At Energia AI we work on that gap. Energia AI is an AI control platform for commercial HVAC that sits on top of the BAS a building already has, instead of replacing it. This post is about one principle that shapes everything we do: understand how the building responds before touching a single setpoint.

Why fixed schedules leave value on the table

A typical BAS sequence says something like "occupied mode starts at 7:00, supply air setpoint 55°F, zone setpoints 72°F." That logic is predictable and safe, which is why it exists. But it treats every Monday the same and every building the same. It doesn't know that the east zones heat up fast on sunny mornings, that a lobby takes two hours to recover after a cold night, or that half a floor sits empty on Fridays.

Operators know a lot of this from experience. The controls usually don't.

Learning the building first

Before any control decision, we build a model of how this specific building responds:

• how zone temperatures move with outdoor temperature and solar gain
• how quickly spaces recover after setbacks
• how occupancy changes internal loads across the day and week
• how the plant and air side respond when setpoints change

That model comes from the building's own trend data, read through the protocols already on site such as BACnet, Modbus or LonWorks. A model you can't check against measured outcomes isn't useful, so we judge it by how well it predicts what the building actually did, not by how attractive its savings estimate looks.

Comfort is reported next to energy, every time

It is easy to save energy in a building: let it get uncomfortable. Any honest evaluation of an HVAC control strategy has to show occupied comfort and ventilation right next to energy and cost, for the same period, with the same data. If a strategy saves energy by drifting zones out of range, that is a failure, not a result.

We hold ourselves to that internally too. When a simulated policy looks great on cost but poor on comfort compliance, the comfort number is the one we work on.

Safety and trust come before control

Writing to a live BAS is a serious step, and facility teams are right to be careful. The path we follow is gradual:

1. Read-only inventory. Map points, trends and equipment without writing anything.
2. Shadow mode. Compute what we would have done and compare it with what the building actually did.
3. Isolated testing. Exercise commands against test hardware, not occupied space.
4. A bounded field pilot, only with the owner, the BAS integrator and a qualified engineer involved and the commissioning steps approved.

A few details matter more than they look:

• Command priority is site-specific. Which BACnet priority level a supervisory layer writes at has to be agreed per point with the integrator, not assumed.
• Network loss does not release a command by itself. A latched command can stay in place after the connection drops, so a local watchdog or command lease, plus reading back the effective value, has to be tested before anything goes live.
• Operators keep the final say. Overrides have to be simple, visible and respected.
• Savings need measurement and verification. Claims should rest on metered purchased energy, measured against a site-specific M&V plan, not on model output.

Beyond control

Once the building data is connected, the same foundation supports energy and carbon tracking, tenant cost allocation from metered usage, and an assistant that answers questions from the building's own data, so operators don't have to dig through trend logs to find out why a floor was warm yesterday.

What we'd like to hear

If you commission, integrate or operate building automation systems, we'd value your view. What would you need to see before trusting an AI layer to write setpoints in your building? What has gone wrong with supervisory control products you've worked with?

You can find more about Energia AI at energiaigroup.com.