AI Agent Test Purchase Light

An organisation puts an AI agent on its website, in a messenger or on a phone line. Inside, the agent was checked by its developer, the party paid to make it work. Outside, nobody checked it. AI Agent Test Purchase Light closes that gap. One demanding visitor goes through the agent in one continuous session and compares what the agent said with what the organisation publicly promised on its own site. The principle is simple: one black swan is enough to disprove that all swans are white. The method does not assess the agent's quality as a whole and does not monitor it; it looks for errors that matter to a consumer, and if it finds one, others are possible. This document sets out the whole method in the open: what counts as the reference and why the organisation's own site is enough; the five layers every check covers (who am I talking to, how to reach them, organisation and service, core function, boundaries and data); the rules of conduct, including the single permitted objection and the language test; what is never done, from emergency scenarios to extracting system instructions; the four states an item can receive and the four severity classes that decide what goes first in the report; what the organisation gets back; and the limits of the method, stated plainly. The questions themselves are selected for each organisation and are not published. Three pages.