 2022,https://doi.org/10.1038/s41586-022-04643-y. - Ella Dagan, Anita Sarma, Alison Chang, Sarah D'Angelo, Jillian Dicker, and Emerson Murphy-Hill, Building and Sustaining Ethnically, Racially, and Gender Diverse Software Engineering Teams: A Study at Google, ESEC/FSE, 2023,https://research.google/pubs/building-and-sustaining-ethnically-racially-and-gender-diverse-software-engineering-teams-a-study-at-google/. - Bogdan Vasilescu et al., Gender and Tenure Diversity in GitHub Teams, CHI, 2015,https://doi.org/10.1145/2702123.2702549. - Joy Buolamwini and Timnit Gebru, Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification, Proceedings of Machine Learning Research, 2018,https://proceedings.mlr.press/v81/buolamwini18a.html. - Atlassian, Lessons learned: 1,000 days of distributed at Atlassian, 2024,https://www.atlassian.com/blog/innovation/distributed-work-report. - United States Equal Employment Opportunity Commission, Work at Home/Telework as a Reasonable Accommodation,https://www.eeoc.gov/laws/guidance/work-hometelework-reasonable-accommodation. - Accessibility Standards Canada, Best practices for accessibility when working from home,https://accessible.canada.ca/resources/best-practices-accessibility-working-home.

source & further reading

jasondoyle.ie — original article