We built a warm, patient, multilingual AI companion for people who live alone, put it on a tablet, and watched almost nobody talk to it. Not "engagement was lower than hoped" — measured across the conversations on record, 88.8% contained zero messages from the senior. They opened the app. They looked at it. They put it down.
The instinct is to blame the prompt, or the voice, or the onboarding. The actual cause was in the response to a single endpoint: starting a conversation returned an opening message of `null`, and the app discarded the field anyway. So an eighty-year-old opened the app and was met with silence and a microphone button.
Silence is a question
A microphone with nothing above it is a question: what do you want? That is a fine question to ask someone who came with a task. It is a terrible one to ask someone who is lonely, because loneliness does not arrive with an agenda. It arrives as a quiet afternoon.
Every other product in this category waits. The lonely person is expected to initiate the conversation that is supposed to cure their loneliness. Written down like that it is obviously the wrong way round.
Variety cannot be a prompt instruction
So she speaks first. The harder problem is what she says. A model told to "vary your opening" converges on the same two or three sentences within a week. No test catches it. The person notices long before you do, and what they notice is that the thing does not really remember them.
The answer was to take the choice away from the model. There is a deck of moves — a callback to yesterday, a memory that matches today’s date, a photograph the family posted, a question she asked last time and never got an answer to — and the code picks which one is available for this person right now. The model only phrases it.
What it runs on was the surprise
The obvious fuel is the profile — interests, era, where they grew up. We measured it. Of the users on record, three had interests filled in. Three had an era. The person who accounted for 857 of 882 conversations had neither.
An opener resting on the profile degrades to "good morning, how are you?" for four users in five. What there was plenty of: 258 daily conversation summaries, every one of them distinct. So the openers are built from what she was actually told, which is also the only thing that could make them feel like being remembered.
One more rule fell out of it, and it is the one we would have got wrong: a thin summary is skipped rather than used. A callback to "you asked what time it was" proves she was not listening. Worse than no callback at all.