Artificial intelligence has begun reshaping nearly every industry. For much of commercial real estate, however, it initially felt distant — abstract, perhaps even overstated.

No office floor has yet been physically vacated by an algorithm.

But financial markets rarely wait for physical confirmation.

Shares of SL Green Realty Corp., Manhattan’s largest office landlord, have weakened this year amid growing concern that generative AI may reduce white-collar employment and, by extension, office demand. Several major investment banks have trimmed their price targets for U.S. office REITs. Revenue growth expectations have softened. The logic driving the selloff is straightforward: if AI reduces headcount, fewer workers will require fewer desks.

It is a clean equation.

It may also be incomplete.

For decades, office space has been underwritten according to a simple formula: labor demand determines headcount; headcount determines square footage.

Developers, landlords and tenants all worked within this framework. Modern office design reflected it. Meetings belong in meeting rooms. Focus work occurs at assigned desks. Breaks happen in designated lounges. Each function receives a labeled space. Anything unlabeled is deemed inefficiency — “dead space.”

People, too, were classified functionally. Engineers engineered. Accountants accounted. Sales teams sold. Roles were discrete; architecture followed.

If people can be counted, space can be calculated.

Generative AI complicates this arithmetic.

The assumption that AI simply eliminates tasks — and therefore labor — is increasingly being questioned.

A recent article in Harvard Business Review, based on an eight-month field study conducted by Professor Aruna Ranganathan of UC Berkeley’s Haas School of Business and doctoral researcher Xingchi Maggie Ye, examined how generative AI tools reshaped workflow inside a U.S. technology company.

Rather than compressing work, AI expanded it.

Employees accelerated output, widened their scope and crossed traditional functional boundaries. Product managers began writing code. Designers engaged in engineering tasks. Researchers assumed responsibilities previously assigned elsewhere.

The authors described a “self-reinforcing cycle”: AI increases productivity; heightened productivity raises expectations; rising expectations intensify workloads, driving further AI use.

In this model, headcount may decline. But the density, simultaneity and breadth of work increase.

If one individual can now perform the functional output of several roles, what does that imply for space calibrated purely to attendance?

The office has traditionally been valued as a container of seats — measurable, divisible, rentable.

Yet work in the AI era appears less linear and less compartmentalized. A single day may include deep cognitive focus, real-time collaboration with AI systems, cross-functional decision making and iterative development requiring heavier digital infrastructure.

Work becomes layered.

Roles blur.

Activity intensifies.

The relevant metric may gradually shift from seat density to activity density. 

Architecture, however, has not always been organised around strict programmatic labels.

Traditional Korean courtyard houses — hanok — formally designated certain rooms for specific purposes. Yet between the primary living quarters and the outer walls often existed ambiguous transitional spaces. These areas were neither strictly corridors nor fully rooms; neither inside nor outside.

They had no official function.

Architectural scholars describe these spaces as deliberately undefined — left unprogrammed to accommodate informal life, unplanned gatherings or emotional moments that could not be formally assigned a name. Their ambiguity was not inefficiency. It was flexibility.

Modern office design largely eliminated such spaces in pursuit of optimization. Undefined areas became “dead space.” Efficiency demanded calarity.

Yet as human roles themselves grow less defined and more fluid, rigidly programmed environments may prove less adaptable than spaces capable of absorbing uncertainty.

Office towers optimized for predictability — with low ceiling heights, dense structural grids, inflexible cores and limited electrical upgrade capacity — were built for stable role definitions and steady occupancy patterns.

Fluid work environments expose those assumptions.

Assets with large, simple floorplates, generous ceiling heights, scalable mechanical systems and adaptable power distribution offer a different proposition: structural optionality.

This is not solely a prime-location thesis. Location remains essential. But in an era of technological acceleration, physical flexibility may become equally decisive.

Occupancy measures how many seats are filled.

Adaptability measures how many futures a building can support.

Artificial intelligence is frequently portrayed as the executioner of office real estate. It may instead operate as a filter.

AI does not necessarily erase the need for space. It alters how space is used — and exposes which buildings were designed for a headcount-driven era that may be fading.

If labor intensity rises even as total labor contracts, square footage will not disappear uniformly. It will differentiate.

Some assets, constrained by structural rigidity, may struggle to justify reinvestment. Others — defined by spatial flexibility and mechanical scalability — may function less as containers of attendance and more as infrastructure for cognitive production.

The central question is no longer whether offices survive.

It is whether the headcount formula used to measure them survives.

Artificial intelligence may not empty skyscrapers overnight.

But it may decisively reveal which buildings were built for yesterday’s arithmetic — and which are prepared for what comes next.