Start with the process, not the model
Document the current workflow, the people involved, the systems touched, the transaction volume and the point where work becomes slow, repetitive or error-prone. Discovery should start from observable process facts, not a preferred AI product.
Identify decisions and exceptions
Separate deterministic rules from judgement calls. Identify mandatory approvals, common exceptions, escalation paths and the action the system should take when confidence is too low to proceed safely.
Bring sanitized real-world inputs
A small set of appropriately sanitized forms, documents, or data samples can expose integration and quality constraints early. Sensitive information should only be shared through an approved handling route.