Every smart home product now claims AI. Most of the time the claim covers a schedule with a variable in it. It is worth separating what genuinely changes when you put learning into a house from what is a rebranded timer.
The real distinction
A rule fires when a condition is met. Someone had to think of the condition, write it down, and keep it current as the household changes.
A learned model infers the condition from what has actually happened. Nobody writes it down, and it updates itself when the family's routine shifts.
That difference sounds academic until you count the rules in a real installation. A well-automated house accumulates hundreds. Every one is a small maintenance debt, and the reason so many smart homes quietly degrade is that nobody pays it. The family's routine changes, the rules do not, and eventually somebody turns the automation off because it keeps getting things wrong.
Where it genuinely earns its place
Occupancy prediction. Motion sensors tell you a room is occupied. A model trained on months of sensor history tells you a room is about to be occupied, and — more usefully — that it will be empty for the next four hours. Heating a slab takes a long time; knowing in advance is worth far more than knowing now.
Thermal modelling. Every room has a different thermal response, and it changes with outside temperature, wind and solar gain. A learned model of how each zone actually behaves beats any hand-tuned schedule, particularly with underfloor heating where the lag is measured in hours. This is where the energy savings genuinely are.
Anomaly detection. This is the one clients end up valuing most, and it is the least exciting to demonstrate. A model of normal power draw notices a freezer that has started running continuously. Normal water flow notices a leak overnight. Normal network behaviour notices a camera that has begun talking to somewhere it never talked to before. None of this needs to be clever; it needs to be watching constantly, which is precisely what people are not.
Useful camera events. The difference between a camera that fires forty notifications a night and one that fires two is on-device classification. Person, vehicle, parcel, animal. This is genuinely mature, runs locally on modern hardware, and is the single change most likely to make a client stop muting their security system.
Where it does not
Lighting scenes. People want the same light in the same room at the same time of day. This is a schedule and an astronomical clock. Learning adds nothing and introduces the possibility of getting it wrong, which is far worse than being predictable.
Anything safety-related. Locks, alarms, gates. A probabilistic system should never be the thing deciding whether a door opens. Inference can inform; it should not actuate.
Voice, mostly. Better language models make voice control more forgiving of phrasing, which is a real improvement. They do not solve the actual problems with voice in a house — that it is slow compared with a keypad, that it fails in a room with music playing, and that nobody wants to talk to the ceiling when guests are over.
Where it runs matters more than what it is
The important architectural decision is not which model. It is whether inference happens in the house or in a data centre.
Local inference means the house keeps working when the line drops, responses are immediate rather than round-tripped, occupancy data — a genuinely sensitive record of when a family is home — never leaves the building, and there is no subscription attached to your own lighting.
Modern hardware makes this straightforward. Camera classification, occupancy prediction, thermal modelling and speech recognition all run comfortably on a small accelerator in the rack. There is no technical reason for any of it to be somebody else's server, and several good reasons for it not to be.
The test worth applying
Ask what the feature does when the internet is down. If the answer is "nothing", the intelligence is not in your house. That may be an acceptable trade for a given feature — but it should be a decision, taken knowingly, and not a surprise discovered during an outage.