Walk through any electronics shop and you will find kettles, doorbells and light switches described as intelligent. Very few of them are. The word has been stretched to cover anything with a companion app, which makes it almost useless as a buying signal.
That is a shame, because underneath the marketing there is a real shift happening. A handful of things in a modern home genuinely benefit from a model that learns, and they are not the things the packaging tends to shout about. This guide separates the two.
The difference between automation and learning
A rule is a sentence you write once: if the hallway sensor sees motion after sunset, turn on the lamp at 40 per cent. It runs identically forever. It is predictable, easy to debug, and costs nothing to run.
A learned behaviour is different. The system watches what happens over weeks, builds a statistical picture of your household, and makes a guess. It might notice that the lamp gets switched off manually every Tuesday because that is when you work late, and stop turning it on.
Both are useful. The important thing is that rules are better than learning for most household jobs, because household jobs are small, stable and safety-adjacent. You do not want a model getting creative with your front door lock.
Learning is worth the complexity when a decision has many inputs, changes gradually over time, and a wrong guess is cheap. Heating, energy scheduling and camera alerting all fit. Locks, alarms and anything involving water do not.
Four places machine learning genuinely helps
1. Heating and cooling
This is the clearest win, and it is the oldest. A thermostat that learns has three things a timer does not: a model of how quickly your specific building loses heat, a picture of when the house is actually occupied, and access to the local weather forecast.
Put those together and the device can start heating before you ask, at the cheapest moment, and stop early because it knows the rooms coast for twenty minutes after the boiler shuts off. Trials of learning thermostats have repeatedly reported single-digit to low-double-digit percentage savings on heating energy, with the biggest gains in poorly insulated homes and the smallest in flats that barely need heating at all.
The honest caveat: savings depend far more on your starting behaviour than on the device. If you already run a tight schedule, a learning thermostat mostly buys you convenience.
2. Camera alerts that are worth reading
The original problem with home cameras was never recording. It was that every passing cat generated a notification, so within a fortnight you stopped looking at them.
On-device classification fixed this. A small model running on the camera itself, or on a nearby hub, decides whether a moving blob is a person, a vehicle, an animal or a wind-blown branch, and only bothers you for the categories you care about. Because the classification happens locally, the video does not have to leave your house for that step, which is both faster and considerably better for privacy.
This is the feature to look for when you buy: on-device detection, with cloud upload optional rather than mandatory.
3. Energy shifting
If your electricity tariff varies through the day, there is real money in moving flexible loads to cheap hours. Dishwashers, tumble dryers, hot water tanks, home batteries and car chargers are all flexible. Ovens and kettles are not.
A scheduler that knows the next day's prices, your car's charge level, your typical departure time and the solar forecast can plan a night that a human would not bother to plan. This is closer to operations research than to anything glamorous, but it is one of the few smart-home features with a payback period you can actually calculate.
4. Voice that survives normal speech
Early voice assistants required you to speak in a stilted command grammar. Modern speech models handle interruptions, accents, background noise, and sentences where you change your mind halfway through. That is not a gimmick. It is the difference between a whole family using the system and one person in the household using it.
Where it is still theatre
Some categories have adopted the vocabulary without the substance.
- Fridges that suggest recipes. The camera sees packaging, not contents. It cannot tell you how much milk is left.
- Washing machines that optimise cycles. Load sensing is real, and it predates machine learning by decades. It is a weight sensor and a timer.
- Lighting that reads your mood. It is reading the clock.
- Anything with a screen bolted on. A screen is not intelligence, and it is one more component to be unsupported in four years.
None of these are scams exactly. They are ordinary features wearing an ambitious label.
What a sensible setup looks like
You do not need many devices. You need a few that cooperate.
| Layer | What it does | Buy carefully because |
|---|---|---|
| Hub | Runs automations locally, speaks to everything else | This is the piece that decides whether the house still works when the internet is down |
| Sensors | Motion, contact, temperature, humidity, power draw | Cheap and boring; more sensors beats smarter devices almost every time |
| Actuators | Lights, plugs, valves, thermostat | Prefer devices with a physical control that still works if the app dies |
| Voice | Optional front end | Never make it the only way to control something |
The single most important compatibility question is whether your devices speak a common local standard rather than each phoning its own manufacturer's cloud. A house built on one vendor's cloud is a house that breaks when that vendor changes its business model.
The three risks worth taking seriously
Devices outliving their support. A light bulb has a ten-year life and a two-year software support window. Anything that requires a live cloud account to function has an expiry date. Local control is not a nice-to-have; it decides whether you are throwing the thing away in 2029.
Microphones and cameras as a category. Every always-listening device is a judgement call about who you trust. That is a personal decision, but make it consciously, room by room, rather than accumulating them by accident.
The network itself. A cheap connected socket is a small computer with an unpatched web server sitting on your home network. Put those devices on a separate network from your laptop and phone. We go through the details in our guide to IoT security risks at home.
Starting sensibly
If you are beginning from nothing, resist the urge to buy a kit.
- Pick one annoyance. The hallway that is always dark. The heating that runs when nobody is home. Solve that.
- Buy the boring standard. Choose devices that work with a widely supported local protocol, not the one with the best advert.
- Add sensors before you add intelligence. A simple rule with good inputs beats a clever model with bad ones.
- Insist on manual fallback. If the switch on the wall no longer works, you have made your home worse.
- Only then automate. Live with manual control for a fortnight, notice what you actually repeat, and automate that.
Where this is heading
The interesting near-term development is not more devices. It is fewer, better-integrated ones, with the model running in the house instead of in a data centre. Local inference is now cheap enough to run on a hub the size of a paperback, which removes the latency, the monthly fee and most of the privacy objection at once.
The version of the smart home worth wanting is not a house that talks. It is a house you stop thinking about, because ten small irritations you used to manage by hand have quietly gone away, and nothing important depends on a server you do not own.