An AI draft is not an authorized action.
Separate the model, the application and the workflow around them.
Dash-Squared / Brick learning lab
A useful draft.
A separate decision.
Try the fictional follow-up. The workflow stops for your explicit example choice before its illustrated tool phase.
A synthetic workflow
A draft is not an action.
Prepare the fictional follow-up, then make an explicit example choice. Nothing is sent.
Fictional follow-up / not generated yet
Choose Play or Step to prepare the fictional follow-up.
Checks in this example: commitment, recipient and scope.
Play is finite. Pause or use Step to read at your own pace.
Demo only. Nothing is sent.
Read the complete explanation without playback
- Model proposal.A fictional draft contains a promised completion time that has not been verified.
- Workflow waits.Play, Step and reduced motion cannot bypass the explicit example decision.
- Two review choices.Approve this exact example to illustrate the tool phase, or return it for revision. A revised draft needs a fresh approval.
- Tool authority is separate.Only the demo highlight changes after approval. No email, API or external action runs. Review does not guarantee accuracy.
This example uses a human review policy. Approval permits the illustrated tool action; it does not guarantee the draft is true.
Primary source: n8n human review for tools ↗
Checked November 9, 2026. Illustration and interaction are synthetic.
A good draft is not permission to send.
AI and automation / separate a proposed result from an authorized action.
Follow the idea
- A model proposes a result inside an application.
- The example workflow validates the draft and obtains human approval.
- Only the authorized action crosses to the sending tool.
Keep the limit visible
This is one approval policy, not a definition of every agent. Models can produce errors; tool permissions, allowed data and failure handling remain application responsibilities.
May the model draft send itself?
Not in this example. The workflow requires review and approval before the tool may send. Other systems need their own explicit policy.
“If an AI answer is convincing, it is safe to let it act.”
Evaluate the output and separately constrain the system’s authority to take actions.Source for this graphic: developers.openai.com ↗ docs.n8n.io ↗ · Checked November 7, 2026 · Illustrative example.
Name the roles
A model generates or interprets information. An assistant is an application someone uses. A workflow coordinates steps; it can be deterministic and may use AI. Agent designs can allow a model to choose steps or tools within defined limits. These terms do not certify a product’s reliability.
A concrete example
Synthetic example: a model drafts a follow-up, checks identify missing facts and a person reviews the message before the sending tool is allowed to act. Human review is the policy chosen for this example, not a definition of every AI system.
What needs verification
Test useful output and failure behavior separately. Evaluate accuracy, allowed data, access, retries and whether the system can take an action beyond the intended scope. A convincing demonstration is not proof of dependable ongoing operation.
The next useful question
What may the system propose, what may it execute, and who owns exceptions?
See it in motion
Independent, unsponsored education. This is an explanation of a tool and a possible approach, not a claim of delivery experience, certification, endorsement or partnership. Verify the actual product, plan and configuration.
Published .
Useful reference. Untrusted instructions.
Layered defenses reduce risk; they do not guarantee safety.