Skip to main content
Project Intelligence

Agent action control

The safety and integrity of data in corporate systems have the highest priority. To rule out errors, "hallucinations", or unwanted changes, Project Intelligence works on the principle of strict confirmation ("human in the loop").

Working principle: Suggestion → Review → Execution​

The system never makes changes to Jira, Confluence, or other project-management tools in automatic mode without the user's knowledge. The entire process is divided into three stages:

1. Suggestion stage​

When you ask the agent to change something (for example, "Move all tasks for module X to next week"), the AI does not do it immediately. Instead, it:

  • Analyzes the current state of the project.
  • Forms a change plan.
  • Displays a suggestion in the interface (for example, a list of the tasks that will be changed and their new dates).

2. Review stage​

The user sees a clear description of exactly what the agent is about to do. At this stage you can:

  • Approve: if the suggestion is correct, you click the confirmation button.
  • Edit: you can manually correct the suggested data right in the interface before it is applied.
  • Reject: if the agent made a mistake, you simply reject the suggestion.

3. Execution stage​

Only after the user's explicit confirmation does the agent call the system API and make the changes.

Verification of the result​

After the action has been executed, Project Intelligence does not simply report "Done" — it runs automatic verification:

  • It calls the system again to make sure the data was actually changed to the intended values.
  • If an error occurred during application (for example, the task was locked by an administrator), the agent analyzes the error and proposes a corrective option.

Why this matters?​

This approach solves three critical problems of using AI in business:

  1. Trust: you know exactly what will happen to your data.
  2. Accountability: responsibility for changing data stays with the person.
  3. Quality: you act as a filter, cutting out any inaccuracies in the AI's logic before they reach the production environment.