Fnac: a novel and a tablet should not receive the same sales pitch

A novel may deserve a recommendation because it leaves a peculiar aftertaste. A tablet probably should not. I like the meeting of culture and technology in one assortment, but it exposes a problem: an assistant applying the same reasoning to every purchase can become impressively unconvincing across several departments at once.

Fnac materials include selections and recommendations from its advisers. That is the documented context for my independent AI sales proposal. I am neither claiming that this workflow is deployed nor suggesting that a language model should replace somebody's cultivated taste.

Two modes of one conversation

For a book, the assistant asks what someone enjoyed in recent reading: pace, language, atmosphere or subject, not merely genre. It suggests a few directions and explains the connection. It must not invent plots, quotations or awards. Where information is limited, publisher descriptions and an identified adviser's opinion are enough.

For a tablet, the questions change to tasks, applications, budget, existing accessories and compatibility requirements. 'You might enjoy it' is no longer sufficient. The recommendation needs verified specifications for the actual model. AI must distinguish a preference from a claim the customer will later test with a cable.

An interesting overlap appears when somebody wants a device for reading. The assistant combines the modes: discuss habits, then check display, formats and restrictions. It cannot borrow a function from a similar device. Sharing shelf space does not give products a shared technical biography, however convenient that would be for the comparison table.

Taste is not a fact

Separate editorial judgement, publisher information and technical specifications in the knowledge base. Each needs a different presentation. Recommendations are described as opinions; capabilities as checkable facts. A commercial ranking must not disguise itself as an intimate understanding of what is perfect for the customer.

A pilot could cover one book section and a few reading devices. Advisers review recommendations and flag where the model too confidently reduces a person to one established interest. Customers are allowed to want something completely different from their last purchase. That is human freedom, not a broken profile.

Evaluation needs both accuracy and perceived usefulness: recommendations without invented facts, voluntary customer ratings of the explanation's usefulness, and device returns caused by a mismatch with the stated task. A delayed book purchase is not automatically a failed recommendation. People may save advice and return later rather than conform to a reporting period.

Useful AI in this setting changes its level of certainty, not just its vocabulary. With a novel, it opens a possibility. With a device, it verifies a condition. The associate remains responsible for the conversation. Otherwise we build a highly educated machine that recommends fiction as though it were choosing the amount of memory in a laptop.