How to use AI agents for work you still review

How to use AI agents is a handing-over problem, not a prompting trick. You write the outcome, you point at the files, and you read what comes back before it leaves. What can AI agents do sits in the middle: the steps you currently do yourself after a chatbot replies.

What can AI agents do

They can carry a task across steps and leave a file you can check. In practice that looks like:

• Research that comes back with the source next to the claim
• A first draft of a plan, a post, an email, or a memo, saved in the folder
• A checklist with owners, from a process you already run
• A meeting brief, or a weekly update, built from notes you provided
• A list of mismatches between files, with the rows flagged for you to open

A chatbot can help you think. An agent is for the part that has to land as a file. The difference, and the test for knowing you have outgrown chat, is AI agent vs chatbot.

What they should not do is the part with a consequence you cannot undo from the board. Sending, paying, publishing, and any claim you have not checked stay with you. If the files do not contain a number, the agent does not get to invent one.

How to use AI agents

1. Write the outcome, not a vibe. "Some ideas for the launch" gives you a chat. "A one-page launch plan and three social posts, from this brief, in this folder" gives you something to edit.

2. Put the files in the folder once. The transcript, the ICP note, the last report, the brand lines you actually use. Pasting the same context into a new chat every morning is the habit you are leaving.

3. State the limits in the brief. What must not be claimed. What must not be sent. Whose voice the draft has to sound like, with an example in the folder.

4. Split the work by role when the job has more than one kind of step. Research, a draft, and a check are different passes. The marketing, sales, founder, and ops versions of that split are the squad pages linked from xSquads for Work.

5. Set the review before anyone starts. The work waits there. You approve it, or you send it back with a note. Nothing that leaves the company skips that step.

6. Keep the correction. A rejected line, a wrong account, a figure you removed: the next brief should start from that, not from zero.

That is the method. It is the same shape as briefing a careful teammate.

One brief, walked through

The brief: "Turn the onboarding notes in this folder into a checklist with owners. Flag any step that has no owner."

You do not ask for "help with onboarding." You name the file you want back.

A process role reads the notes and writes the checklist. A second pass marks the steps that have no name next to them. You open the checklist. You assign the two steps the notes left blank, because those names are a decision, not a transcription. You delete a step the notes do not contain. Then you approve.

The operations version of this job, with the named roles, is the Operations Squad. This walk-through is the method. The squad page is the roster.

Where the first week usually goes wrong

People ask for ideas, then copy the answer into the real document by hand. The agent never saw the document, so the next question starts over.

People skip the review because the draft looks finished. A finished look is not a check. Open the claim, the number, and the name.

People hand over sending on the first day. Draft the email. Read it. Send it yourself until you trust the pattern.

Where this runs

xSquads is a desktop app. You point a mission at a folder you already use, and the squad works a shared task board. You read what landed. Mac now, with Windows coming soon. The product page is xSquads for Work, and the waitlist is how you get the download. A personal AI agent is the single-assistant version of the same idea. A squad is that idea with roles.

Frequently Asked Questions

How do I start using AI agents?

Start with one outcome and the folder it belongs in. Name the file you want back, hand over the source files once, and read the result before anything is sent. One checklist is enough for the first week.

What can AI agents do that a chatbot cannot?

They can carry the steps and leave a file, a list, or a draft in the folder, instead of a reply you copy out by hand. You still check the result. They should not send, spend, or publish before you have read it.

Do I have to prompt every step?

No. Write the outcome and the limits once, and let the roles split research, drafting, and a check. You come back at the review, not after every sentence.

How should I review an agent's work?

Open the file. Check any claim, number, or name against the source you provided. Edit what you would not sign. Send it back with a note when the miss will happen again.

What should I not hand to an AI agent?

Anything you cannot check, and anything that leaves the company before you have read it: a send, a payment, a publish, or a legal or pricing decision. Drafts of those are fine. The action is not.

Can I use AI agents if my work lives in files?

Yes. Point the mission at the folder the work already lives in. xSquads for Work is the desktop app built for that folder.

Related Reading

• AI agent examples: marketing, sales, a founder, and ops, in the same shape.
• AI agent vs chatbot: the difference between a reply and a finished file.
• Personal AI agent: one assistant, then a squad.
• Operations Squad: the checklist example as a roster.
• xSquads for Work and the waitlist.