The future of AI personal assistants

The idea

Most predictions about assistants are really predictions about model capability, which is the least uncertain variable we have. Models get better, cost per token falls, and running something capable on a phone stops being remarkable.

The open questions are structural. Where does personal context live, who is allowed to reach it, and what is an assistant for once it is no longer novel.

From prompted to grounded

The first generation was prompted. You supply the context in the conversation, the model reasons over it, the context is thrown away. Fine for self-contained tasks, poor for life, because life is not self-contained.

The second is grounded. The assistant has standing access to some slice of your real material. The difference is not subtle: a prompted assistant answers, a grounded one can raise something you had not asked about. That is the behaviour people wanted all along, and it is a much harder thing to build well.

The shift from answering to noticing is the main thing to watch, because it moves who starts the conversation. Once the assistant initiates, being smart is not enough. It has to be right about you and it has to know when to stay quiet. We feel this weekly: every proactive thing we ship, from daily recaps to charts that open on their own, sits somewhere on a line between useful and presumptuous, and the line is not in the same place for everyone.

The platform question is not settled

Three plausible endings, no consensus.

The operating system ending: your assistant is whichever one ships with your phone, because it is the only software allowed to see everything. Convenient, and it makes your context a feature of a hardware purchase.

The super-app ending: one AI company accumulates enough of your material that leaving becomes unthinkable. Fast, capable, and it recreates the platform dynamics of the last decade with far more intimate data.

The personal data ending: your context is something you hold, and models and applications connect to it with your permission. Slower to get good, harder to build, and the only version where switching stays cheap and you stay the customer.

We are building for the third. It is a bet, not a forecast, and worth saying plainly.

Four things that will change

• Ambient over conversational. Less typing into a box, more surfacing at the right moment. The chat window will look like an early form factor, the way the address bar does now.
• Context becomes portable. There is real pressure, regulatory and commercial, toward personal data that can move. Whoever makes that genuinely work changes the shape of the category.
• Small models get personal. On-device and regionally hosted models handle the everyday work, frontier models get the hard cases. The privacy argument and the cost argument point the same way, which usually means it happens.
• Evaluation moves from benchmarks to trust. When the assistant proposes rather than responds, the question stops being is it clever and becomes is it right about me.

What probably will not happen

Assistants will not replace deciding, because the bottleneck in most personal decisions was never information. Full autonomy over consequential actions will stay rare, because the tolerance for error in someone's actual life is far below what a demo needs. And one assistant will not do everything, for the same reason one app never did.

Continued

The version worth wanting is quieter than the marketing. Not an agent running your life. Something that holds the context so you do not have to, says when something needs you, and stays out of the way when it does not. Getting there is less about the model than about who is trusted with the material, and that is being decided now.