Comfort speaks in more than one way: What a new BrainBox AI co-authored research paper reveals about buildings

There’s a small moment most of us barely register anymore.

You walk into a room.
It feels too hot.
You adjust the thermostat a notch.
And move on with your day.

That action lasts seconds. But what if it mattered more than we’ve ever allowed it to?

A newly published scientific paper in Energy & Buildings suggests it does. Co-authored by BrainBox AI President, Founder & CTO, Jean-Simon Venne, the paper, Adaptive thermostat preference learning using behaviour nudging and multi-armed bandits: A field implementation, demonstrates that buildings can begin to understand occupant comfort by paying closer attention to what people already do.

At the heart of this paper is something refreshingly human: buildings can learn what people actually prefer, simply by paying attention to how they already interact with their space. No surveys. No wearables. And no loss of occupant control.

Just observation, done carefully.

So, what did the research explore?

For decades, most buildings have relied on fixed temperature setpoints chosen to avoid complaints. While surveys and apps can help determine these setpoints, these methods of feedback are hard to sustain in real buildings.

Let’s be honest. Survey fatigue is something most of us have experienced, often leading to participation becoming rushed, delayed, or skipped altogether. That reality makes this kind of feedback difficult to rely on in buildings and ultimately reveals a gap in how comfort data is collected within building science. That gap is exactly where this research begins. Instead of asking occupants how they feel, it listens to what they do.

From assumed comfort to learned comfort

Rather than relying on solicited feedback or fixed assumptions, the system observes everyday actions, such as thermostat adjustments. Known as adaptive occupant behaviour, this approach allows comfort preferences to be learned gradually through real-world interaction.

A natural companion of this concept is behaviour nudging. Instead of making big or abrupt changes, the system introduces small, careful adjustments and watches how occupants react. If they are comfortable, nothing happens. If they are not, they adjust the thermostat, and that response becomes part of what the system learns.

Behind this process is a learning method known as a multi-armed bandit (MAB), which helps the system compare outcomes over time and learn which temperature settings work best based on occupant responses.

When applied in real buildings, MAB is optimizing for both comfort and efficiency. It “wins” when energy use can be lowered without prompting an occupant response, and it “loses” when an override signals discomfort. Over time, this allows the building’s temperature controls to learn the most efficient setpoint that maintains comfort while achieving energy savings.

A glimpse into the findings

The paper was based on a study that was implemented in an occupied academic facility located in Ottawa, Ontario, Canada. Using naturally occurring thermostat interactions, the system was able to learn preferred temperature ranges in different zones. Importantly, those preferences were not uniform.

In one zone where occupants were comfortable with slightly lower winter temperatures, energy savings of up to 12.7% were achieved. In another zone, where occupants preferred warmer conditions, the system respected that preference. An example of how behaviour nudging and adaptive occupant behaviour work together to shape data comfort learning.

Redefining intelligence in buildings

This research challenges a long-standing assumption in building operations: that comfort must be predefined rather than learned.

With that in mind, the study reframes how building science understands intelligence in buildings.

Not as rigid control.
Not as forcing optimization.
But as adaptive learning applied carefully, responsibly, and with humility.

As building operations face increasing pressure to reduce emissions and improve efficiency, there is a temptation to prioritize energy targets at the expense of human experience. This work shows that the two are not inherently at odds, provided comfort continues to guide how systems learn. By treating thermostat adjustments as meaningful feedback, the system continuously adapts to real occupant behavior.

Smarter buildings start with better listening

For BrainBox AI, intelligence is human-centered by design. The systems we care about are those that support decision-making, remain responsive to real-world conditions, and respect occupant control. The research makes clear that learning only works when people remain part of the loop.

It shows that progress in smarter buildings doesn’t always come from adding more hardware or tightening occupant constraints. Sometimes it comes from paying closer attention to signals that have been there all along.

A note from Jean-Simon Venne

“When people are in a building, whether at home or at work, they rarely describe comfort. They act on it. They adjust the thermostat and move on.

This research asks a simple question. What happens when we stop overlooking those small, everyday actions and start paying closer attention to them?

There isn’t one perfect temperature. Comfort is multi-dimensional, shaped by context, activity, and individual preference. Intelligence in buildings, and building science more broadly, comes from learning and adapting to how people actually live and work in these spaces.”

—Jean-Simon Venne
Co-author | President, Founder & Chief Technology Officer, BrainBox AI

Interested in learning more?

The scientific paper is available online and will appear in Volume 355, Issue 15 of Energy & Buildings (issue dated March 15, 2026).