What AI Takes Out of the Building

The first subtraction is desk volume. Drafting, formatting, reconciling and first-pass analysis are moving to software. The floor space that housed that work, rows of workstations and the storage and back-office functions around them, has less reason to exist.

What remains is denser in meaning but not necessarily in people. Rooms for decisions, client work, teaching and the conversations where trust is built. These ask for different things from a building: better acoustic separation, flexible meeting space, good air and light, reliable connectivity, and services that cope with uneven occupancy, empty on Friday and full on Tuesday.

For the base building, uneven occupancy is a services problem. Plant sized for steady full occupancy runs inefficiently at partial load. Older systems without good zoning heat and cool whole floors for a handful of people.

The second subtraction is in operations. Building management has been largely a matter of schedules, set points and someone responding to complaints. Analytics and AI-driven controls can now predict loads, detect faults and tune plant continuously. That only works if the building can be read and controlled: sensors, metering, networked controllers and plant that responds to commands.

Software-defined operation exposes the age of the controls. A building with a closed, obsolete management system cannot benefit, however good the software above it.

Many buildings from the 1990s and 2000s carry management systems whose controllers are no longer supported, built on proprietary protocols that are hard to integrate. Their replacement used to be discretionary. It is becoming the condition of participating in how buildings will be run.

Reading a building’s readiness for software-defined operation comes down to a few questions. What is the building management system, which version, and is it supported? Are the controllers on an open protocol, such as BACnet, or a closed proprietary one? How much of the plant is metered and sub-metered, and how much reports its state? Do the major items, chillers, boilers or heat pumps, air-handling units, accept commands at variable output, or only on and off? And is there a secure network path from the building’s systems to the software that would analyse them, separated from tenants and the internet in the way current cyber guidance expects?

Many buildings of the 1990s and 2000s answer poorly on most of these. That is not a crisis. It is a capital line, and it is often one of the cheapest ways to improve both performance and energy ratings in an older building.

AI also adds. It adds compute, which needs power and cooling, whether in a data hall or a server room on a floor. It adds connectivity requirements, from fibre into the building to coverage across it. It adds monitoring, which turns parts of the building into data sources. The building-as-host argument in The Building as Host to Machines develops this further.

If AI were simply emptying offices, the evidence would show it in vacancy. What it shows instead is concentration, and in places AI is adding demand rather than removing it.

The subtraction this paper describes is real at the level of tasks and desks. At the level of buildings, the first effect has been to sharpen the divide between space that people want to be in and space they do not. The capital question is the one this paper ends on: whether the base building’s services and controls can support fewer, more purposeful occupants.

The capital consequence is a shift in where money goes: less into floor-plate fit-out at scale, more into the base-building systems that make a smaller, better-used building work. Zoned and responsive mechanical plant, modern controls, electrical capacity, connectivity, and the quality of the shared spaces that bring people in.

The dividing line in office stock may not run between prime and secondary at all. It may run between buildings whose services and controls can adapt to fewer, more purposeful occupants, and buildings that cannot without a deep refurbishment. Much of this depends on how far and how fast firms actually change the way they work, and occupancy evidence over the next few years will test it.