Silence Rather Than Error

What does it mean for a customer-facing AI agent to answer correctly? In August 2026, Sergei Ponomarev and Andrey Ekhmenin, founder of ANDEKS, examined the same live public support agent through two independently applied approaches. Each froze his findings before seeing the other's. Ponomarev compared the company's public promises with the customer's experience through a test purchase; Ekhmenin assessed which conclusions the available evidence could actually support. Neither approach identified incorrect information in the answers examined. But questions about restrictions, refunds and personal data exposed a different problem: the agent did not convey conditions its own company had already published. The paper distinguishes a legitimate admission of uncertainty from silence about an available public rule. It also explains where the approaches differed over response speed, round-the-clock availability and the source of an answer, and proposes assessing factual correctness, absence of invention, completeness of relevant published rules and resolution of the customer's question separately. The six-page article sets out the findings, practical implications and evidence limitations, including incomplete preservation of the test-purchase recordings and limits on session independence. This is a study of one agent in specific scenarios, not a certification, legal opinion or general quality rating.