What "AI in the home" actually means

“AI” now appears on the box of devices that contain a timer. That is not a complaint about marketing so much as a practical problem: if you are spending serious money on a house, you need a way to tell the difference between things that are actually different.

What follows is the ladder we use internally. Five levels, each defined by one specific capability the level below does not have. It is deliberately unkind about where things sit, including our own.

---

Level 0 — Manual

You operate the house. A switch is a switch.

Not a failure state. A well-designed manual house with excellent switches in the right places is more pleasant than most automated ones, and every level above this must preserve it. If your automation stops working and the house becomes unusable, the automation was built wrong.

Level 1 — Scheduled

The house acts on the clock. Heating comes on at six, garden lights at dusk, irrigation on Tuesdays.

What’s new: action without a person present.

The ceiling: a schedule is a guess about the future, made once. It is right about the average day and wrong about every particular one. It does not know that you are away, that it rained, or that the room is already warm. Most disappointment with “smart” heating is a Level 1 system being blamed for Level 3 expectations.

Level 2 — Scened and triggered

The house responds to events and stored configurations. Motion triggers light. A button recalls eight dimmers, three blinds and an amplifier input.

What’s new: response to the world, not just the clock.

The ceiling — and this is the important one, because almost every luxury home is here: every response is a stored photograph, and every rule is an independent writer. There is no memory, so nothing improves. There is no model of the building, so behaviour is hand-wired per address. There is no arbitration, so conflicts resolve by whichever instruction arrived last.

Where this sits in the market: Control4, Savant, Crestron Home, and KNX with logic modules. All of them. A flagship Savant installation and a supermarket motion sensor are on the same rung of this ladder — the Savant is enormously better executed, more reliable, and better looking, and it is doing the same category of thing.

That is not a knock on those companies. Level 2 done superbly is worth a great deal. But it explains why spending more inside the category does not produce a qualitative change, and why owners who have spent a fortune still describe their house as dumb. They are not confused. They correctly perceive that nothing on offer was a different kind of thing.

Level 3 — Assisted

A person’s intent is interpreted rather than encoded. Natural language, context across a conversation, compound requests.

What’s new: you no longer have to know the system’s vocabulary. Make it cosy in here resolves to actions without anyone having pre-defined “cosy”.

Where this sits: Josh.ai is the best example, and it is genuinely good. The current wave of LLM assistants — including Home Assistant’s own — is here too. Homey and similar bridges sit between 2 and 3: Level 2 mechanics with a Level 3 interface bolted on.

The ceiling, and it is a hard one: assistance is still instruction. The house does nothing until asked. A perfect Level 3 system in an empty room does nothing at all, forever, because its entire function is to convert requests into actions.

This is where the industry’s current excitement is concentrated, and it is worth being clear-eyed: putting a language model in front of a rules engine improves the front door and changes nothing behind it. The house has not become more aware. It has become more articulate.

A specific warning. An LLM that can call service functions directly — a pattern being shipped by several vendors right now — is a Level 3 interface with Level 4 authority and no Level 4 safeguards. It can act on your house, it has no precedence model, and its reasoning is not recorded in a form you can audit. That is the single most dangerous configuration currently available in this market, and it is being sold as the most advanced.

Level 4 — Autonomous

The house forms and acts on its own view, without being asked, within rules you set.

Three capabilities distinguish this from everything above, and all three are required — any two produce something worse than Level 2, not better.

A model of the building. Spaces, capabilities, occupants. Behaviour is expressed once about a kind of space and applies wherever it fits, instead of being hand-wired per address. This is what makes the system portable between houses and maintainable within one.

A durable memory. A record of what happened, what was decided, and what the household did about it. This is what makes usually computable — and usually is the whole content of anticipation.

An arbiter with explicit precedence. Every policy proposes rather than writes. Proposals are compared in one place, on stated tiers, and a lower tier structurally cannot beat a higher one. Safety above security above habitability above explicit human instruction above comfort above efficiency — enforced in the engine, not observed by convention.

What Level 4 is not: it is not the system being clever, and it is not the system being confident. Precedence tiers and append-only logs are constraints on cleverness. The autonomy is in acting unbidden; the safety is in never acting unaccountably.

Where this sits in the market: essentially nowhere, commercially. This is the gap we are building into, and we are not going to pretend it is crowded or that we have finished. As of today there is one instrumented house running this, it is ours, and the evidence page says so along with everything wrong with that as a basis for a claim.

Level 5 — the one we are not claiming

A house that revises its own goals. Not just how to keep you comfortable, but deciding what it should be optimising at all.

We include this to be explicit that we are not building it and do not think anyone should yet. A system that changes its own objectives in a building where people sleep is not a product, and the governance for it does not exist. If a vendor implies they are here, ask to see the audit log.

---

Using the ladder

Three questions will place any product on it, and they take a salesperson about thirty seconds to fail:

“Does it act without being asked?” No → Level 3 or below, whatever the demo showed. Every impressive demo is someone asking for something.

“What does it do when two automations disagree?” If the answer is anything like the last one wins, you set priorities in the rule order, or a pause while they work out what you mean → Level 2 mechanics.

“Can you show me why it did something last Tuesday?” If there is no decision log with the competing options and the reason, then nothing is accumulating, and nothing that does not accumulate can learn.

The honest summary

Most of what is sold as AI for the home is Level 2 with a Level 3 interface. That combination is genuinely useful, widely available, well supported, and far cheaper than what we make. For a large number of households it is the right purchase and we will tell you so during an assessment.

The reason to want Level 4 is narrow and specific: you are tired of operating your house, the thing you want is for it to stop asking, and you are willing to let it act on its own provided you can see exactly what it did and overrule it instantly.

If that is not your complaint, buy the cheaper thing.