Designing an AI assistant people actually trust

The first version of our assistant replied to every receipt with a cheerful 'Done!' — even when it had guessed half the fields. Users loved it for a week. Then someone exported a report, found three invented totals, and trust evaporated overnight.

We rebuilt the personality around one principle: the assistant must be visibly honest about its own uncertainty. When a field is unclear, it says so — 'I read HK$248.50 but the total is faded, is that right?' — instead of silently filing a guess.

The counterintuitive result: admissions of uncertainty made users trust the confident answers more. A system that sometimes says 'I'm not sure' is believable when it says 'Done.' A system that never doubts itself teaches users to doubt everything it says.

We also deleted the personality theater. No fake typing delays, no 'Great question!', no emoji cascades. The assistant answers like a competent colleague: briefly, precisely, and only when it has something to say.

The metric we watch is not engagement — it is the correction rate trending down while the check-the-details rate trends down too. People verify less over time, and the data says they are right to.