The answer is not the only output that matters when an AI agent queries a database. The system also needs evidence. What data was touched? Which scope was applied? How many rows came back? Was the result truncated? Was the schema context current? Did the agent summarize raw rows or approved aggregates? If that information disappears before the final response, the answer becomes hard to trust and harder to debug. That is why AI database tools need result contracts. Raw rows are not enough A database tool can return rows and let the model summarize them. That works for demos. In production, raw rows alone leave too much ambiguity: Was a row limit applied? Did the query time out? Were some columns redacted? Was tenant scope enforced? Which metric definition was used? Was the result fresh enough for the question? The model may produce a confident summary while important caveats are lost.…