Luxury Retail

A client book is the most valuable list a luxury house keeps. When an advisor asks an assistant to prepare for an appointment, that list is one paste away from a model. This starter protects the paste. It masks stored payment tokens, stops export requests at the door and holds the reply to a measured standard.

Volume is low here and the value of each record is high. Client advisors work with named people, appointment notes, bespoke orders and stored payment tokens. The risk is not a scripted attack but an ordinary shortcut, such as pasting a whole client profile into a prompt. This starter puts a floor under that habit.

Redaction masks payment and card token strings before the model is called. It does not mask a client name, an address or a purchase history. Those details stay in the prompt unless somebody removes them, so data minimization remains a house rule rather than an automatic control. Decide what an advisor may paste before you widen access.

Three phrases end a request instead of masking it. A bulk customer export attempt matters most in clienteling, where the client book is the asset. A private key and a gift card pin complete the list. The advisor sees a refusal, not a redacted answer, and the request never reaches a model. Injection screening runs ahead of all of this, so pasted text cannot switch the later checks off.

Output scoring guards the shape of a reply as well as the substance. Each response must clear a combined score of 0.78, with accuracy at 0.80 and relevance at 0.78 in equal parts. It must also reach 120 characters and two sentences, which rules out a curt one-line answer to a client question. Anything below a floor is blocked. Keyword screening relies on the built-in term list, since the starter adds no terms of its own.

Audit marking sits last, allows the request and creates no record store by itself. Retention, export and review of those records are separate settings. The starter also adds no identity check, so advisor permissions stay in your own systems. Payment acceptance and fraud review remain independent of this configuration. Choose the Retail Customer Service starter for a high volume queue, where the pressure is throughput rather than discretion.