What an AI Assistant Actually Does in a Household

Every advertisement for these things shows a family being handed a week of dinners.

Eleven of our fourteen households tried exactly that. Three were still doing it after three months.

The failure is not that the meal plans are bad. They are fine. The failure is that a generated plan does not know there is a half-used bag of arborio in the cupboard and a chicken that has to be eaten on Tuesday, so it produces a shopping list that buys things the household already owns and ignores the thing that will otherwise be thrown away — which is the grocery problem with extra steps. The three households still doing it had all independently arrived at the same fix — type in what is already in the house first — which takes about four minutes and is the step nobody does, because the entire promise was that you would not have to.

What people actually used it for was more interesting and much less photogenic.

The three real uses

Drafting the message you are dreading. The largest single category. The email to the teacher about the thing that happened. The message to a landlord. The note to a sister-in-law about Christmas. In almost every logged instance the household did not send the output — they rewrote it, sometimes heavily — and several people said explicitly that the value was in no longer facing a blank box.

Decoding a document. The insurance denial. The school’s phone policy. A clause in a lease. Households pasted in official language and asked what it meant, and then — this is the important part — went back to the document itself to act. Used this way it is a translation layer over a text you already hold, which is a substantially different thing from asking it what to do.

A rough first version of something tedious. A packing list for a family of five going camping. A chore rota. A list of questions to ask at a school open evening. All immediately edited; all better than starting from nothing.

The common thread: the hard part of each of these is beginning, and a human checks the result within seconds.

The use nobody predicted

Six households used it to rehearse.

Not to write anything. To practise a conversation before having it: with a school about a child’s support plan, with an insurer about an appeal, in two cases with a parent or sibling about care. One mother described running the school conversation three times, playing herself, with the tool playing an administrator who kept saying no — and then going into the actual meeting with the third version of her own argument rather than the first.

She was not asking it what to say. She was using it as a wall to hit a ball against, and she came out with her own words, sharpened.

I found this the most defensible use in the whole log, and it is the one nobody is advertising.

Where it broke

Anything requiring knowledge of the household’s real life. Scheduling, consistently: it does not know that Thursday is impossible. Local recommendations: it does not know the neighbourhood, and two households were sent to a business that had closed.

And two households tried leaving something to run unattended — one a recurring shopping order, one a set of automatic replies. Both stopped within a fortnight. The recurring order had bought the same item three weeks running.

The rule that fell out of it

Across fourteen logs, the uses that survived three months have one property in common: the household supplies the facts and keeps the judgement.

Where the household supplied the facts — here is what is in my cupboard, here is the letter I received, here is what I want to say — it kept working. Where the household asked it to supply facts about their own life, or to make a decision on their behalf, it was abandoned, usually within a month, usually after producing something confidently wrong.

That is not a limitation anybody will put on a billboard, and it is a genuinely useful line to run a house on.