Delegate the inbox by defining four categories, reply, draft, delegate and archive, then have your assistant run one morning and one afternoon pass so you only see the messages that genuinely need you.
Handing over an inbox fails for predictable reasons. There are no rules, so everything gets escalated. Or the rules are too rigid, so something important sits unanswered. The fix is a small set of categories and a rhythm you both trust.
Anything that does not fit goes into a short escalation list reviewed once a day. That list should shrink every week.
A morning pass before your first meeting and an afternoon pass before you finish. Between passes, the inbox stays closed to you. What arrives is either handled or waiting in a single, ordered queue with drafted replies ready.
Style guides rarely work. Twenty of your own sent messages do. Pick examples across contexts, a warm reply to a client, a firm decline, a short internal instruction, and let your assistant work from those. Correct drafts in the first fortnight and the accuracy compounds quickly.
Classification, summarising long threads and preparing first drafts are safe and save real time. Sending anything sensitive, commercial or personal without a human read is not. The rule worth holding is simple: AI prepares, a person decides.
Spend ten minutes each Friday on the escalations. Most will turn into a new rule. After about six weeks, an assistant running this properly handles the large majority of the inbox without you seeing it at all.
Triage fails when it depends on judgement that only exists in your head. The fix is a small, boring set of categories that someone else can apply consistently, with a clear action attached to each one.
Four buckets is enough for almost every inbox. More than that and the system becomes a decision in itself.
Continuous triage feels responsive and quietly destroys focus for both of you. Two fixed passes, usually mid-morning and late afternoon, cover almost every realistic response expectation while leaving long blocks of uninterrupted work.
The exception list should be written down and short: named clients, named investors, anything from your own team flagged urgent. Everything else waits for the next pass.
Classification and first-draft writing are genuinely good uses of AI here. A model can sort by intent, extract the request from a long thread, and produce a reply in your register once it has seen enough of your previous messages.
What it should not do is send. The value of triage comes from you trusting the queue, and one badly judged auto-send costs more trust than a month of saved minutes recovers.
A triage system built for yourself and a triage system someone else can run are different things. The first can rely on instinct. The second needs the instinct written down, because every ambiguous message is otherwise a question back to you, which defeats the purpose entirely.
The handover takes about two weeks. In the first, the assistant shadows: they categorise everything and you review the categorisation daily, correcting what they got wrong and explaining why in one line. Those corrections are the actual training data, and they should be captured somewhere durable rather than left in a chat thread.
In the second week the flow reverses. They triage and act, you review the exceptions only. By the end of it, the rules that felt uncodifiable have usually turned into a page of specific guidance: this client always gets a same-day reply, anything from this supplier goes to finance, anything mentioning the board comes straight to you regardless of the hour.
Keep that page current. Inboxes change as the business does, and a triage system that is not revisited quarterly slowly drifts back into a queue that only you can interpret.
Should my assistant have full access to my inbox? Usually yes, with delegated access rather than shared credentials, and with clear rules about what can be sent without approval.
How long before inbox delegation feels safe? Most leaders are comfortable within four to six weeks, provided drafts are reviewed and corrected consistently in the first fortnight.
Does AI send the replies? No. AI helps classify, summarise and draft. A person reviews and decides what goes out.
More insights | Alvara Partners