If unsure, ask. Never guess. — AI Agent Pre-Execution Checklist

One important point: When AI produces a wrong result, the discussion often goes back to model capability. Regulation and safety discussions also tend to focus on the model. But this document separates responsibility across the execution structure. If the Provider does not supply the necessary Checklist for an Action, the Provider’s scope of responsibility becomes relevant. If the user instruction is incomplete, the system should not treat it as complete. If the Agent executes without confirming unresolved unknowns, that is not a model error. It is a pre-execution failure. A better model may fill the blanks more convincingly — but it does not replace the structure that checks and stops at those blanks. AI is not the source of authority. It is an execution layer that performs the verification process. Final authority must remain with the user. One of the core challenges in AI regulation is that technology moves faster than law. By the time a rule is written, the technology may have already moved to the next stage. This document proposes a way around that problem. The Provider supplies the question list for each Action. The Agent checks whether the answers exist. If not, it asks the user. If still unanswered, it does not execute. With this structure, regulators do not need to follow every technical implementation. They can define standards — what Checklist items must be included for each industry or Action type. Law and regulation can then review after the fact whether the structure was sufficient and whether the Agent followed the verification process. The execution structure outside the model can be changed now. Responsibility is not only examined after an incident. It is declared inside the structure before execution.