What does an AI assistant in your text messages actually get used for?
Ten days in: the use cases that stuck were small, constant, and mostly admin, and the one that surprised us was two assistants scheduling a meeting between them.
After ten days, the AI assistant use cases that stick are small and constant: calendar changes by voice, a morning brief that flags conflicts, standing watches on cold deal threads, and routine setup work like DNS records. The deep work stays in coding tools. The assistant earns its place on admin.
The use cases that stuck were the small ones
This is the second pass on Instinct, an AI assistant that lives inside iMessage. Same disclosure as last time: no relationship with the company, no payment in either direction, and it is free during their beta.
Ten days in, the pattern is clear. None of the work that stuck was impressive on its own. A friend mentions a party on Saturday at seven, and an audio message adds it to the calendar without the phone ever being unlocked. A rejected company card gets chased down. A refund request gets filled out and filed, confirmed the next day. A new site needs DNS records, a domain, and a workspace, so it does the whole setup while the real work continues elsewhere. One email needed a line cut before sending, and that was a single message.
Each of those is a two to five minute task. The honest math is that there were roughly ten of them a day, and the cost was never the two minutes. It was picking up the phone and losing twenty to whatever else was on the screen.
Two assistants can schedule a meeting between them
The feature that was genuinely new is the trusted network. Two people each run their own assistant, and the assistants are allowed to talk to each other.
In practice it looked like this. A client came back with a time that worked for them. The reply was one line: check whether that works on the other calendar first, then confirm and send the email. The assistant checked with the other person's assistant, got a yes, and sent the confirmation. No further involvement.
That is the back and forth that scheduling actually is. Does Tuesday work, let me check, no, how about Thursday. Handing that to software is not a novelty feature for a household. It is the first time an AI assistant has behaved the way a real assistant behaves, which is to say it talked to someone else's assistant and came back with an answer instead of a question.
Standing watches beat one-off requests
The bigger shift was moving from asking for things to setting up watches that run without being asked.
A list of birthdays went in once, and now a text arrives on the morning of each one. Everyone thinks they could be better at that, and this removes the part that was actually hard, which was remembering.
On the business side, a daily pipeline watch now reads across email, calendar, and meeting transcripts and reports what needs attention. With several calls a day, threads go cold quietly, and commitments made on a call never make it to a list. A standing watch catches both. It also surfaces decisions rather than making them. One morning it flagged that a domain warmup had paused and asked whether to keep it running, which is exactly the right shape: the agent notices, the human decides.
Why the channel decides whether it gets used
The reason any of this got used is the delivery channel, not the model behind it.
Count the apps on your phone you have opened once in two years. Now count the text messages you have left unread. Almost nobody ignores a text. An assistant that lives in a separate app is one you have to decide to open, and that decision is where adoption dies.
This is the same finding as deploying agents inside companies. The ones that get real use run inside Slack or Microsoft Teams, where people already are. The capability is rarely the constraint. Everything here could be done in a coding tool or a chatbot, and most of the day is still spent in those. The assistant wins the small stuff because it is already open.
What we still would not hand it
Privacy is the real cost, and it deserves a straight answer rather than a shrug. The scoping is deliberate: read-only access to most email, no write access to business systems, no access to bank accounts. That mirrors how we set agents up for clients, which is read-only anywhere reading is enough, with the actions an agent can take on its own named explicitly.
It is still a genuine trade. An assistant that reads your calendar and your mail knows a lot about you, and each person has to decide whether the convenience is worth it. There is also a limit worth naming: the assistant lives in your messages but cannot read the rest of them. Plenty of commitments get made over text and then dropped, and that thread stays invisible to it.
The vendor question worth asking out loud
The last honest caveat is durability. This is a free beta with no revenue behind it, and that has to change eventually. It may not survive, and several competitors are building in the same space.
That is a reason to be careful about what you build on, not a reason to wait. Treat it as a category rather than a product. Messaging-native assistants are going to be normal, and the useful thing to work out now is which parts of your week are small, constant, and safe to hand over. That answer stays true no matter who ends up serving it.
The short version
- The work that sticks is small and constant: calendar edits, admin chores, setup tasks. Deep work stays where it already lives.
- Standing watches beat one-off requests. Set them up once and let them report.
- An agent that surfaces decisions instead of making them is the right default.
- Adoption is decided by the channel, not the model. Inside a company, that means Slack or Teams.
- Scope access deliberately: read-only where reading is enough, and no write access to money or business systems.
Questions this raises
What is a trusted network between AI assistants?
It lets two people's assistants talk to each other directly. When a meeting time comes in, your assistant can check the other person's calendar through their assistant and confirm without you relaying anything. It removes the back and forth that scheduling normally requires.
Does an AI assistant like this replace tools like Claude or ChatGPT?
No. Most of the day is still spent in coding and chat tools for real work. The assistant handles assistant-level tasks: bookings, forms, calendar changes, follow-up tracking. The value is that it is reachable when your laptop is not open, not that it is more capable.
What access should an AI assistant be given?
Start read-only. Reading email and calendars is enough for most of what it does. Withhold write access to business systems and bank accounts until you have watched it work for a while, and name explicitly which actions it may take without asking.