AI, AI, Oh…
It was the summer of 2023, and everybody was getting excited about the robots. Skyscanner was no different.
We threw ourselves into the world of AI by designing our first Large Language Model-fuelled proof-of-concept... A destination search feature! What else?
What’s everyone else doing?
It all began with good old competitor analysis.
After snooping out other Large Language Models (LLMs) in the travel space, we found that most were using AI to level-up their chatbot experiences. Some were even developing whole avatars and fresh GUIs for users to interact with.
On the other hand, we wanted to be thoughtful and intentional with our first foray into AI. Analysing what else was on offer in our space showed us what we didn’t necessarily want to do.
We decided to keep things tightly scoped.
2. Ask and ye shall receive
My lead designer and I co-ran some feasibility and desirability workshops with our engineering colleagues to align on a solid direction for our PoC.
The winning concept was a powerful, yet intuitive search experience, allowing users to ask anything travel related and then receive destination suggestions.
Cue a hectic week of wireframing, design reviews, and iteration between my co-designer and myself. Content-wise, I wanted our UI to combine the aspirational, dreamy aspect of travel with the functional efficiency of AI.
3. Working promptly
I worked with our engineering team to sculpt the first iteration of our prompt (first image, in black). We put together the guard rails and foundations for the LLM to follow, mainly covering issues of comprehension and style (formatting, verbiage, length of output etc.) I embedded our content design guidelines at this early stage.
However, our first tests of the prompt showed that these initial guidelines were not enough. It still needed more of a human touch.
It was time for me to pull out the raw text of our prompt and iterate, iterate, iterate in a word document to ensure we were covering all potential roadblocks. The primary concern was that the feature should not recommend travellers visit locations that were dangerous or ill-advisable for tourism.
4. Playing in the sandbox
After many revisions of the prompt guard rails, including multiple cross-collaborative meetings with our SEO, legal, marketing, brand, and engineering teams, I was finally able to wrangle the LLM’s output.
It was now giving users short, readable destination summaries written in the Skyscanner tone of voice, and able to turn (almost) any input from into useful travel insights.
The prompt was honed, the UI was marked up, and the team was feeling good. Now it was time for my bread and butter: UX writing.
To tie up the loose ends and get the PoC to handover, I ran through a number of unhappy paths that might occur during users’ experimenting with the feature.
These ranged from standard error messaging, empty states, no-results copy, right through to advice on how to use the feature. A free text input field is asking for trouble—especially when AI is being so rigorously scrutinised by the media.
I worked closely with User Research to compile a list of potential bad faith searches, and wrote effective messaging for each.
5. Business as usual
6. Getting Savvy
Within weeks, the “Dream and discover with AI” PoC reached 6 million users without any extensive marketing or push.
More crucially, we encountered zero instances of inappropriate or insensitive content thanks to the rigorous iterations in the prompt guidelines.
This early attempt at Skyscanner AI was so successful that we quickly decided to refine it. The functionality was ported to the app, as we noticed stronger user intention on smaller screens.
We took this opportunity to rebrand the feature, now called Savvy Search (all credit due to our wonderful brand designers and copywriters).
Savvy Search currently sits proudly on the Skyscanner app homepage. Why not try it out, now that you’ve seen what lies behind the scenes?
“Your focus on making things better for travellers and your ability to pull off such a great outcome in such a short time is seriously impressive.
I’m happy to have worked with you on building this product! Thank you!”
“Phil has been excellent partner on the AI PoC project. His content designs are quality, well thought through and made perfect sense to users and stakeholders. This project was quite a learning curve for us all and Phil did a great job of fine tuning all the edge cases.”