Using AI as an Executive Assistant: What to Hand Over and What to Keep
AI can genuinely do a slice of executive-assistant work — triage, drafting, meeting prep — but not the judgement calls about priority and access. A worked example and what to check first.
An executive assistant’s job is not one skill — it is several, bundled under one title: reading a flood of email and knowing what matters, drafting replies in someone else’s voice, protecting a calendar from things that do not deserve a slot on it, and quietly knowing who gets access to what. AI genuinely helps with the first two. It cannot do the third or fourth, because both require knowing things about the person and the organisation that were never written down anywhere a model could read them.
That split is the whole answer to "can AI be my executive assistant" — parts of it, done well, and one part that stays a person’s call regardless of how good the drafting gets.
What it is genuinely good at
- Email triage. Reading a long inbox and flagging which messages need a same-day reply, which can wait, and which are noise — using rules you actually state, not a vague sense of "urgent".
- First-draft replies. A response that confirms receipt, answers the parts that have an obvious answer, and flags the parts that need the actual executive’s input — ready to edit, not ready to send unread.
- Meeting prep. Pulling together the last few exchanges with an external contact into a one-paragraph brief before a call, so nobody walks in cold.
- Turning a rough note into a structured document — a set of scattered bullet points into an agenda with a clear ask against each item.
- Travel and logistics drafting. Assembling a first-pass itinerary from stated constraints — dates, cities, preferred airline — for a person to check and confirm, not book unsupervised.
The pattern across all five: the model produces something you would have had to write or read in full otherwise, and a person still decides whether it is right before it goes anywhere that matters. That is the same shape covered for a different role in how to use AI as a banker — the drafting moves, the accountable decision does not.
A worked example: triaging a morning inbox
Weak: "Go through my inbox and tell me what’s important."
Better: "Here are today’s unread subject lines and senders. Flag anything from [list the three people whose messages always take priority] as urgent. Flag anything mentioning a deadline in the next 48 hours as urgent. Group the rest into Needs a reply today, Can wait until Friday, and No action needed. Do not draft any replies yet — just sort."
The second version gives the model your actual priority rules instead of asking it to guess at what "important" means for this specific person’s job, and it separates sorting from drafting into two steps rather than one — so a wrong guess at priority does not also produce a wrong reply sent in the wrong tone.
Sorting and drafting are checkable in seconds. Deciding who gets fifteen minutes with the executive this week is not — that judgement stays with a person who knows why.
Where it needs a person, not a better prompt
- Deciding priority between two genuine competing asks. A model has no way to know that the vendor call matters more this week because of a deal closing quietly in the background — that context lives with the person doing the job, not in the text of either email.
- Anything touching access or permissions — calendar visibility, who gets copied on a sensitive thread, who can see a draft before it is final. Get this wrong once and the fix is a difficult conversation, not an edited message.
- Representing someone in a reply without it being reviewed first. A drafted reply that goes out unread is not saving time, it is transferring risk — what AI is actually bad at covers the general pattern, and these systems are well documented to produce a fluent, confident-sounding answer regardless of whether it is actually the right call, which is exactly the property that makes an unreviewed send dangerous.
- Dates, times and numbers in anything final. A drafted itinerary or invite with a wrong time zone reads exactly as confident as a correct one — check every date against the source before it is confirmed to anyone.
Do not let it grade its own sorting
It is tempting to ask the same tool "did I categorise these correctly?" after it triages a batch. That tends to produce a reassuring answer regardless of whether the sort was actually right — models are documented to favour a response that matches what the user seems to want over a genuinely independent check. The reliable version of this is a person spot-checking a sample of the sorted messages each week against what actually turned out to matter, not asking the model to mark its own work.
Where the tools themselves fit
Most of this does not need a dedicated "AI executive assistant" product — it runs on the assistant features already built into the tools the role already lives in. Google Workspace’s own AI features cover drafting and summarising inside Gmail and Docs directly, and a general-purpose chat tool handles the rest once you paste in the relevant thread or notes. The product category matters less here than the habit of stating the actual rule — whose messages are always urgent, what counts as a deadline — rather than leaving the model to infer it, which is the same discipline Anthropic’s prompting guidance recommends for any request, not just this one.
What to do Monday
- Write down your actual priority rules — the names, the deadline threshold, the senders who always jump the queue — before asking a tool to sort anything against a rule it has to guess.
- Start with sorting alone, not sorting plus sending. Add drafting once you trust the sort.
- Pick one recurring task to test first — finding the repetitive part of a job is the general test for which one — and run it against a week you can already check by hand.
- Keep access decisions and anything sent under someone’s name a manual step, and treat every date or number in a draft as something to verify, not trust — checking an AI answer when you are not the expert is the general habit that applies here.
The wage premium for people who can actually direct these tools well is real and growing, and the same broader skills shift that shows up across knowledge work generally applies just as much to an admin-heavy role as to any other — the edge goes to whoever can state the rule precisely, not whoever has access to the most tools. This sits alongside the same split covered for a whole process rather than one inbox in workflow AI: automate the sorting and the draft, keep the decision that actually closes something.
Coursium teaches exactly this kind of practical skill — stating the actual rule instead of a vague instruction, and the habit of checking a tool’s output before it goes anywhere that matters. Stay ahead of AI by learning the tools on your phone.