Working with others
How far to delegate to an agent
The line between saving work and shifting a risk that stays yours.
An agent can do almost anything a person with a browser can. The useful question is not what it can do but what it should — and the answer turns on one thing: who answers if it goes wrong.
The line
| Kind of task | Delegate? | Why |
|---|---|---|
| Log into a portal and download a document | Yes | Mechanical and verifiable: either it brings the document or not |
| Fill in and submit a familiar form | Yes, reviewing at first | The result can be checked |
| Check whether a certificate meets a criterion | Have it propose | It is a judgement, however mechanical it looks |
| Approve a supplier | No | It has consequences and you answer for them |
| Decide to block someone | No | It affects a third party who did not choose this |
Important
If an agent approves a supplier who did not comply and that ends at an inspection, the answer "the system approved it" does not exist. You answer exactly as if someone on your team had approved it, because you were the ones who let it decide.
How to release it, in practice
- 1
Everything reviewed
For the first weeks, a person confirms every result. That is where you see whether it is right and where it fails.
- 2
Release the mechanical
Downloading, uploading, filling in the usual. If it fails, you notice quickly.
- 3
What judges stays proposing
Permanently, not as a temporary phase.
Watch out
Watch the point where the log stops being read. An agent running well for six months is exactly when nobody checks, and also when a change in the target portal can go unnoticed for weeks.
What delegating does not change
Worth knowing
The healthy way to see it: an agent is someone new on the team being taught a task. You would not hand over every key on day one, and you would not let them sign decisions you answer for.
›Can I undo what it did?
It depends on the action. What was submitted on an external portal, usually not.
›What if the portal asks it something it does not know?
It stops and reports. It does not improvise.
›Is it worth it for few tasks?
Rarely. It pays off on the repetitive, where time accumulates.
A real case
The situation
A company lets an agent approve supplier clearances to move faster.
What you do
- Reverses it: the agent proposes and a person confirms
- Releases only fetching and submitting
What you get
Keeps the genuine time saving and returns the decision to whoever answers for it.
The situation
Full supplier onboarding is delegated to an agent and a week later files are approved unreviewed.
What you do
- Delegates first what touches nobody outside
- Keeps approval behind human confirmation
- Widens only after reviewing what it did in the first week
What you get
Delegation grows on verified behaviour rather than on an expectation.
The situation
The agent decides something that required business judgement.
What you do
- Keeps those decisions out of its scope
What you get
Responsibility stays where it should.
The situation
It is delegated and nobody ever reviews what it does.
What you do
- Reviews a sample periodically
What you get
Trust rests on checks.
The situation
The agent acts out of hours and nobody expects it.
What you do
- Configures when it may act
What you get
Its actions happen when somebody is around.
The situation
Its scope is widened without telling the team.
What you do
- Communicates scope changes
What you get
Nobody is surprised by what they no longer do by hand.
The situation
The agent and a person do the same thing at once.
What you do
- Splits who does what
What you get
The supplier does not receive two requests.
The situation
The agent is switched off and the work piles up unowned.
What you do
- Assigns somebody what it was doing
What you get
Switching off leaves no gap.
This article answers
- can i let ai approve documents
- how far to automate with agents
- who is responsible if the agent gets it wrong
- supervising an ai agent