Practical AI
Decide what AI must hand back to a person.
Define the conditions that stop an AI-assisted workflow and return the next decision to an accountable person.
In a hypothetical distribution business, an AI-assisted workflow summarises supplier exceptions for the operations team. Most records are routine. One contains conflicting delivery dates and a request to change payment details. The summary is fluent, but the next step is not an ordinary processing decision. Before automating a task, decide which situations the workflow must hand back to a person. A good hand-back rule protects important decisions without turning every case into manual work.
Separate assistance from authority
Write down what the AI may prepare, classify or recommend, and what remains a human decision. Use the actual business task rather than broad language such as ‘human in the loop’. The useful question is who can approve the next action when the case is unusual or consequential.
In the hypothetical supplier workflow, the AI might group routine delivery issues and draft a summary. It should not approve amended payment details or resolve conflicting commercial information. Those actions need the person with the relevant authority and access to the approved source records.
Define observable hand-back triggers
Choose conditions the workflow can identify and the team can test. Useful triggers include missing required information, conflicting sources, an action outside an agreed limit, a request involving sensitive details, or an output that cannot be supported by the approved evidence.
Avoid relying on a vague confidence score alone. A high score does not make an unauthorised action acceptable, and a low score may not explain what the reviewer needs to inspect. Connect each trigger to a clear reason that can be shown to the receiving person.
- What condition stopped the workflow?
- What information has already been checked?
- Which decision or action is still required?
- Who is authorised to take it?
Make the hand-back usable
Send the case to an agreed queue or owner with its reference, the stopping reason and links to the relevant approved sources. Preserve the work that can safely be reused, but do not present an incomplete action as completed. The receiver should be able to distinguish a new exception from a repeated notification.
Set an expected response route for time-sensitive cases. If nobody is available, the safe fallback may be to pause, continue through the existing manual process or escalate to a named role. Do not let the workflow quietly choose a different action because the preferred reviewer is unavailable.
Test the boundary, not only the normal case
Create synthetic examples for an ordinary case, missing information, conflicting evidence and a consequential request. Check that routine work continues, each stop occurs for the intended reason and the right person receives enough context to decide what happens next.
Review the exception record after real use. Repeated hand-backs may expose unclear guidance, poor source information or a process that needs redesign. They are not automatically evidence that the AI should receive more authority.
Rethink's AI review and enablement work can help define these boundaries, test the workflow and turn the result into practical operating guidance.