A retail company deployed a Copilot Studio customer service agent three months ago. Post-launch telemetry in the analytics dashboard shows a high volume of sessions, an escalation-to-human rate of 62%, and a session outcome breakdown where most conversations end with 'abandoned' status shortly after the agent responds to product-return questions. Knowledge sources include a returns policy document and a Dataverse case table. As the solutions architect responsible for tuning, what should you prioritize first to improve agent performance?