File AI of the SME · Governance

When may AI carry out work autonomously?

Autonomy is agreed task by task. Clear permissions and review points show which work AI may prepare or carry out.

One hand controls a checkpoint between AI tasks and further execution

Start with the action

Summarize a document, write a concept and send an external campaign have different consequences. Therefore, do not give a general mandate just because a task is carried out by the same AI.

Describe the purpose, the required data, the owner and the possible impact. Also see if an error can be repaired and who can notice that error.

Make three levels recognizable

A low-risk task can be performed within a given mandate. Another task may only be prepared. A sensitive or external action may first require explicit human approval.

These levels should be visible in the interface. A button with a vague label does not make clear what the user allows. A good review identifies the concrete scope and the expected effect.

DSC SolutionAI Management PortalInteractive demo
Chat and AI executionYour AI Assistant
Olivier, CEO
Olivier (CEO) · your question

What actions are you allowed to perform yourself?

DSC AI Assistant

This example mandate allows reading and preparing. An external change requires separate approval.

  • Voorbeeldmandaat
  • Proceseigenaar
  • Reviewlog
  1. Context Summary

    Allowed within the reading mandate.

    Within example mandate
  2. Preparing draft action

    Allowed without external changes.

    Within example mandate
  3. Execute External Action

    Blocked until the competent owner approves the proposal.

    Approval required

May I prepare this proposal? First, look at the steps and boundaries.

Status: proposal ready

Example interaction · no live business data or external execution.

Approval is part of the workflow

The competent owner must have sufficient context to assess a proposal, identify the action that has been approved and the conditions that apply, and prior approval is not an unlimited consent for subsequent steps.

Where architecture supports this, logs help to reconstruct actions, errors and review moments. Not every external action can be reversed; that is precisely why assessment before execution counts.

Let the mandate grow with proof

Start limited and evaluate the operation. Do not extend the mandate until the task, data and results are sufficiently verifiable. New systems or a larger target group may request re-assessment.

Good governance supports continuous implementation. The customer determines the pace and limits; DSC helps translate these agreements into design, guidance and quality assurance.

Frequently asked questions

Does a person have to approve every concept?

Not necessarily. Preparation may fall within a pre-arranged mandate. External or sensitive execution is a separate decision.

Can any action be reversed?

No. Design recovery wherever possible and use prior review when the consequences are not fully reversible.

Translate this into your own work question.

See the appropriate module or discuss which step will help your organisation.

When may AI carry out work autonomously? | DSC Solution