Human in the loop: how to design meaningful approval of AI actions

Start with the consequences of the operation

Preparing a draft, saving data, and sending a message have different effects. Determine what can be easily reversed and what creates an obligation to the client. Define the required level of control for each class of actions. Do not add identical approval to every minor step, as the user will quickly stop reading notifications carefully.

In the quoting process, a human can check the scope, price, and recipient before sending. They do not need to separately approve the retrieval of each document available to them. This distinction shortens the work and focuses attention on the decision that matters. It should result from process analysis, not implementation convenience.

Context and references: NIST: Generative AI Profile

Show the ready change, not a general intention

The user needs to see what will be added, changed, or sent. For a record, show the values before and after. For a message, show the recipient, subject, body, and attachments. Highlight missing information and elements requiring special attention. Do not make them reconstruct the entire task from the conversation history.

Demonstration example: an assistant proposes changing the delivery date. The approval view shows the previous date, the new date, the order, and the source of information. If the data changes after the preview is prepared, the previous consent should not unconditionally cover the new variant. The application must verify the currency of the approved operation.

Control requires time and accountability

Designate a person or role authorized to make the decision. Consider what happens during absence and how long the matter can wait. An approval queue without an owner becomes a new bottleneck. The user should be able to reject the proposal, correct it, or pass it on with justification.

Measure waiting time and reasons for rejection. If most proposals require the same correction, fix the source or rule. If approvals are done in bulk without reading, check for overload and preview readability. The mere fact of clicking is not proof of effective oversight.

Consent and execution must be linked

Record which operation was approved and whether it was executed. Do not create a second effect of the same consent upon retry. If execution fails, show the status and a safe next step. “Approved” and “sent” are different pieces of information that should not merge into a single success message.

In simple cases, it is possible to later reduce manual control, but this requires data on that class of tasks. Retain the ability to restore approval. A change in data quality or business rules may cause previously safe automation to require a human again.