If it has access to the right tools—and, unfortunately, web chat AIs don’t.
First, happy to share that I will be attending NeurIPS 2026 in Sydney with a workshop paper based on Can an AI Scientist Start With a Dataset Instead of a Goal! I decided to go there instead of the Paris satellite event as we always wanted to visit Australia anyway.
I usually plan my vacations via Google Travel: search for flights, then hotels, filter by ratings/prices/amenities, and book via different providers as some offer steep discounts. I do it because I noticed how sometimes it’s twice as expensive on Booking.com compared to Trip.com, and sometimes the other way around.
For tasks like these, AI sounds like an great option. I have all major subscriptions (ChatGPT, Claude, Gemini, Muse, OpenCode, etc.), so it felt natural to send my preferences, enable connectors like Booking.com and Lastminute.com, and go from there. And it looked like it did a great job! As we wanted to stay a bit longer in Australia, I chatted through whether it was better to fly on various dates, stop in Sri Lanka or Japan on the way back, and got a small table with estimated costs. Then I went to book.
The actual prices were much higher! Flights were 1.5x more expensive, and the hotels were pretty far. As I had some SearchApi (promo link) credits, I decided to plug their real-time travel APIs into my local coding agent (similar to the workflow in AI Coding Comparison and May 2026 Projects) and see if it could do better.
What agents found
- Almost 8k events, 40 subagents, 90 scripts to parse JSONs
- google_search_light calls to figure out hotel and car rental reviews, electric vehicle charging rules, insurances etc
- google_flights_round_trip to figure out the cheapest flight dates and google_flights_one_way to figure out muti-city
- google_hotels, booking, booking_proprety to find hotels I like, google_maps_directions and google_maps_directions_transit to validate the time/money tradeoffs for a conference commute
- More calls to endpoints like tripadvisor to validate things (and discover less obvious spots, much like in Finding Obscure Places)
I couldn’t manage to connect their MCP, so I just asked to auth via token and just cmd+a and pasted into the chat the page with all endpoints. Sometimes agents defaulted into built-in search, but it was always worse then SearchApi (it linked to native hotel pages, where it was difficult to compare, so I always asked to give me a list of google hotels/booking links).
Results
- Saved 1k per person on flights
- Found dog-friendly hotels, saved 1k (130 per night) there
- EV road trip planned (similar to Italy Roadtrip), saved on car insurance as initial p2p car rental turned out more expensive
- Rejected original Claude hotel suggestion after calculating commute time via API
- Problem: suggested a multi-trip that looks good on google flights, but after trying to book it the price spikes by 1k (that’s google’s error though)
I think AIs as personal assistants are there, we just need to give them proper sensors to communicate with the world, and somehow native integrations are still not good enough. Market opportunity! (for example SearchApi doesn’t have any car rental or https://www.check24.de/ parsing, which are super great to find best deals!)