HVAC concepts, in plain English.

The equipment layer is where the physics happen. Four categories cover almost everything you'll encounter: DX (direct expansion) systems — split AC and VRF/VRV — where refrigerant moves heat directly; hydronic systems — chillers and boilers — where chilled or heated water carries heat to air handlers and fan coils; heat pumps, which do both heating and cooling with one refrigerant circuit; and ventilation systems — AHUs, ERVs, and VAV boxes — that manage fresh air and pressurization. Most commercial buildings use two or more categories in combination.

Controls sit at two levels. At the equipment level you have the unit controller (the board in a Daikin outdoor unit, the unitary controller on a Carrier RTU), a thermostat or zone sensor, and sometimes a protocol gateway that translates the unit's proprietary bus to BACnet or Modbus. At the building level sits the BMS or building automation system (BAS): software and hardware that reads setpoints, schedules, and alarms from every piece of equipment and lets a facilities manager see it all in one place. Common BMS platforms include Siemens Desigo, Johnson Controls Metasys, Schneider Electric EcoStruxure, and Trane Tracer.

AI sits above the BMS — it is not a replacement for it. An AI layer reads the telemetry the BMS already exposes (or, where no BMS exists, reads directly from brand cloud APIs), learns normal patterns, detects anomalies, and either surfaces recommendations or, with explicit write-back authorization, makes small adjustments through the existing control chain. The safety interlocks, setpoint limits, and override logic already baked into your BMS stay in place. AI adds a faster, smarter observer on top — it does not bypass what you already have.

The defining feature of a VRF system is its inverter-driven compressor. Traditional split systems run at a fixed speed — on or off. A VRF compressor modulates continuously, delivering exactly as much refrigerant as the load demands at any moment. Multiple indoor units (cassettes, wall units, ducted handlers) share a single outdoor unit over a two- or three-pipe refrigerant circuit. Two-pipe systems heat or cool all zones simultaneously; three-pipe heat-recovery systems let some zones heat while others cool — capturing waste heat from one side to serve the other, which is why three-pipe VRF is common in buildings with mixed orientation or mixed use.

VRF is the right choice when you need independent zone control without central ductwork, are retrofitting a building where adding ducts is structurally difficult, or need high part-load efficiency across zones with very different occupancy schedules. Typical commercial installations range from 2 tons (single outdoor) up to 60–80 tons for large multi-frame systems. Key manufacturers: Daikin VRV (the original, commercial flagship), Mitsubishi Electric City Multi (favored in North American commercial), Fujitsu Airstage (strong in light commercial), and LG Multi V (aggressive on price/efficiency at mid-capacity). All four publish open-protocol integration via BACnet or their own cloud APIs.

AI adds the most value on VRF at two points. First, per-zone setpoint optimization: the agent learns each zone's occupancy pattern, pre-conditions before arrival, and relaxes setpoints during unoccupied periods — compounding efficiency gains the building operator would never have time to tune manually. Second, compressor envelope monitoring: VRF compressors run continuously at variable speed and show fault signatures (discharge temperature drift, suction pressure anomaly, EXV hunting) days before a hard failure. An AI monitoring layer catches these early, routes a ticket to the technician, and often avoids an emergency call-out.

In a split system, the outdoor unit contains the compressor and condenser coil; the indoor unit contains the evaporator coil and the air handler. Refrigerant circulates between them through copper or aluminum line sets, typically 15–50 feet long. The outdoor unit rejects heat to the outside air; the indoor unit absorbs heat from the room. In cooling mode, refrigerant enters the indoor unit as a cold low-pressure liquid, absorbs room heat, evaporates, and returns to the outdoor unit as a warm gas to be compressed and condensed again. Reversing this cycle with a reversing valve gives you a heat pump.

Ducted split systems — sometimes called 'central air' — connect to a network of ducts that distribute conditioned air throughout the home. The indoor unit is a large air handler in a closet, attic, or basement. Ductless mini-splits skip the ductwork: each indoor unit mounts directly in the room it conditions, connects to the outdoor unit via a small hole for refrigerant lines and electrical, and conditions only that room. Mini-splits are common in renovations, additions, and any space where running ducts is impractical. Multi-head mini-splits connect two to five indoor units to a single outdoor unit, giving you zone control at residential scale.

AI helps split systems in three ways. Scheduling: the agent learns occupancy from connected sensors or calendar integrations and automatically pre-conditions before the space is needed, avoiding both over-conditioning and cold starts. Occupancy-based setbacks: during extended absences the agent relaxes setpoints to minimum safe levels — not the default coarse setbacks a homeowner sets and forgets, but fine-grained adjustments matched to actual absence duration and outdoor conditions. Fault prediction: compressor current draw, refrigerant pressure envelopes, and coil temperature differentials all carry early fault signals. Baseline learning over 14–30 days gives the agent enough context to flag drift weeks before a failure.

The chilled-water loop architecture has four main components. The chiller itself sits in the plant room (or outdoors for air-cooled chillers) and produces chilled water — typically 44–48°F supply, 54–58°F return. Primary pumps circulate that water through the chiller's evaporator barrel. Distribution pumps (in a primary-secondary or variable-primary arrangement) push the chilled water through the rest of the building. At the load side, air handlers (AHUs) and fan coils use chilled water coils to cool the air. Return water flows back to the chiller to be re-cooled. Cooling towers or dry coolers reject the heat to outdoors on the condenser side.

Chillers come in four main types. Centrifugal chillers use a spinning impeller — high-efficiency at full load, common in 200-ton-and-up ranges, vulnerable to surge at low load without inlet guide vanes. Screw chillers use helical rotors — more tolerant of part-load conditions, common in 100–500 ton range. Scroll chillers (used in smaller packaged units, under 100 tons) stack multiple scroll compressors for staging. Absorption chillers use heat (steam or hot water) instead of a compressor — found where waste heat or district heat is available, or where electrical demand is a constraint. Magnetic-bearing centrifugal chillers (Daikin Turbocor, Carrier AquaEdge) eliminate oil lubrication and dramatically improve part-load efficiency.

AI is highest-value on chiller plants because small improvements in plant efficiency translate to enormous energy savings at scale. Chilled water reset: instead of always maintaining 44°F supply water, the agent raises the setpoint as load decreases — every degree of rise saves roughly 1–2% compressor energy. Free-cooling enable: when outdoor wet-bulb temperature drops low enough, a waterside economizer can produce chilled water without running the compressor at all; the agent optimizes the crossover point in real time. Compressor fault detection: discharge pressure anomaly, oil temperature drift, and power factor deviation are early fault indicators. In data centers, the agent also monitors for hot/cold aisle drift — supply air temperature creeping up in a row — and adjusts CRAC/CRAH setpoints before IT equipment throttles.

The key component that distinguishes a heat pump from a plain air conditioner is the reversing valve. In cooling mode, the refrigerant cycle is identical to a conventional AC: outdoor coil rejects heat, indoor coil absorbs heat from the room. When the reversing valve switches, the cycle reverses: the outdoor coil absorbs heat from outside air (even cold air contains extractable heat), and the indoor coil delivers that heat to the room. This is why a heat pump can replace both a furnace and an air conditioner. The measure of efficiency is COP (coefficient of performance): a COP of 3 means 3 kWh of heat delivered per 1 kWh of electricity consumed. At moderate outdoor temperatures, modern heat pumps achieve COP 3–5. At very low temperatures, COP drops.

The cold-climate boundary is where heat pump conversations get nuanced. Standard air-source heat pumps lose capacity and efficiency as outdoor temperatures fall below about 35°F, and most have a minimum operating temperature around 0–5°F. Below that, a backup heat source (electric resistance strip heat, or a gas furnace in a dual-fuel system) takes over. Cold-climate heat pumps — Mitsubishi Hyper Heat, Carrier Infinity Greenspeed, Bosch IDS, and others — push the efficient operating range down to -13°F to -22°F using enhanced vapor injection (EVI) or other cycle modifications. Ground-source (geothermal) heat pumps bypass the outdoor-air problem entirely: ground temperature 6 feet down stays near 50–55°F year-round in most of North America, giving a stable heat source even in extreme winters.

AI targets two heat pump use cases. Defrost optimization: air-source heat pumps frost over their outdoor coil in cold, humid conditions and must periodically run a defrost cycle (reversing to cooling mode briefly to melt the ice). Default defrost logic runs on time-and-temperature schedules that are often too frequent. An AI agent monitors coil temperature differential and outdoor wet-bulb to trigger defrost only when actually needed — reducing defrost events by 30–50% and recovering that heat capacity. Dual-fuel switchover: in a dual-fuel system (heat pump + gas furnace), the balance point — the outdoor temperature below which gas is cheaper than electricity — shifts with utility rates and weather. The agent optimizes the switchover setpoint in real time against current electricity and gas prices, not a fixed balance point set at installation.

Read first, write later. The right way to deploy an AI layer is read-only for the first 7–30 days. During this period the agent builds a statistical baseline of every point it monitors — supply air temperatures, setpoints, runtime patterns, alarm frequencies, compressor envelopes. This baseline is what makes later write-back safe: the agent knows what normal looks like, so it can detect when its own actions produce unexpected outcomes and self-limit. Skipping the baseline phase and going straight to write-back is the most common mistake in HVAC automation deployments. It removes the agent's ability to distinguish its own effects from underlying system behavior.

Safety interlocks always win. A well-designed AI system never has a code path that bypasses a freeze stat, a fire damper position, a smoke control mode, or a high-limit cutout. These interlocks exist because the consequences of bypassing them — frozen pipes, smoke inhalation, equipment destruction — are severe and fast. Technically, this means the AI agent's setpoint writes should always be bounded by the hardware safety limits already present in the controller, and the agent should treat a safety interlock activation as an alert requiring human review — not an obstacle to route around. If your automation vendor cannot clearly explain how their agent handles a smoke control override, that is a red flag.

Audit everything. Every write-back action should produce an immutable log entry containing: the point that was written, the previous value, the new value, the timestamp, and the agent's stated reason for the action. This audit log serves three purposes: it lets you verify the agent is doing what you expect; it gives your technicians the context they need when they see an unexpected system state; and it provides the rollback path when you want to undo a sequence of changes. Alert suppression is a related discipline: when the agent makes a change it expects to trigger an alarm (e.g., a scheduled setpoint shift that crosses an alarm threshold), it should pre-suppress that specific alarm for the expected duration and log the suppression. Uncontrolled alert suppression — suppressing whole categories of alarms because they're noisy — masks real faults and should never be permitted.