Consider an assistant asked to prepare a purchase request. It finds a suitable supplier and completes the paperwork. The output looks useful. But the organization still needs to know whether the assistant was allowed to submit the request, approve a payment, or change the terms.
These are questions about authority and responsibility. They belong inside the design of the workflow.
Make permission specific
We begin by separating the actions involved. Reading a document, drafting a recommendation, sending a message, and committing funds each carry different consequences. A broad instruction to “handle it” leaves those boundaries unresolved.
A useful design question is: what can the system do on its own, under which conditions, and when does it need a person’s approval?
Make the sequence reviewable
After an action, someone should be able to examine what was requested, which permissions applied, and what actually happened. The record needs enough context to support a meaningful review.
A record alone cannot establish that an action was appropriate. It gives people a basis for investigation, correction, and learning. The quality of that evidence matters as much as its existence.
Give responsibility a place to land
Who can interrupt the workflow? Who responds when the outcome is disputed? Who decides whether the system’s authority should change? Answering these questions helps connect technical behavior to organizational responsibility.
If this action is challenged tomorrow, what can we show, and who can respond?
Agent Proof explores these relationships between permission, activity, and accountability. Our aim is to make them easier for people to examine in the work itself.
Further reading: NIST’s AI Risk Management Framework offers a broader, voluntary framework for considering trustworthiness throughout the design, development, use, and evaluation of AI systems.