← All insights
AI

How Agentic AI Controls HVAC: Supervisory Optimization Explained

How Agentic AI Controls HVAC: Supervisory Optimization Explained

Supervisory optimization is the layer of HVAC control that decides which setpoints and operating modes a building’s equipment should run at, while local controllers hold each setpoint. AI control platforms typically work at this layer, recalculating setpoints from operating data and writing them to the building automation system (BAS).

Supervisory setpoints decide much of what HVAC equipment consumes. An air handler running at a fixed duct static pressure setpoint uses more fan energy than one whose setpoint follows what the zones need, and a chiller plant held at a fixed chilled water temperature works harder than one reset to the load. The ASHRAE Handbook puts the penalty for fixed setpoint control typically at 5 to 20 percent of cooling system energy.

The industry has long known which supervisory strategies recover that energy. Supply air temperature reset, static pressure reset, and chilled water reset are standard practice, and ASHRAE Guideline 36 publishes them as uniform, high-performance sequences of operation. In a simulation of a medium-sized commercial building, Lawrence Berkeley National Laboratory (LBNL) found that Guideline 36 sequences cut HVAC energy by an average of 31 percent.

Rule-based resets still leave savings behind, because they apply fixed logic to conditions that change. A reset schedule tuned at commissioning reflects the occupancy, weather, and equipment of that moment, and it drifts out of tune as all three change. Rules also struggle with trade-offs between systems, where the best setpoint depends on the energy use of several components at once. The ASHRAE Handbook describes condenser water control in those terms: the optimum balances cooling tower fan energy against chiller energy.

Supervisory optimization software addresses both limits by recalculating setpoints continuously from trend data, weather forecasts, and utility rates, then writing the results to the BAS. Agentic AI platforms can add AI agents that diagnose why a system misbehaves before optimizing it, and that check whether each change holds. Local controllers, safeties, and life-safety systems stay in place, because the software changes only the targets they work toward. Since the software writes to live equipment, owners need to know what it writes, how the BAS arbitrates its commands, and how savings are verified. This article explains the control hierarchy, the supervisory strategies AI builds on, how AI writes to a BAS, what field studies show about savings, and how to evaluate a platform.

How is HVAC control organized in a commercial building?

AI supervisory control works in the top layer of a control stack that runs from field controllers up to software that sets targets for the whole building.

The ASHRAE Handbook describes the split: the lowest level provides “local-loop control of a single set point,” and the upper level, “typically called supervisory control, specifies the set points and other modes of operation.” Between the two sit the sequences of operation programmed into the BAS, which define operating modes, interlocks, and the reset logic that adjusts setpoints.

Layer What it decides Where it runs Example
Local loop control How far to open a valve or how fast to run a fan to hold one setpoint Field controllers A PID loop modulates a chilled water valve to hold 55°F supply air
Sequences of operation Operating modes, interlocks, and rule-based resets BAS controllers An air handler switches to economizer mode when outdoor air is cool enough
Supervisory optimization Which setpoints and modes the building should run at BAS server, on-site gateway, or cloud software Software raises the chilled water setpoint by 2°F because forecast load is light

Which supervisory strategies save the most HVAC energy?

Air-side resets and schedules typically save more energy than central plant measures. In PNNL modeling of a large office building, many single air-side and scheduling measures cut annual HVAC energy by more than 20 percent, while measures that changed central plant conditions saved less than 5 percent in most climates. The following strategies form the core of supervisory control:

  • Supply air temperature reset: raises the supply air temperature to the highest value that keeps every zone at setpoint, which cuts both cooling and reheat energy.
  • Duct static pressure reset: lowers the supply fan’s pressure setpoint until the most-open VAV box is nearly fully open, which saves fan energy. Guideline 36 implements this reset as trim and respond, in which zones that need more air send requests and the setpoint rises only when enough requests arrive.
  • Chilled water supply temperature reset: raises the chilled water temperature when cooling load allows, at fixed intervals such as every 15 minutes.
  • Condenser water optimization: sets cooling tower and condenser water conditions to balance tower fan and pump energy against chiller energy.
  • Optimal start: starts equipment as late as possible while still bringing zones to comfort conditions at occupancy.
  • Demand-controlled ventilation: matches outdoor air to occupancy, measured through CO2 sensors or occupant counts.

Combining the resets returns more than applying them one at a time. In a PNNL emulation of a large office in Chicago, the combined Guideline 36 resets saved up to 41 percent of energy in the heating season, 18 percent in the shoulder season, and 20 percent in the cooling season, more than the sum of the individual strategies. Field results vary more widely: an LBNL study of 17 Guideline 36 retrofit projects measured HVAC savings from 2 to 49 percent, with a median of 18 percent.

How does AI write setpoints to a BAS?

AI supervisory software writes setpoints through the BAS’s own protocol, most often BACnet, and the BACnet priority array decides whether each command takes effect.

The software computes setpoints with one of several methods. Model predictive control (MPC) uses a model of the building to choose the setpoints that minimize energy or cost over a forecast horizon. Reinforcement learning (RL) learns a control policy from operating data and feedback. Agentic AI platforms add AI agents that diagnose equipment problems before the optimization runs and verify results after it.

BACnet carries each setpoint change as a WriteProperty request with a priority from 1 to 16. Under the BACnet standard, each writable point keeps a 16-slot priority array and acts on the highest-priority command present, and a write without a stated priority defaults to 16, the lowest. The standard's command priority table assigns priorities 1 and 2 to life-safety commands and priority 8 to manual operator commands, so the priority a platform writes at decides which commands outrank its writes.

A BACnet command stays in the priority array until something relinquishes it by writing a null to its slot, at which point the point falls back to the next active command or its relinquish default. A platform that loses its connection can therefore leave its last setpoints in force. Fallback to the building’s own sequences has to be designed in, through a heartbeat the BAS monitors, commands that release on a timer, or software that runs on site and keeps operating when the cloud connection drops. Google’s AI control of data center cooling sets a useful reference design: the local control system verifies each AI action against its own constraints, and operators can return control to on-site rules at any time.

How much energy does AI supervisory control save?

Rigorous field evidence puts typical savings from advanced supervisory control in the mid-teens, below the figures that single-building trials tend to report.

A 2025 review in Applied Energy examined 104 papers reporting field demonstrations of MPC and RL control. The authors found that 71 percent of the demonstrations used experiment protocols that may produce unreliable estimates, including simulated baselines, control of only part of a building, and short test periods, all of which can inflate savings. After filtering for those problems, commercial buildings showed duration-weighted savings of 15 percent in energy and 13 percent in cost.

Single-building results run higher and vary widely. In an LBNL field demonstration, MPC saved approximately 40 percent of HVAC energy in one office building over a two-month trial. The researchers also reported that collecting data and deploying the controller each took as much effort as building the model.

How should owners evaluate AI supervisory control?

Owners should evaluate an AI control platform on what it writes, how it behaves when something fails, and how its savings are measured. The following questions cover those areas:

  • Points and limits: which setpoints the platform writes, and the minimum and maximum it can set on each point.
  • Priority: which BACnet priority the platform writes at, and which operator and safety commands outrank its writes.
  • Fallback: what happens to the platform’s commands if the platform or its connection stops, and how the BAS returns to its own sequences.
  • Advisory mode: whether the platform can recommend setpoints before writing them, so engineers can compare its decisions with the existing sequences.
  • Overrides: how engineers override the platform, and how the platform responds to a manual command.
  • Measurement and verification: which IPMVP option the vendor uses. Option B measures the affected system, and Option C measures the whole facility at the utility meter against a regression baseline.
  • Trial design: how long a trial runs and which baseline it uses, given the biases the Applied Energy review found in short and partial trials.
  • Cybersecurity: how commands travel from the cloud to the BAS, and whether they are encrypted and authenticated end to end. NIST warned in 2026 that the cyberattack risk to building automation systems rises significantly as they connect to corporate networks and the cloud, and BACnet Secure Connect adds TLS encryption and authentication to BACnet traffic.

How does Noda approach supervisory control?

Noda is an agentic AI platform for commercial building operations that works as a virtual building engineer. Noda’s AI agents diagnose equipment and control problems from BAS trend data, and the AI Project Generator turns them into Projects, typically changes to setpoints, schedules, and control sequences, with root causes identified and annual savings estimated.

For peak demand, Automated Demand Management runs supervisory control automatically. As building power nears its limit, ADM raises occupied zone cooling setpoints in small steps, never past each zone’s maximum, and hands every setpoint back to the BAS when the event ends. ADM writes at BACnet priority 8, which leaves life-safety and critical commands at higher priorities in charge. ADM runs on the Noda LaunchPad on site and keeps operating if the cloud connection drops.

Frequently asked questions

Does AI supervisory control replace the building automation system?

No. AI supervisory control writes setpoints to the existing BAS, and the BAS controllers, sequences of operation, and safeties keep running underneath. Some products replace field controllers with their own hardware, which puts them in a different category from supervisory software.

Is supervisory control the same as fault detection and diagnostics?

No. Fault detection and diagnostics (FDD) finds equipment and control problems and reports them, while supervisory control changes setpoints to optimize operation. A supervisory optimizer working on top of a stuck damper or a failed sensor optimizes against bad data, so fault diagnosis has to come first.

How is AI control different from Guideline 36 sequences?

Guideline 36 sequences apply fixed reset logic, such as trim and respond, inside BAS controllers. AI supervisory control computes setpoints from models and forecasts, and can weigh trade-offs across air handlers and the chiller plant that fixed request logic only approximates. Well-tuned Guideline 36 sequences also give AI control sound local logic to fall back on when the supervisory software stops.

Which BACnet priority should an AI platform write at?

No standard assigns one. The BACnet standard assigns priorities 1 and 2 to life-safety commands and priority 8 to manual operator commands, so a platform that writes at priority 8 or lower leaves safety commands in charge. Owners should confirm the priority in writing and test how a manual override interacts with the platform’s commands.

How long should a trial of AI HVAC control run?

A trial should run long enough to cover the weather and occupancy conditions that drive savings. The 2025 Applied Energy review excluded tests shorter than 7 days and found that short tests showed higher and more variable savings. Whole-facility verification under IPMVP Option C compares energy use against a regression baseline, typically built from a full year of data.