An AI for work can prepare a great deal. The decisive question lies elsewhere: at what point does a person give their approval, and to what exactly? Well-placed human approval separates preparation from action and makes every approval verifiable.
IN SHORT
- Show what has been prepared before proposing an action.
- Make approval cover one specific action: its content and its destination.
- Then show the real status: done, declined, failed or unknown.
Make the preparation visible
Before showing an action button, show what has been prepared. A precise title, the full text and the chosen sources help people understand what they are about to review.
A prepared message must stay clearly marked as a draft. Nothing in its appearance should suggest that it has already gone out.
Make approval cover a specific action
“Approve” can mean very different things. For an external action, the summary must make it possible to identify what it is and where it is going: the sender, the recipient, the subject, the text, the attachments.
A button that opens a document must not trigger a send. Approval to prepare a version is not permission to publish it. If the content changes, a new approval is needed.
A simple rule: approval covers only what is shown. Nothing more, nothing else.
Show what happens next
Visual confirmation must match a known result. Each outcome has its own status: done, declined or failed. An action whose outcome is unknown deserves an explicit status, not a reassuring message.
In a multi-step workflow, separate automatic steps from those that wait for a person in the same way.
In Inkore
Inkore gathers these decisions in “Review”. Each proposed external action — sending an email, creating an event, uploading a file — opens on its exact content and on what the approval covers: one single action, nothing else. You confirm, postpone or decline, and what has been done can be viewed afterwards. See the product journey.
Frequently asked questions
What does “human in the loop” mean?
Human in the loop means that a person steps in before an AI action takes place. What matters most is to state exactly what the approval covers — the content, the recipient, the single action concerned — and to leave a genuine option to refuse.
Should you approve every step an AI takes?
No. Internal steps — reading a source, preparing a draft — can go ahead on their own. It is the actions that leave the tool, such as sending or uploading, that require explicit approval. Asking for approval at every step would soon turn it into a reflex, and a reflex protects nothing.