Ask anyone who has worked with a good executive assistant what they actually did, and you rarely hear "answered my questions". You hear things like: they knew which fire to put out first, they chased the thing I forgot, they said no on my behalf.
Four things a real EA does that most tools do not
- They hold context you never gave them explicitly — who matters, what is political, what you always forget.
- They own outcomes, not turns. The meeting gets booked, not "here are three times you could propose".
- They follow up without being asked, days later, on the thing you both knew was pending.
- They protect your attention, filtering aggressively rather than forwarding everything.
Why chatbots cannot do those
A chatbot is a turn-taking machine. It has no memory of Tuesday, no visibility of the conversation where the promise was made, and no way to act three days later. Every one of the four behaviours above requires persistence, context and the ability to act unprompted — none of which a chat window gives you.
What has to be true instead
- Context by default: the assistant sits where the conversations already happen.
- Durable state: commitments, tasks, and multi-step work survive restarts.
- Scheduled autonomy: it can act at 7am on Friday without you opening anything.
- Written memory you can read and edit — not an opaque embedding store.
- Hard guardrails, so unprompted action does not become unwelcome action.
The uncomfortable trade
All four behaviours require access, and access is exactly what people are right to be cautious about. The answer is not less access — an assistant without context is a search box. The answer is a permission model you can see: read and write as separate decisions, blocked chats that stay blocked in every mode, draft-first sending, and caps on the irreversible actions that no setting can unlock.
A good assistant is not the one that can do the most. It is the one whose limits you can state from memory.