週一闢謠：「人工智慧助理只是更便宜的虛擬助理」（我的助理寫了314份草稿，但一篇也沒發）

Every week someone sends me the same plan: fire the VA, hire an AI agent, save two grand a month. On paper it is unbeatable. A part-time assistant costs you $600 to $2,000 a month. An agent costs a few dollars a day in API calls, works at 3am, never forgets step four, and never takes a holiday.

I run 22 scheduled agent jobs on this brand alone — 12 every day, 10 every week. So I am the last person who is going to tell you agents do not work. But the number that killed this myth for me was not a cost. It was a zero.

314 drafts. Zero posted.

I measured this in my own container an hour ago, not from memory:

• Since 16 June, one of my agents has written 314 social reply drafts. Good ones — researched, on-voice, ready to paste.
• Zero have ever been posted. Not one. The downstream log table that records a posted reply has been empty since the day it was created.
• 274 of them quietly expired unread. 40 are still sitting there waiting.
• Every single one of those runs exited 成功. Across 1,337 logged runs since June, 1,014 came back clean.

The agent did its job perfectly, 314 times, into a room with nobody in it.

Then I checked the other pile

Same morning, same container. My agents have filed 1,215 tasks into my project board. 699 are still open. Of those, 287 are explicit “a human has to decide this” items — finance notices, flagged emails, partnership enquiries.

The oldest one has been sitting there since 15 June. That is 98 days of an agent politely raising its hand at a wall.

A VA would have asked on day two

Here is the part the cost comparison hides. What you were actually buying from a good assistant was never the hours. It was one sentence: “Hey, quick one — I’ve done eight of these and nobody’s touched them. Should I keep going?”

That sentence is the whole product. It is judgement, backpressure and a stop button bundled into a person who gets bored. A VA will not silently write 314 drafts into a void, because a human being finds that unbearable by about number six.

My agent found it completely fine. It has no opinion about whether its work is being used. It was asked to produce, it produced, it reported success, and it went back to sleep. That is not a bug in the agent. That is the deal.

AI agent vs virtual assistant: what you are actually buying

Set the price tags aside and the two things are structurally different purchases.

A VA is capped output with escalation included. They can only do so much in a day, and when something looks wrong they interrupt you for free.

An agent is uncapped output with escalation sold separately. It will produce until you run out of credit, and it will only interrupt you if you build the interruption, route it somewhere, and then personally read it. All three. I built the first two and skipped the third, and that is exactly how you get 699 open tasks.

So you do not delete the cost when you swap a person for an agent. You move it. It stops being a salary and becomes your attention — and unlike a salary, nobody sends you an invoice when you fall behind on it.

Three things that actually fix this

1. Name the consumer before you build the producer. For every output your agent creates, answer: who or what opens this, and how often? “It goes in a table” is not an answer. A table is storage, not a reader. If nothing downstream consumes it, do not automate it yet — you are just paying to generate landfill.
2. Give the queue backpressure. My agent should have refused to write draft number 315 while 274 sat expired behind it. Make the producer count its own unconsumed output and stop — or at minimum shout — when the pile crosses a threshold. A queue with no ceiling is not a queue, it is a leak.
3. Keep exactly one channel you genuinely read. Twelve hundred tasks is not escalation. It is a landfill with timestamps. Pick one place — a phone alert, one inbox, one board column — and rule that anything filed anywhere else is an archive, not a request. Then actually clear it, weekly, on a timer.

要點： an AI agent is not a cheaper VA. It is an uncapped VA that never once asks whether any of this is working. The cheap part is the production. The part you still have to pay for — in attention, in design, in a weekly ten minutes — is knowing when to stop.

If you are weighing this swap right now, I answered the five questions people actually ask me before replacing a VA with an agent. The backlog problem above is the one I warned about in clear the approval queue before it clears you — and then demonstrated on myself for 98 days. It is the same family of failure as “sent” not meaning “published”, and the same reason prompt engineering is never the hard part.

Want the version of this where the output actually lands somewhere? That is the unglamorous plumbing behind agents you can genuinely leave alone. 預約自動化策略會議 and I will show you where my own pile is buried.