Agentic AI · 28 September 2026
AI employee: what the phrase hides
An AI employee is not a new colleague who can do everything. It is software with a small job, a small set of permissions, and a person who still owns what leaves the room.
Somebody tells you the company has hired an AI employee. Ask what the software may touch when nobody watches it.
It may read a folder. It may write a draft. It may send an email, change a refund, or update a customer record. Those are different jobs. Calling all of them an employee hides the difference.
The question hiding behind "AI employee"
An employee has a job, a manager, and rules. Software needs the same three things before it touches work that belongs to somebody else.
Start with a copied inbox. The software reads the messages, puts them in piles, and writes draft replies. A person still reads the draft and sends the email.
Do not point it at the whole company drive because the sales page says it works all day. Leave out passwords, bank details, and the browser that can send something.
Tell the software what it may do, what it must leave alone, and what needs a person. That written instruction is a permission rule.
The 3 keys you are really handing over
There are 3 keys worth naming before you call software an employee.
Read files. Reading a price list or a folder of old replies can save you time. It also means the folder must contain only files the software may see.
Write drafts. A draft can be useful because you can compare it with the message and the source before anyone receives it.
Reach outside the folder. Sending an email, approving a refund, changing a record, or moving money reaches somebody else. Keep that key with a person until the job has earned more trust.
The current saved-choice label is "Yes, and don’t ask again." It applies to the rule that matches the tool or command, not unrelated actions. It does not allow another program to send email.
A draft is not a decision
A draft needs a visible rule and a person who checks it before it leaves the company. Give software work that leaves that trail.
Let it sort copied messages and write drafts. Ask it to put the source line under every number. Then read the list before the draft.
The safe first test
Make a small test folder. Put copied messages inside it. Add only the files it may read. Remove the email connection, browser, or any other tool that can send something.
What may it read? The copied messages and the price list.
What may it write? A draft reply. Every number names the matching price list row or original message.
What may it send? Nothing. You read the draft and choose whether it leaves.
Run the job. Open one draft. Open the price list row or message named under it. Check that both say the same thing.
If the software guessed, the folder is the right place to find that out. Nothing can leave the test while no permission path can send it.
What may it read, write, send, and who checks what leaves?
That is the useful meaning of an AI employee. It does not mean a new colleague who can do everything. It means a small job, a small set of keys, and a person who still owns what leaves the room.
Where do you stand?
15 short situations about handing a job to an AI tool. In each one you pick what you would do next.
- ▸Whether you sized the job right
- ▸What you handed over
- ▸How you checked what came back
- ▸When you stopped
About 5 minutes. You see your score, your weakest area and the reasoning behind every answer before anybody asks you for an address.
Questions people ask next
answered in one line eachWhat is an AI employee?
It is a sales phrase. Before you use it, name the files the software may read, the drafts it may write, the outside actions it cannot take, and the person who checks the result.
What does always allow mean in Claude Code?
An allow rule approves the matching tool or command. It does not approve unrelated tools or actions.
What is a safe first AI employee job?
Use copied messages in a small folder. Let the tool sort them and write drafts. Keep sending and other outside actions with a person.