An AI payment agent can finish a task and still leave the merchant with more work. I would therefore evaluate the work left for merchants, not just whether the agent completed its task.
For a Razorpay case study, I would compare merchants using the assistant with merchants using their existing process. That baseline should include the tools and automation they already use.
Use comparable tasks and the same authority limits in both workflows. I would include ordinary requests as well as exceptions, so an agent cannot look successful by refusing everything difficult.
Keep task completion as an operational metric. Add decision quality, review and correction effort, unnecessary escalation, and cost per correct resolution.
The approval-card concept and synthetic scenarios for this case are proposed research artifacts. I have not tested Razorpay's product or run a merchant study, so these ideas are not evidence of product performance or merchant-study results.
A useful evaluation should show the business trade-offs clearly enough to support a decision, including keeping the existing process.