Tiffany: remembering a client does not mean reciting everything you know

There is a narrow gap between 'How lovely that you remember me' and 'How exactly do you know that?' An expensive purchase makes the distinction particularly visible. A customer expects attention, not the sensation that somebody has ceremoniously opened a database folder alongside the jewellery box.

Tiffany offers appointments with client advisers. That is a useful setting for thinking about preparation and tact. What follows is my AI training proposal, not a description of Tiffany's internal technology or a project I have delivered for the company.

Rehearse what is appropriate

I would give an employee a short fictional history: a customer explored a category, agreed to receive an email and mentioned a convenient month to visit. The task is to prepare an invitation without adding assumptions about relationships, occasions or wealth. An empty CRM field sometimes deserves to remain empty.

The second mechanism is a role-play in which expectations change. The customer now wants something different, or does not want to discuss a budget. AI lets the adviser practise adapting without making the visitor feel guilty about the preparation already completed. Customers are not obliged to validate our homework.

A third exercise covers follow-up. The associate proposes a next contact, confirms the channel and purpose, and records only what was agreed. The trainer checks whether the message has a concrete benefit: answering a question, confirming availability or arranging a visit. Merely announcing continued existence is not always sufficient.

Practise a request to stop contact about a particular topic. The adviser acknowledges it and updates the permission correctly. An elegant final message must not become one last attempt to work around a preference the customer has already made perfectly clear.

Discretion is not a volume of notes

The three measures would be correct handling of communication permissions, independent ratings of message appropriateness and the proportion of agreed next steps that actually happen. That final measure matters: polite assent and genuine interest are not identical. Complaints about intrusive contact should pause expansion, not disappear beneath a conversion chart.

Training records should be fictional. Access to real client information would require a separate project with retention limits and role-based permissions. The model should neither send invitations automatically nor enrich profiles with guesses about private life. A person reviews the content and decides whether to send it.

My pilot would compare performance across three consistent situations: preparation, a changed request and follow-up. Senior advisers would score examples using an agreed rubric. Otherwise one prefers elaborate messages, another values brevity, and the AI receives contradictory lessons before lunchtime.

Useful AI in clienteling should teach restraint as much as fluency. It helps an adviser prepare without turning preparation into a performance of omniscience. Attention feels valuable when the customer is comfortable, not when the brand demonstrates every detail its memory can retrieve.