AI Executive Assistant: I Don’t Use an App. Here’s the Agent That Triaged 573 Emails (and Never Answered a Finance One)

Search for AI executive assistant and you get a wall of app roundups, and most of them rank their own product first. Calendar tools, inbox tools, note-takers. Useful, sure. But an executive assistant was never an app. It’s a set of rules: what gets through to you, what gets handled without you, and what nobody touches until you say so.

I don’t use an AI executive assistant app. I run an agent that works my business inbox twice a day, every day, from a container on a server while I’m in Taipei. In the last 30 days (14 October to 12 November 2026) it went through 573 emails across 60 runs. It wrote about nine reply drafts and sent none of them. It flagged 55 finance emails and answered none of those either. That last part isn’t a bug. It’s the most important rule it has.

This post gives you the rules, the receipts (including the mistakes), a short and honest map of which apps cover which part of the job, and the things you should never hand to software. No “I tested 47 tools” claims. Just one operator’s delegation log.

What an AI Executive Assistant Actually Does (and Which Duties Software Can Hold)

Before you pick a tool, write down the job. A good human executive assistant does five things, and only some of them move cleanly to software.

1. Gatekeeping. Sorting what comes in so the principal sees the 5% that matters, not the 95% that doesn’t. Software can hold this. It’s pattern-matching, and models are very good at it.
2. Protecting time. Defending focus blocks, rescheduling around conflicts, saying no to meetings that shouldn’t exist. Software can hold most of this. Calendar tools like Reclaim and Motion were built for exactly this.
3. Drafting in your voice. First drafts of replies, follow-ups, polite declines. Software can hold the draft. You keep the send button.
4. Escalating what only you can decide. Money, commitments, anything where the wrong answer costs a relationship. Software can hold the routing rule (“anything with a price in it goes to me”). It cannot hold the decision.
5. Judgment and relationships. Knowing that this investor hates long emails, that this client is about to churn, that two people in the same meeting shouldn’t be. Software can’t hold this yet. Not reliably, and not without context you’d have to write down anyway.

That’s the whole frame for this post. An AI executive assistant is great at duties 1 to 3, good at the mechanics of 4, and bad at 5. Every product page blurs that line. Your rules shouldn’t.

If you’re coming at this from the other side (“can I just replace my VA?”), I covered that swap, task by task, in what transfers when you replace a virtual assistant with AI. This post sits one layer up: the executive part, which is mostly about what doesn’t get done without you.

The Rules My AI Executive Assistant Follows

My email agent is a scheduled skill: a plain-English instruction file that Claude Code runs headless at 11:05 and 21:05 every day against one business inbox. There’s no app and no dashboard. Here are the rules, almost word for word.

Rule 1: Every email gets exactly one of four labels

• Customer. Someone asking about a service or a problem. Gets a draft reply.
• Partnership. Collaboration, guest post, referral, integration. Gets a draft reply that asks exactly one qualifying question.
• Finance. Invoices, receipts, payment failures, anything with money in it. Never gets a reply. It becomes an Asana task for me.
• Misc. Notifications, newsletters, cold pitches, the rest. Labelled. If it’s ambiguous or might matter, it becomes an Asana task too.

Rule 2: Draft, never send

Replies go into Gmail Drafts, written as me and signed “— Jon”, 3 to 5 sentences. The agent has no send step. I read, edit, and press send myself, or I delete it.

Rule 3: No commitments, ever

The instruction is blunt: never commit to pricing, timelines, or partnerships. Express interest, ask a question, let the operator decide. An assistant that agrees to a price on your behalf isn’t saving you time. It’s creating a negotiation you didn’t choose.

Rule 4: Leave everything unread

The agent never marks mail as read. The label is how it knows it already handled a message. The unread count is mine. This sounds small, but it’s the difference between an assistant that sorts your inbox and one that quietly hides things from you.

That’s it. Four rules, about a page of plain English, and that page is the AI executive assistant. Nothing here is clever. The value is that the rules are written down, so they run the same way at 11pm on a Sunday as they do when I’m watching. That’s also why I think of these as AI employees with a job description rather than chatbots.

30 Days of Receipts: 573 Emails, 9 Drafts, 0 Finance Replies

Here’s what the AI executive assistant actually did, counted from its own run logs, 14 October to 12 November 2026.

Category	Emails	Share	What the agent did
Misc	508	88.7%	Labelled, left unread. Anything ambiguous became an Asana task.
Finance	55	9.6%	Asana task for me. Zero replies.
Customer	7	1.2%	Draft reply as Jon, or a task when it couldn’t answer from the record.
Partnership	3	0.5%	Draft with one qualifying question.
Total	573	100%	60 runs: 59 success, 1 skip (inbox already processed)

Three things jumped out when I counted.

1. Almost nothing needs me. 10 of 573 emails (1.7%) were from a human who wanted a human answer. That’s the real case for an AI executive assistant: it isn’t writing your emails, it’s sparing you from opening the other 98%.

2. Drafts are rare, and that’s correct. Across the month I count about nine reply drafts. At least one I rewrote before sending. That’s the system working: the agent handles the first 80% of the reply and I handle the 20% where tone matters.

3. The escalations are where the value is. A few real examples from the logs, lightly anonymised:

• A referral partner replied to an outreach campaign asking for rates, terms and a commission structure. The agent didn’t draft anything (“rates and commission are your call”) and opened a task with a due date instead.
• A monthly software receipt arrived with a changed plan name on the line item. It became a finance task, with a note pointing out the name change so I could confirm it was expected.
• A customer asked a question about an admin console the agent can’t see. Instead of guessing, it drafted a short holding reply that answered only what was already on record and left the rest to me.

The mistakes (because there are always some)

Two worth sharing, because they’re the kind every AI executive assistant will hit:

• The reply that would have gone to myself. Contact-form submissions arrive from my own address to my own address, with the real sender tucked into the Reply-To header. A normal threaded reply would have landed back in my own inbox. The fix is now a rule: for form submissions, draft a fresh email to the Reply-To address, never a threaded reply.
• The draft pile-up. When a partner followed up on a thread where my earlier draft was still sitting unsent, the naive move was to add a second draft. The agent rewrote the existing one instead. I only know to check for that because it once nearly happened.

Neither mistake was dangerous, because of Rule 2. Nothing goes out without me. That’s the whole point of keeping the send button: your mistakes become drafts, not incidents. (If you want the longer version of why “it ran successfully” isn’t the same as “it did the right thing,” I wrote about it in an empty result is a claim, not a fact.)

AI Executive Assistant Apps, Mapped to the Duty They Cover

I haven’t run these AI executive assistant apps in production, so this isn’t a “we tested them” ranking. It’s a map: which duty each tool is built for, and what its pricing page says today. Prices were checked on the vendors’ own pages on 13 November 2026. Check them again before you buy, because this category reprices often.

EA duty	Tool	Listed price (13 Nov 2026)	What stays with you
Protecting time (calendar)	Reclaim	Free Lite plan; Starter from $10/seat/month	Which meetings deserve to exist
Protecting time + tasks	Motion	Pro AI from $19/seat/month (credit-based)	Priorities when everything is “urgent”
Inbox speed	Superhuman	Pro $12/month billed annually ($15 monthly); Business $33 ($40)	The send button
Inbox triage + drafts	Fyxer	Pro $37/seat/month annually ($49 monthly)	Tone with people who matter
Custom rules across inbox, calendar, Slack	Lindy	From $49.99/user/month for 5,000 credits; $50 free credits for 7 days	Writing the rules (same as my setup)
All of the above, your rules, your infra	Claude Code + Gmail API + Asana (my setup)	Runs on the Claude subscription my fleet already uses, plus setup time	Maintenance. When it breaks, it’s yours to fix

Two honest notes on that table.

First, notice that every tool covers one or two duties. A calendar tool doesn’t triage your inbox, and an inbox tool doesn’t know your finance rules. Most people who say “my AI executive assistant didn’t work” bought one tool and expected all five duties.

Second, the last row isn’t for everyone. I build my own because I already run a fleet of agents on cron, so one more skill file costs me almost nothing. If you don’t, an off-the-shelf tool with good rules support (Lindy, or Fyxer for inbox-only) is a saner start. For the wider app roundup, see my comparison of AI personal assistants.

If you’d rather have the rules written and the agent wired up for you, that’s the kind of build I do. Book an automation strategy session and bring your inbox numbers.

What You Should Never Hand an AI Executive Assistant

This is the section every AI executive assistant roundup skips, because it’s the opposite of a sales pitch. Here’s my list, and every item on it comes from a rule I actually run.

1. Money, in either direction

No replying to invoices, no confirming payments, no “yes, go ahead and renew.” 55 finance emails in 30 days, and every one became a task. The cost of a wrong yes on a payment is real money. The cost of me reading a task is about 30 seconds.

2. Commitments: price, dates, scope, partnerships

Anything that creates an obligation. An AI executive assistant can say “thanks, Jon will get back to you on rates.” It must never say “our rate is X.” That’s not because the model can’t do the maths. It’s because a commitment is a decision, and decisions are yours.

3. Anything relationship-ending

Complaints from your best client. A tense thread with a partner. A message to a board, an investor, or family. These are exactly where a slightly-off tone does lasting damage. Let the agent flag them. Write them yourself.

4. Access it can’t explain

If you can’t say in one sentence why the assistant needs a permission, don’t grant it. My email agent can read, label, and draft. It can’t send, can’t delete, and can’t mark things read. Fewer permissions mean smaller mistakes.

5. “Proactive” actions without a log

Proactive sounds great until you can’t tell what it did. If an assistant acts on your behalf, you need a record of every action. Mine writes one per run. I dug into whether proactivity is worth it in this FAQ Friday on proactive agents.

Can AI Replace a Human Executive Assistant?

Partly, and the split is more useful than the yes/no.

The US Bureau of Labor Statistics puts the median pay for executive secretaries and executive administrative assistants at $76,590 a year (May 2025), before benefits. Most AI tools in the table above cost under $50 a month. On price alone it’s not close, and that’s exactly why the question gets asked badly.

Here’s the honest split from running an AI executive assistant every day:

• AI wins on volume and consistency. 573 emails sorted the same way at 11:05 and 21:05, every day, including weekends. No human EA wants that job, and no human should have it.
• A human wins on judgment and relationships. Duty 5 from the top of this post. Knowing who matters, reading the room, rescuing a trip that falls apart halfway. My agent can’t do any of that.
• The best setup is usually both. If you have a human EA, the AI should work for them: it clears the 98% so they spend their day on the 2% that needs a person.

For a solo operator like me, there’s no human EA to protect, so the agent takes duties 1 to 3 and I take 4 and 5. That’s still a big win. I open the 1.7% of email that needs me plus a short Asana list, instead of 573 messages. But it’s not “replaced.” It’s delegated, with the decisions kept. I made the same argument against the “AI agent is just a cheaper VA” line in this Monday Myth post.

AI Executive Assistant FAQ

What is an AI executive assistant?

Software, usually built on a large language model, that handles parts of an executive assistant’s job: triaging email, drafting replies, managing a calendar, and routing decisions to you. The good ones follow written rules about what they may handle alone and what they must escalate.

How much does an AI executive assistant cost?

On the vendor pricing pages I checked on 13 November 2026, most plans run from about $10 to $50 per user per month, often with a credit allowance. A custom-built agent costs whatever your model subscription and setup time cost you. Either way it’s a small fraction of a human EA’s $76,590 median salary.

Is it safe to give an AI assistant access to my inbox?

It’s as safe as the permissions you grant. Start with read, label, and draft only. Don’t allow send or delete, and don’t let it mark mail as read. Then nothing leaves your inbox without you, and nothing disappears from it.

Can an AI executive assistant reply to emails for me?

It can, but I don’t let mine. It drafts, and I send. In 30 days it wrote about nine drafts and I still rewrote at least one before sending. The time saved comes from not having to open the other 563 emails, not from skipping a 30-second review.

What’s the difference between an AI executive assistant and an AI personal assistant?

A personal assistant app helps you do your own tasks (reminders, notes, scheduling). An executive assistant protects your attention for other people’s requests: it gatekeeps, escalates, and keeps you from committing to things by accident. Same tech, different job description.

Do I need to code to build one like yours?

No. The rules are plain English, and most of the “code” is the agent writing its own API calls. You do need to be comfortable reading what it did. If you’re still deciding between building and buying, this comparison of agent runtimes covers which one I’d trust to run unattended.

Final Thoughts: Write the Rules Before You Buy the App

An AI executive assistant is only as good as the rules you give it. Before you compare a single pricing page, write a page of plain English answering four questions: which categories of email exist in your inbox? Which ones may be drafted? Which ones must never get a reply? And what can the assistant not touch?

Once you’ve got that page, the tool choice gets easy. It’s whichever one can actually follow it. Mine’s a scheduled agent because I already run a fleet, and it sorted 573 emails last month while I answered 10. Yours might be Fyxer or Lindy, and that’s fine. What matters is that the finance email never gets an automatic “sure, sounds good.”

Start strict. Loosen the rules only when the receipts tell you it’s safe.