The generative AI topic switched at the end of January almost exclusively to agents, with a sprinkling of when we can expect AGI. It’s the agents that will be the biggest disrupting force.
The initial conversation after the launch of ChatGPT was all around trip planning. That hasn’t gone away, but with the launch of practical agents, it’s the transactional piece of the journey that’s in play.
Let’s imagine you’ve done your research, and you’ve decided on a two-week vacation to California, a difficult choice to beat. In your head, you have lots of parameters, but you’re limited to searching for one product at a time. So, accommodations one location at a time and activities all individually. The search you have in your head is:
- Two weeks, from July 1-14
- Three to four nights in San Francisco – Alcatraz, Hop-On Hop-Off bus tour, those big trees, hotel near walking area. Warriors game. No car.
- Approximately two nights in Yosemite – hiking. Glamping or cabin ok. Stay in park ideal.
- Three to four nights in Los Angeles – hotel near beach. Splurge if sea view. Parking. Universal, Disney. Surf lesson.
- Budget = $200/night for hotels, Warriors $150
That’s a few hours manually. When you do it manually, a lot more parameters run through your head for each part, and you could easily add those to the prompt. At first glance, you might think you can get most of what you need from a major online travel agency or two. Throw in some metasearch and you’re there. So let’s look at what is different with an AI agent today.
Subscribe to our newsletter below
Run all searches in parallel – above is your prompt, add more details where you like. Mention loyalty, preferred standards, breakfast options, budget, whatever is in your head. It will research the entire trip with that single prompt.
Deep search — you can go as granular as you like. Think of all of those left column filters but also things you can’t filter for — the fuzzy stuff like “walking distance from restaurants.”
The OTA filters are great — they are just the most common parameters we all think about, but in reality, we have a lot of niche or personal preferences which don’t fit neatly on an OTA’s web page.
Metasearch needs APIs and contracts. AI agents will use those APIs, but can also browse products not available on OTAs, like Airbnb and local non-distributed products. As far as the AI agent is concerned, a website is now an API.
Dates are flexible. It can move things around. Start in San Francisco around a Warriors home game. Avoid a conference in L.A. because prices are double. This is the reasoning part, which you try to do in your head. The AI agent has no problem with a few thousand parameters.
Unlike metasearch it can check all room-types, options, inclusions, hidden fees etc. Before the final booking, it can negotiate. Think of the old “name your price” Priceline days. Via phone or live-chat the agent will say “Listen I found this same room for $180 elsewhere, give me your AAA, Costco or AI agent discount for $175 and we have a deal.” The agent on the other end will negotiate it through in 0.05 secs.
Once all this is complete, you’ll be presented with whatever you want. A hundred options on a spreadsheet, five choices with images for each product or one recommended option, just accept or decline. Or just book it all and send me an itinerary, along with a 15-minute video preview of my trip narrated by MrBeast.
What’s missing from where we are today?
- The large language model technology is there today, but it needs to be pieced together. Especially where it gets stuck. It would need a human in-the-loop constantly right now, but task by task, it will learn to be autonomous. Autonomous doesn’t mean it needs to go end-to-end every time. Humans still like options. It’s not the worst thing to have the humans make various choices or tweaks along the way.
- Logins, payments and privacy need to be solved. I would expect that to be a Google /Apple thing, but who knows where that will end up. Blockchainers will have some ideas.
- APIs are easy and efficient. Browsing is slow and expensive. There will be new global distribution system which are just an API of products and links to booking pages (like a Google Travel feed). These will constantly crawl and cache products, acting as the API for everything and giving a shortcut to much of the browsing required for each search. That’s pretty scary if you’re a current intermediary.
- Flexible booking. I imagine there might need to be more widespread booking conditions, where the AI Agent can easily lock in a booking and hold a price because it still needs final approval from its human. Today’s book > cancel is probably not a good hack to make that work.
Industry impact
Searching using APIs is easy and available today — an AI agent could even just use an OTA affiliate account. When you need to deep-search (information buried in descriptions) it gets harder because the APIs were not built for that.
When you add in browsing, it will probably get expensive in the short run. Then again, expensive might mean 1-2% of a transaction. Not close to the 10% each for an OTA and for Google.
Overall, you’d expect these agents to drive demand. Taking the hassle out of booking can only increase volume. On price, it has to be deflationary, and it has to squeeze the intermediaries. They are making money with services that are no longer required in many cases. Competition only drives prices lower.
It’s taken just two years to get from the launch of ChatGPT to this point. Next we have AGI. How will that change the industry? I don’t think anybody has any idea. But it’s certain that it will.
About the author…
