When an AI agent consistently fails to summarize support tickets, a crucial first step is to refine the prompting strategy. Ensure the prompt provides clear instructions, examples of good summaries, and specifies key information to extract, alongside evaluating the training data quality. Consider breaking down lengthy tickets into smaller, manageable chunks or implementing pre-processing steps to remove irrelevant conversational noise. Establishing a robust human feedback loop is essential, allowing agents to correct summaries and provide labels that reinforce learning. Additionally, exploring different AI models or adjusting model parameters like temperature might yield better results. Implement a fallback mechanism where human agents review failed summaries, and analyze common failure patterns to inform further adjustments.