I have a slightly greedy expectation of personal AI: it should do things for me, and know which things not to bother me with today.
The first half makes a good demo. An email goes out. A hotel gets booked. A calendar rearranges itself. The second half is harder to show. The assistant notices that I already have enough going on and decides not to add another suggestion. Nothing appears on the screen. That may be the useful outcome.
The question I keep coming back to is whether an increasingly capable assistant makes my life easier—or gives me another colleague to manage.
The assistant no longer starts from scratch
Meta describes Muse as a personal agent that can work across applications, remember context, and suggest actions. Today describes its product around persistent memory and proactive help throughout the day. These are their published descriptions, not a comparative test I have run.
This is a meaningful direction. Nobody wants to reintroduce themselves to an assistant every morning. But it also makes “we have memory” a less complete answer for a small company. What does that memory let a person stop doing?
More context creates more plausible interventions. A saved restaurant becomes a reservation suggestion. An unread book becomes a recommendation. An unanswered message becomes a reminder. Each can be reasonable on its own. Together, they can turn a phone into a very enthusiastic project manager.
As execution gets cheaper, deciding what is worth doing matters more.
Two weeks without calling Dad
Consider a hypothetical assistant that notices you have not called your father for two weeks.
Perhaps you forgot. A reminder could help. Perhaps you want to call but do not know what to say. Another reminder adds very little. Perhaps you want some distance right now. Persistent encouragement would make things worse.
Now imagine one more piece of context: yesterday, while cooking, you noticed that you chop vegetables in exactly the same order as he does.
For someone who wants to reconnect but lacks an opening, that might be enough: “I just realized I cook the same way you do.” It offers a way into a conversation. “You have not contacted your father for fourteen days” sounds more like a late-payment notice for a family relationship.
The distinction is not tone. It is an understanding of what is getting in the way. The same logic applies to an unsent proposal: the person may lack information, remain undecided, or have decided not to send it. A task list can flatten all three into “overdue.”
A starting point for a small team
At Spiro, I am exploring everyday records and sharing updates with family. Sometimes the person already wants to share. Choosing a moment, reconstructing the background, and explaining it feel like too much work.
That apparently small problem contains concrete decisions. Which moment belongs in this message? What does the recipient already know? Did the sender previously ask us not to turn ordinary life into a grand reflection?
I would rather begin with something people already want to do and repeatedly struggle to complete. If we can reduce the total effort, there is a reason to return. What else the system might eventually do should follow from that experience.
Count the work around the demo
When I watch a personal-agent demonstration, I want to look before and after the satisfying green checkmark.
How long did the person spend explaining the task? How much did they delete afterward? Did the recipient need to untangle an accidental promise? After a month, do the same instructions still need repeating?
An assistant can finish five tasks while creating three new ones for its owner. Those costs do not necessarily appear in the completion metric. They will appear in the decision to open the product again.
The things an AI can do will keep expanding. A person's day still has twenty-four hours.