What an AI chief of staff actually does in a week

The phrase "AI chief of staff" now turns up in software marketing about twice a week. Most of what it is attached to is a chat window with a new label on it: you still ask, it still answers, you still close the tab. What changed in August is smaller and more specific than the marketing. An agent can now have a computer of its own.

That is the whole difference. A cloud machine with a browser, files and a terminal, signed into your accounts with your logins, running while your laptop is shut. Grok Bot has been bundled into every SuperGrok and Cursor plan since 26 August 2026, which puts the entry price around HK$160 to HK$240 a month. The intelligence is roughly what you already have in ChatGPT. The persistence is new.

What separates an AI chief of staff from a chatbot

A chatbot waits. A chief of staff agent runs on a schedule and on standing orders, and it manages the other agents.

That last part is the bit people miss. One agent doing everything becomes a mess within a fortnight, because the instructions for reading your inbox and the instructions for filing your quarterly numbers have nothing to do with each other. The working shape is a small fleet: three or four specialists that each do one narrow job well, and one agent above them that knows your week, decides what is worth your attention, and writes you a single briefing instead of four.

The value is not the work it does. It is the forty small decisions a week it takes off your desk without asking.

An ordinary week

Concretely, on a normal Tuesday:

• 06:40. The overnight inbox has been read. Forty-one messages in, three flagged as needing you, each with a two-line summary and a draft reply already written in your wording.
• 07:15. The school circular that arrived at 11pm has been turned into a form, filled in, and parked. It is waiting on your yes, not on your attention.
• 12:30. The dinner nobody had booked has a table held for Friday, chosen from the three places you have been before.
• 21:00. The week's numbers are filed in Notion, in the same format as last week, so the comparison takes ten seconds instead of an evening.

None of that is impressive individually. A competent assistant does all of it. The point is that none of it reached you as an interruption, and the total is roughly six hours you did not spend.

What it holds for your yes

The rule that makes this survivable is written down before anything is switched on. Standing orders: what the agent does alone, what it drafts and holds, what it never touches.

Reading, summarising, drafting, scheduling, filing and researching go in the first bucket. Anything that leaves your name on it in front of another person goes in the second. Two things go in the third by default, and I would not move them for anyone: your employer's systems, and any client's confidential data. If you work in a Hong Kong bank or a law firm, the fleet gets built entirely on your personal accounts and stays away from work systems, and the standing orders say so in writing. Start an agent on your own admin. Never on the client file.

This is also the honest answer to "is it safe to give it my logins". It is safe in the way giving an assistant your logins is safe: with rules, a log, and the ability to revoke in about a minute. It uses your accounts, so nothing is copied anywhere else. The questions worth asking first are the same ones I set out on whether ChatGPT is safe for work data, and they apply here with more force, because an agent acts rather than answers. If those answers do not satisfy you, the correct move is not to do it.

The Hong Kong parts

Three things that never appear in the American write-ups.

The first is the locked laptop. A large share of professionals here cannot install anything on a work machine, and the compliance team is not going to make an exception. Agents on cloud computers sidestep this entirely, because nothing installs anywhere. Your dashboard is a Notion page you open on your phone.

The second is code-switching. If half your messages move between English and Cantonese inside a single sentence, an agent will summarise them badly and confidently. It is the same failure mode I described in AI note-taking on a bilingual meeting. Test it on a real week before you trust it on anything that matters.

The third is that the useful jobs here are often not work jobs. Domestic helper contract renewals, school admin, Mandatory Provident Fund paperwork, the visa run. Nobody is going to build you software for those. An agent is written instructions plus the accounts you let it use, so anything you can describe in a paragraph can be built.

What it still gets wrong

I run a fleet on my own business, so this is first hand rather than theoretical.

It over-summarises. Given a long thread with one buried sentence that changes everything, it will hand you a tidy paragraph that leaves the sentence out. The fix is a standing order to quote directly from anything it flags, not to paraphrase it.

It is confidently wrong about scheduling across timezones roughly once a fortnight. Check anything it books with someone in London before you rely on it.

And it degrades quietly. An agent whose instructions were written for the job you had in March is still doing that job in September, just less usefully. Rewriting the standing orders takes twenty minutes and almost nobody does it.

There is also a version of this that fails completely, which is building the fleet before you have proved the underlying work is worth automating. I turned off my own first agent on 9 August because I had spent more time maintaining it than it had ever saved me. If a job is not already costing you an hour a week, automating it costs you more than it returns.

Where to start

Not with the tools. With one week of honest observation: write down every task you did that a competent stranger could have done with your logins and ten minutes of context. Most people find between six and twelve. Three of them are worth handing over immediately, and the rest are noise.

Then pick the platform. My comparison of Grok Bot, OpenClaw and Hermes Agent covers the trade-offs, and the free prompt library is a reasonable place to practise the instruction-writing before an agent has any access at all.

If you would rather not lose a weekend to it, this is now the thing I do most. A chief of staff plus three specialists, set up on your real inbox and calendar in one two-hour session, standing orders written down, running before the session ends. The catalogue and the prices are on the AI agents page. The first conversation is fifteen minutes and free, and I will tell you if the honest answer for your week is "not yet".