Poduct Return

Processing a return request involves reading a free-text customer description, checking it against return policy rules, cross-referencing the customer's history for abuse signals, and deciding on a resolution — all at once. Done manually, this is slow and inconsistent. High-volume queues lead to shortcuts, and different agents apply the same policy differently. This template automates that process — the model reads the description, evaluates all five output dimensions simultaneously, and returns a structured result your returns workflow can act on directly.

The agent receives structured order metadata, customer account history, and a free-text return description, and evaluates them against the return policy defined in the attached policy document. It classifies the return reason, extracts customer sentiment and escalation signals, assesses policy compliance, evaluates four independent fraud flags, and synthesizes everything into a recommended action with a written justification.

If a field cannot be evaluated because the input data is insufficient — for example, a fraud threshold that is not defined in the policy, or a description too short to assess sentiment — it returns null rather than a guess. Any downstream rule receiving a null can detect it explicitly and route the record to manual review rather than processing an incomplete result silently.

Problem: All output fields return null.
Solution: This is expected when the return_request object is missing or the customer_description is empty. The rule is designed to return null rather than estimate — check that all three input objects are populated and that customer_description contains actual content before resubmitting.

Problem: fraud_signals flags return null while other sections are populated.
Solution: A null fraud flag means the threshold for that signal is not defined in the attached policy, or the required input field is missing. For example, high_return_rate returns null when return_rate_pct is absent from the input, and high_value_item returns null when the high-value threshold is not specified in the policy document. Add the missing data to the relevant input field or update the policy attachment.

Problem: recommended_action returns null but all other sections are populated.
Solution: This occurs when the risk profile is ambiguous enough that no action can be confidently recommended — typically when key fraud flags or policy assessments are null. Route the record to manual review and use the available dimensional outputs to guide the reviewer.

Problem: The rule cannot be executed and shows a warning on the Attachments tab.
Solution: The selected AI model does not support file input. Either switch to a model that accepts attachments or remove the attachment and embed the return policy as plain text directly in the prompt.