A $6,880 AI phone shows why the boring details matter

TechCrunch asked the Alphafold's assistant to handle a realistic leaving-for-the-airport sequence: tell a contact the reviewer was running late, navigate to the airport, turn on Do Not Disturb and set a reminder to call the hotel in 15 minutes. The assistant sent the message, changed the phone setting and opened Google Maps. It did not start navigation, and the reminder landed at 9:08 p.m. even though the request was made at 2:32 a.m.

A second task asked for a Mumbai-to-Pune business trip with a morning flight, a hotel recommendation and calendar entries for July 18–19. The assistant correctly found that there was no suitable direct morning flight and offered a human-concierge handoff, but it created the calendar event for July 7.

The document test cut the other way. Hermes Agent initially analyzed a locally saved spreadsheet and summarized the second-quarter figures. Days later, in the same conversation, it no longer recognized the file and asked for another upload. Google's Gemini needed the initial upload too, but retained enough context to answer a later question.

The comparison exposes a basic design tradeoff. Vertu's assistant was more willing to act. Gemini asked more questions and completed less of the original request on its own, but produced the correct reminder after the reviewer chose an airport and reminder app.

Neither behavior is automatically better. An assistant that asks about every harmless choice becomes another form to fill out. One that guesses through dates, destinations and recipients can move quickly in the wrong direction. The interface has to know when a missing detail changes the consequence.

On a phone, that means the final check should be short enough to read while putting on a coat: airport, arrival time, reminder time, calendar dates, person being messaged. Five concrete fields beat a paragraph saying the task is complete.

Vertu says Hermes Agent remembers context, works across meetings, messages, documents and travel, and keeps significant actions under user approval. Those are sensible promises. The harder product job is deciding what the approval actually shows.

A generic confirmation such as “ready to proceed” does not help much. The user needs to see the parts the assistant inferred: which airport, which date, which calendar, who will receive the message and whether navigation will merely open or actually start. The inferred fields should look different from details copied directly from the request.

The same rule applies after the action. Do not report four green checks if one step only opened an app and another used a guessed date. Show sent, changed, opened but not started, and needs correction as different states. A busy person should not have to reconstruct the run from five apps.

Priya Rao would count correct finished tasks, not taps performed. For ten ordinary phone jobs, track how often the user has to correct a date, reopen an app, repeat a file upload or check that an action really happened. A fast run that creates five minutes of checking did not save five minutes.

Cass Bell would watch the luxury framing. Nearly seven thousand dollars can make ordinary software roughness sound like an exclusive beta experience. The useful question stays stubbornly plain: did the assistant remove a chore, or did it make the owner supervise a more expensive phone?

Those views pull in different directions but land on the same buying test. Ignore the longest feature list. Pick five tasks from a real Tuesday and look at the after-state. The calendar date, recipient, route and document should still be right when you return later.

Ask the assistant to create a reminder using a relative time, such as 15 minutes from now. Check the clock. Then ask it to add a two-day event next week and read the dates before saving.

Give it a harmless local document, ask one question, close the app and return the next day. See whether the assistant remembers the file, clearly says it no longer has access or quietly answers from stale context.

Finally, ask for a three-step task across apps. Note which steps were completed, which only opened the right screen and which need your approval. If the result collapses all three into one cheerful success message, the assistant is still making you do the audit.

AI assistants on phones are becoming capable enough to touch the day, not just talk about it. That raises the bar. The product does not need to be timid. It needs to make the fragile details obvious before confidence outruns accuracy.