YEMPIK / STUDIOSoftware, projects and training.
All articles
Type
Guide
Topic
Governance
Published
Reading time
2 min read

Which actions may an AI agent take, and when must it stop for human review?

Before an agent goes live, you write down which actions it may take without asking. How far you can trust it depends on that list.

By, AI Architect & Co-Founder, Yempik
In this article
In summary
  • Start from a written list: every action the agent could take goes into one of three columns. It does it alone, it prepares it and waits for a yes, or it never does it.
  • The test is the possible harm: an action that can be undone in one click can run alone; one that reaches a customer or moves money goes through a person.
  • The agent must also stop when it is unsure: data is missing, the request is not on the list, an operation’s outcome is uncertain.
  • Permissions are set in the systems, not only in the instructions: an agent that must not delete has no permission to delete.

All frequently asked questions →

The three columns

Take the task, write down every action the agent could take and put it in a column.

  • Alone: reading, classifying, searching, drafting, updating an internal field that is easy to correct.
  • With approval: emailing a customer, confirming an order, changing a price, writing ERP data that others will use.
  • Never: deleting data, paying, signing, changing other people’s permissions.

The list changes over time. An action moves from “with approval” to “alone” when, for a few weeks, almost every approval was a yes.

When the agent must stop

Beyond the columns you need stop rules. The agent stops and asks when data is missing, when the request resembles no expected case, when two sources disagree, when an operation gives no clear outcome.

An uncertain send, for example, must be checked before retrying, otherwise the customer gets two emails. The email agent page walks through it.

How oversight works

Every agent action leaves a trace: what it did, on which data, who approved. Approvals have a named owner and a time limit, so a draft does not sit for a week. The agent’s account sees only the folders and fields its task needs.

More questions

Who is liable if the agent gets it wrong?

Towards the customer, the company using it, as with an employee’s or a software error. That is why actions that reach customers go through a person.

Don’t approvals slow everything down?

A little, at first. They measure how often the agent is wrong: when it is rarely wrong on an action, that action moves to the ones it does alone.

Transparency note

Answer edited by Raffaele Zarrelli, co-founder of Yempik, written with Claude. Prices, timelines and cases are the ones Yempik publishes; this page was last updated on 8 November 2026.

Shall we look at your case?

Tell us about the work you want to start from. The first call is free, and if the project does not make sense we will say so.

Book a call
AI agents: which actions alone, when to stop for review | Yempik