LeaThe trouble with an AI assistant that never clocks off
An assistant working while you sleep sounds wonderful. But when it can send, delete and decide, convenience starts to look a lot like handing over the keys.
Picture waking up to a shorter to-do list. The dinner is booked, the school email is answered and your calendar finally makes sense. Now picture one more notification: the assistant canceled something you meant to keep. You were asleep when it decided that was helpful.
The appeal is obvious
Personal assistants such as Muse, Instinct and OpenAI’s dots are built around helping with ongoing tasks and connected services. X also describes Grok Bot working through tasks in selected tools. Features and safeguards differ, but the direction is clear: help is moving beyond a reply in a chat window.
That is appealing because real life does not arrive as neat questions. It arrives as a broken boiler, a missed appointment and six people trying to agree on dinner. Someone who can keep working on the loose ends would be useful.
These services already have safeguards. Muse describes approvals for actions such as purchases and emails. Dots has built-in action review, and its proactive research uses read-only tools. Check the controls for the particular task you authorize rather than assuming background work always means unrestricted action.
“Sort it out” contains a surprising number of decisions
Take a simple request: arrange a weekend away. To finish the job, an assistant might need to choose dates, read your calendar, contact someone and make a booking. Each step asks a different question about what you are comfortable letting it decide.
Finding available dates is one thing. Explaining a private appointment to a friend is another. Canceling your existing plans or spending a nonrefundable amount goes further still.
A human assistant would eventually ask, “Are you sure?” With AI, you need to know which moments trigger that question before the task starts.
The risk grows when small steps add up
One wrong calendar entry is annoying. A bot that uses that entry to cancel a booking, send an explanation and rearrange other plans can turn the first mistake into an afternoon of repairs.
This is why reviewing the final summary is not always enough. By then, messages may have reached other people. A private detail cannot reliably be made private again by deleting your copy.
The OpenClaw inbox incident is a real example of a person trying to stop an assistant that had already started unwanted changes. It does not prove all autonomous assistants are unsafe. It does show why “I will check later” has limits.
Decide what can happen while you are unavailable
Give routine jobs some room. Ask an assistant to collect options, organize information or prepare drafts. Keep a shorter list of decisions that should wait for you.
- Messages that disclose personal information or make commitments.
- Permanent deletions and broad changes to accounts or files.
- Payments, new recipients and bookings with meaningful cancellation costs.
Build the boundaries before you build the habit
Use the assistant’s own approvals and account controls. For supported actions routed through Leash, add checks for private details, outgoing messages and certain destructive changes. Leash is an antivirus for AI activity, not a promise that every assistant action everywhere will be caught.
Start with one useful background task and look at what actually happened in the connected account. If the result earns more trust, expand carefully.
The best version of this future is not spending all day supervising a bot. It is knowing which decisions can happen while you are out living your life, and which will still be waiting when you come back.