When you ship an AI assistant for people who didn't grow up with one, you don't have the benefit of decades of normalized behavior. Every assumption you'd make about a "user" has to be examined. These are the five principles we keep coming back to.
1. Speak like a neighbor, not a help desk
Most assistants default to a formal, corporate register. It signals safety to the company shipping it — but it signals coldness to the person on the other end. Mimi's voice is closer to a thoughtful neighbor: warm, direct, occasionally funny, never patronizing.
2. Show your work
When Mimi flags something as suspicious, it never just says "this looks like a scam." It says why: the URL doesn't match, the sender has no history, the urgency language is a red flag. People build long-term skill from explanations, not warnings.
If your assistant can't tell you why, your user is just trading one black box for another.
3. Confirm before you act
For anything that touches money, identity, or family, Mimi asks one more time. Yes, it adds friction. Yes, that friction is the point. We'd rather slow down a routine action than miss the one time it matters.
4. Be quiet on purpose
Mimi is on by default but verbal only on request. The most common feedback in our first cohort was relief that the assistant wasn't constantly chiming in. Presence is not the same as noise.
5. Hand off, don't hold on
When something is beyond what AI should do — emotional distress, complex disputes, anything requiring a human signature — Mimi connects to a real person. The assistant's job is to know its limits and act on them gracefully.
What ties them together
All five principles share a single belief: an AI assistant for non-experts isn't a downgrade of the "real" assistant. It's a different design discipline entirely, with its own tradeoffs and its own ethics.
We don't always get it right. But we publish what we learn, in the hope that the rest of the industry treats this audience with the seriousness it deserves.