Start with the problem, not the product
Every launch starts with a product. I want to start with a number that made us angry.
The average knowledge worker spends 4.1 hours a week looking for documents they themselves wrote. Not reading them. Looking for them. Four point one hours, per person, per week, and the number has not moved in nine years.
We have made the writing faster and the finding slower, and then we congratulated ourselves on the writing.
The constraint we set ourselves
Two years ago we wrote one constraint on a whiteboard and did not erase it: no new interface.
Not "a better interface". No new interface. Because the reason search fails is not that the search box is bad. It is that you have to leave what you are doing to go and ask a question. Every retrieval system we have built for forty years requires you to stop, formulate a query, and go somewhere else.
So the constraint was: the thing you need has to arrive inside the work, without being asked, and it has to be right often enough that you trust it. Not always right — trusted enough.
What we built
Three pieces.
A live model of what you are working on, held entirely on your device. Not a copy of your documents — a model of your intent, which is a much smaller and much more private thing.
A retrieval layer that reads that model and pulls from everything you have ever touched, ranked by what you are doing right now rather than what you typed last.
And a confidence gate. This is the part I am proudest of. The system does not show you anything unless it is at least 80% sure the suggestion is useful. Below that threshold, it stays silent. A system that is occasionally brilliant and often noisy is worse than no system, because it trains you to stop looking.
The proof
We have had this running internally for eleven months. The 4.1 hours is now 1.3. Ninety-one percent of people who try it for three weeks are still using it at week twelve, which is the only retention number that has ever mattered to me.
And the number I did not expect: people report finding documents they had forgotten writing, and 22% of what the system surfaces is something the person did not know they had. That is the part that made me think we are onto something bigger than search.
The promise
Available today, on the four platforms you asked about most, with the on-device model as the default and the cloud version opt-in and off by default.
No new interface. Four hours back in your week. And a system with the good manners to stay quiet when it is not sure.