Everything the pack knows lives in one Google Sheets workbook it creates for you, named after your shop and ending "Commerce Data". Seven tabs: Customers, Orders, Items, Policies, Verification, Cases and Log. You fill in the first four with your own data; the fifth is where you set the rules. The Verification tab is the point of this pack. It has four settings, read fresh on every request: - required_checks - how many things a customer must get right. Two by default. - accepted_checks - which ones count: the order number, the delivery postcode, the last digits of the phone number on the account. - reveal_on_pass - the list of order fields the agent is allowed to state once somebody passes. Anything not on that list stays inside the gate even though it is sitting in the sheet. - on_fail - whether a failed check may be offered a human, or simply refused. Raise required_checks to three and the agent starts asking for more before it will say anything. Remove a field from reveal_on_pass and no agent can mention it again. Neither is a prompt edit, and neither is left to a model's judgement. The agent itself holds no access to your workbook at all. Policy text, identity checks and case opening all run inside a guarded workflow, and what comes back is either a masked summary or a refusal. That is why "this agent cannot read your customer list" is a fact about how it is built rather than a promise about how it behaves. A failed check tells the customer one thing, in one wording, whether the account exists or the details did not match - so the chat window cannot be used to find out which email addresses you hold. When something needs a person, the agent opens a case and emails your support address. Every case carries a key built from the customer, the order and the kind of problem, so the same problem reported twice returns the first case number instead of opening a second one. Once a case is closed, the same problem coming back counts as new.