Planning a family trip can be daunting. Parents and guardians must balance a myriad of preferences across eating, sleeping, and entertainment preferences all while maintaining a budget. A wrong turn or a small mishap can cascade into major vacation disruptions and undermine the best plans and budgets. I recently faced these important analytics challenges while planning a 9-night family trip to Seattle and Vancouver. Looking across the expanse of possibilities, I reached out to friends and colleagues for ideas. Next, it occurred to me leverage generative AI in the form of ChatGPT to help me sort through and organize my options. The following post explains how ChatGPT planned my family vacation.

Tourist and Visit Information from the Locals

Two of six people responded to my request for ideas in time for my advanced planning. Two people responded just in time for us to make time to visit. The combined input of the two timely “respondents” produced a raw list of things to-do and see. I excluded restaurants since meal choices would be highly dependent on availability of options and convenience to the itinerary. I also filtered out entertainment ideas that I knew would not work for my family. Here is the raw itinerary almost verbatim:

• Bloedel Reserve
• Visit several of these Native American tourist locations (you could start with the location mentioned in Pike’s Place)
• If you get a chance, highly recommend traveling up to Vancouver, Canada (will need ID).  While there, highly recommend checking out some of the Native American tourist destinations 9 Unforgettable First Nations Experiences in Vancouver
• Should visit the historic African American community known as the Central District
• Along with the Central District – should also drive the length of Rainer Ave. in Seattle
• Should visit the famous Black high school Garfield High.  Attended by Quincy Jones, Jimi Hendrix, Ernestine Anderson and others Garfield High School (Seattle)
• Should also visit Franklin High Franklin High School (Seattle)
• Should visit the Technology Access Foundation (TAF) Home – TAF
• If you like Jazz or R&B – try to catch a show at Jazz Alley.  The food there is also decent.
• Should definitely take at least 1 or 2 Ferry Rides.  Start with the Brainbridge Ferry to see the Bloedel Reserve and then consider the Ferry to visit the San Juan Island cluster
• Should definitely check out Amazon and Microsoft HQ’s
• I believe Starbucks also still offers a tour (they have an exhibit in there HQ)
• Should visit the Gates Foundation (they are right across from the Space Needle and also close to Amazon HQ)
• Of course visit the normal tourist spots such as Pikes Place, Space Needle, some of the Aquariums, etc..  (To help get acquainted with the layout – taking a Bus Tour is not a bad idea)
• Seattle has a lot of good, niche museums – for example Microsoft has a museum (highly recommended), there is an Aeronautics museum, and others
• Columbia Tower observation deck
• MoPop/Pacific Science Center
• Gasworks park (this place is amazing for the 4th of July)
• Golden Gardens (beautiful and fun sand beach in Seattle)
• Marymoor park (has concerts and movies during the summer; is about 1 hour outside of Seattle)

ChatGPT Plans the Vacation

Logistics presents the biggest challenge for processing all these ideas, especially without knowing relative locations. A blank slate stares like a daunting expanse of permutations in time and space. I turned to ChatGPT for a solution. I braced for an extended conversation to get ChatGPT to understand my objectives. However, I was pleasantly surprised at how well the large language model handled specific constraints and preferences. ChatGPT even reached beyond my list with attractive suggestions that filled space and interesting alternatives that generated new ideas, especially for Vancouver where my raw list included fewer specifics.

My prompt engineering provided specifics of time and place, but I purposely left out specific dates. I did not want ChatGPT to over-interpret the importance of the dates given I had some flexibility in sequencing (in retrospect, I should have given ChatGPT the opportunity to wrestle with the additional information). I tossed in a request for dinner recommendations just to see what ChatGPT would recommend.

Note I used ChatGPT-4 with chat history turned off.

I was absolutely stunned by ChatGPT’s recommendations from my first try! Not only did ChatGPT meet all constraints, but it also organized the days by themes which gave the itinerary an extra layer of purpose. I ran a second pass to get an alternative to the lofty day of heights in Vancouver. My raw list did not include a visit to Victoria, a place that was at the top of my priorities for the entire trip.

ChatGPT prompt (what I typed into ChatGPT as my request): “Provide a travel recommendation using the list below of Seattle and Vancouver tourist spots. I want 5 consecutive days in Seattle. On the 5th day we will leave Seattle by 11am to drive to Vancouver. I want another 4 consecutive days in Vancouver. On the 4th day in Vancouver we will drive back to Seattle to catch our evening flight back home. Provide start and end times for each recommendation. Start no day earlier than 10am and end no day later than 7pm on each day of the itinerary. Minimize the time spent traveling between recommendations. Recommend a popular restaurant to visit at the end of each day using the given list.”

ChatGPT recommended the below itinerary. I added Google map directions for each day as a way to review the quality and practicality of the routing. The only route I disliked was Day 2 with a visit to Microsoft requiring a side trip. I also made a small adjustment to the sequence for the trip to Victoria (see the attached note).

Human Judgement Takes Over

I was so stunned by the ChatGPT results, I almost thought my job was done! The daily routing worked out well. The time constraints worked. The restaurant recommendations looked appetizing. Alas, once I checked into the details of various stops, I realized my raw list included several hidden constraints. Some places are only open certain days. Some places do not offer tours. Most importantly, (Air B&B) accommodations turned out to be far more expensive than expected. Vancouver was even more exorbitant than Seattle. To reduce costs, I shifted a day out of Vancouver and added a “free” day in Seattle. That free day came in handy for visiting friends.

ChatGPT’s restaurant recommendations were appetizing and some were stunning. Accordingly, they were also quite pricey, especially for a family of course with two kids. Every proposal needed a downgrade!

The Human Augments the AI’s Travel Plans

At this point, I could have applied some additional prompt engineering to get an itinerary aligned with more constraints. For example, I could have asked ChatGPT to downgrade the restaurant selections. However, I figured ChatGPT’s initial results provided the majority of the value of the exercise. It was time for human judgement to take the wheel. The Google maps daily routes became an important tool for making manual adjustments. I used the maps for adding new stops and removing others. By looking for nearby restaurants, I found plenty of great culinary options that fit the budget and the family palates.

After some extended work (with several check-ins with my wife), I developed the final plan. I also added stops at Pink Gorilla games for my gaming son. I made more adjustments in “real-time” as the normal contingencies of travel occurred. We also jumped at unexpected opportunities (some free!). Annotations indicate places where we made decisions on the fly. A few instructions provided important reminders for various stops. I left out the Google maps routing because they included the addresses of our Air B&Bs. I did my best to incorporate thematic labels for each day.

The augmented itinerary experienced several iterations. The annotations do not convey the full experience of real-time decision-making, negotiating periodic fatigue, occasional boredom, and magical shopping experiences. The main point here is that I did NOT use ChatGPT as my running travel buddy. I used it to lay the foundations for planning logistics and built from there. Once the trip was underway, I needed access to up-to-date information, driving directions with real-time travel times, transit information, and the latest restaurant ratings. All these travel supports come from the familiar tools accessed by Google searches and Yelp.

Epilogue

I could have called this post “How Google Maps Planned My Family Vacation.” Google Maps was a critical component of my augmentation process. Yet, ChatGPT gets the marquee as the more novel technology and product. Many people are still learning how to effectively use ChatGPT for everyday needs. Moreover, ChatGPT built the foundation that saved me a lot of time in planning.

ChatGPT impressed me as a tool for travel logistics. I did not know what to expect after a prior analytics experience using ChatGPT to schedule OBGYN hospitalists failed miserably. That exercise required complex calculations and an ability to apply optimizing algorithms, a big stretch for a language model. In fact, ChatGPT recommended several times that I seek counsel with scheduling software! However, travel planning apparently uses simple enough math that can be gleaned from understanding the language used in related travel and tourism documents.

In future travel planning, I want to push this analytics exercise as far as ChatGPT will let me. For example, I will provide a comprehensive list of requirements and preferences as well as ask for sources. If you decide to try out ChatGPT for your next trip, I would love to hear about your experiences and lessons learned!