The Multilingual Itinerary Engine: What a Tokyo AI Concierge Has to Get Right

Inbound arrivals into Japan are dominated by Chinese, Korean, Taiwanese and Hong Kong travellers, with English serving as the lingua franca for everyone else. A concierge that only speaks English addresses a minority of the opportunity. English, Simplified and Traditional Chinese, and Korean are the floor.

Language coverage also has to extend past the interface. A recommendation is only useful if the venue can actually serve that visitor, so the underlying data needs an English-menu or foreign-friendly flag, not just a translated name.

The engine should answer from a maintained corpus of Tokyo points of interest: category, ward, opening hours, price band, reservation requirement, accessibility, and whether a vehicle can realistically stop nearby. Retrieval-augmented generation over that corpus keeps recommendations grounded and auditable.

The last field matters more than people expect. Plenty of famous Tokyo destinations have no viable kerbside drop-off. An itinerary that ends at a coordinate a car cannot reach is a broken itinerary.

• Time budget: a three-hour evening is three hours, including transit between stops.

• Opening hours and last orders, resolved for the actual date, not the general case.

• Reservation lead time — some venues cannot be added same-day at all.

• Sequencing that respects geography, so the route does not cross the city twice.

• Weather and season, which reshape any outdoor stop in Tokyo's summer and winter.

The valuable artefact is not the paragraph describing the evening. It is the ordered list of pickup and drop-off coordinates with timestamps, which is what gets handed to a dispatch layer. Prose is the interface; structured route data is the product.

This is a design brief for an early-stage build, not a shipped feature. If you are building in this space, or want the asset and the framework, email andrew.mc@nousdomains.com.