Why AI trip planners get New Zealand wrong
Ask a general-purpose AI to plan ten days in New Zealand and you will usually get something that reads beautifully. Wellington's waterfront on day three. Nelson and the Abel Tasman on day four. A tidy loop, a confident tone, and a problem any advisor who has made the trip spots straight away.
Between Wellington and the top of the South Island is Cook Strait. On a map it is a short hop. In practice it is a three-and-a-half-hour ferry or a flight, with check-in and transfers at both ends. Call it most of a day. The draft treats it as a morning drive.
Distance is not travel time
Most itinerary tools, and almost every AI that writes trips from memory, reason about geography the way a map does: close together means easy to combine. That holds in parts of the world. In New Zealand, Australia and across the South Pacific it fails constantly:
- Islands. Two hubs can be an hour apart by air and a full day apart in practice, depending on which side of the water you start.
- Gateways. Lord Howe Island and Kangaroo Island are reached through one specific mainland hub. Route a client there from anywhere else and you have invented a connection.
- Spokes. The West Coast of the South Island has no practical airport of its own. You get there via Christchurch, and the draft should say so.
None of this is obscure. Every advisor who sells the region knows it. The problem is that it isn't written down anywhere an AI can reliably learn it, so the model fills the gap with the shape of the map.
Write the rules down once
The fix isn't a smarter model. It's putting what advisors already know into the system as rules and letting the software follow them. In Cordovan, every New Zealand hub must record which island it is on. Each hub lists the places it genuinely pairs with, and how: as a gateway, before, after, or as an extension. A route can only move along those links. A hub that is slow to reach costs part of the trip's night budget, so a seven-night trip isn't spread across six stops.
The result is less impressive to watch than a chatbot writing paragraphs. It is also far less likely to send your client to the wrong ferry terminal.
A draft should save you time. One you have to fact-check line by line doesn't.
What to ask of any itinerary tool
If you are evaluating AI itinerary software for your practice, three questions sort the useful from the merely fluent:
- Where do the recommendations come from? A vetted catalog, or the model's memory?
- How does it decide two places belong together? Rules from people who have travelled the route, or proximity on a map?
- What happens when it doesn't know? Does it flag the gap for you, or fill it with something plausible?
We wrote up how Cordovan answers all three, failure by failure, in how our itinerary drafts work.