An AI suggestion and an AI action carry different responsibilities. Define the task, the permitted actions and the human role before allowing a workflow to change records or make commitments.
A helpful answer is one kind of result
A customer asks for a change. AI reads the request and drafts a helpful response. A person checks it and decides what to send.
Now imagine the same system also updates the account and makes a commitment. The capability has crossed a boundary. What was assistance has become delegated authority.
Separate the kinds of permission
Reading information, suggesting a change and applying it are different actions. Give each a clear place in the workflow.
What may the system read? What may it propose? What may it change without approval? Those questions make the delegation understandable.
Make the human role practical
A reviewer needs the original request, the relevant context and a clear account of the proposed action. They also need a way to reject, escalate or stop it.
A person can only exercise useful judgment if the workflow gives them enough information and authority to do so.
Try an exception before calling it ready
Consider a request with missing information, conflicting instructions or an action outside the approved scope. What happens next?
Who notices? Can someone inspect the result and recover? A convincing demonstration becomes more useful when you also understand what happens at its boundaries.
Use the technology to rethink the work
The interesting opportunity reaches beyond a faster version of the same process. Which steps still serve a purpose? Where could people spend more attention?
Start with the work you want to improve. Then decide what role AI should have, what authority it needs and which decisions remain with a person.
IDEAS TO TAKE WITH YOU
- Name the task and the acceptable result.
- Separate information, suggestions and actions.
- Give a person the context and authority to intervene.
Which decision would you want a person to understand before a system acts on your behalf?
Explore the background.
These references offer context for the questions in this essay. The perspective above is personal.
- NIST AI Risk Management Framework
Background on considering trustworthiness throughout the design, use and evaluation of AI systems.
Anas Kamlani
Anas Kamlani hosts a conversation across technology, business, ownership and the places we build our lives.
The story behind this spaceWhich part stayed with you? Bring a question, a lived experience or a different way of seeing it.