easyJet: Rapid chatbot prototyping for an AI Flight Attendant
Over the years, I’ve worked with numerous airlines and flight affiliated products and services, and I can honestly say they always offer the opportunity to write copy that's full of fun and anticipation. When your copywriting is there to capture everything great about travelling abroad, it's often very easy to get caught up in other people’s travel experiences — whether it's the sun, the sand, the last minute holiday preparations, or pre-flight duty free shopping.
That's why I jumped at the chance to create a rapid prototype chatbot solution for easyJet. It offered the opportunity to frame this subject matter within a dynamic conversation design context by harnessing improvisational comms and a cutting edge language model. The core focus of this demonstration was to illustrate how quickly a brand-ready conversational AI could be deployed for easyJet in the form of a virtual assistant designed to help air travellers — from their point of arrival at the airport, to the moment they landed at their destination.
Crafting an easy persona
Before making a start on our technical proof of concept, it was first important to define an overall persona — and a job role. We decided to frame our AI as an extension of easyJet's flight crew, a digital Flight Attendant complemented with a comprehensive knowledge base, as well as live service information and customer booking info.
To help inform the chatbot's look and feel, I turned to some of my previous work for the airline. Having already worked on easyJet's Covid-19 landing page, it's critically acclaimed Look and Book mobile app and its World of Orange (WOO) design system, I already had a good idea as to how this new Flight Attendant should present itself. The design system in particular, complete with an accompanying ToV informing UX copy guidelines, served as a great foundational resource on which to build the bot's persona and company role.
Using a persona canvas, I then started to flesh out the some of the chatbot's prescribed characteristics, providing it with just a little bit of quirkiness to complement the levity and colour of the easyJet brand. It was a careful balance, with the temperature was dialled up on attributes like 'warmth' and 'casuality', offset by a robust sense of responsibility and attentiveness to customer's needs above all else.
I also saw the opportunity to ensure the Flight Attendant's function was aligned with easyJet's existing corporate persona. Having already been in involved in the crafting the airlines brand Tone of Voice, it meant I had working knowledge of how the company spoke across its digital domain — with ready access to a comprehensive set of ToV guidelines encompassed by four key principles: Optimistic, Bold, Playful and Energising.
Using a ChatGPT instance I went through these principles one by one, incrementally refining a prompt designed to yield AI output that was both nicely conversational while emulating each specific sentiment. It’s all very well to tell a language model to be 'optimistic', but in order for it to output a specific intepretation of optimism within a given context it’s important fine-tune a specific prompt component with some qualifying statements and working examples. You can see below how calibrating a prompt with just a little nuance in the desired direction can inflect the system response with phrasing that feels smart, sharper, and more on-brand.
Harnessing the power of Voiceflow
The tech landscape specific to conversational AI is constantly shifting. One constant that has held its ground over the past few years has been Voiceflow — a slick, adaptive chatbot platform that allows for plenty of creative innovation. Voiceflow continues to go from strength to strength, catering for seasoned developers and conversation designers, to novices experimenting with CxD for the first time. Founded on the principles of accessibility and user-centric design, Voiceflow sports an intuitive visual interface with drag-and-drop functionality and it makes it easy to visualise, and course-correct conversation flows without the need for extensive coding knowledge:
Voiceflow boasts, I think, the best design and prototype solutions as well as and fantastic compatibility — with an ability to deploy conversational agents across multiple channels, including Amazon Alexa, Google Assistant, and web-based chatbots. For that reason, Voiceflow was my go-to platform upon first receiving the easyJet brief and it gave me the opportunity to really get under the hood of this versatile bot platform.
Sourcing knowledge on the fly
Compared to that of most conversational Als I’ve put together, creating a knowledge base for our easyJet Flight Attendant proved quickly achievable. This was less to do with the easyJet brand, and more to do with the fact they’re an airline; flight providers commonly consolidate details volumes of service information online — complete with precise figures informing things like baggage allowances, flight procedures, and safety regulations. That meant being able source a well organised, comprehensive knowledge compendium that Voiceflow’s AI response node capably interrogated based on the questions being asked of it.
With the Flight Attendant’s Tone of Voice already defined, a specific prompt was inserted into the AI response node. This would dictate how the attendant should deliver its phrasing — citing the ToV principles, syntax, structure. It essentially means distilling the complete persona into a concise prompt optimised to be suitably interpreted by NLU while keeping token usage to a minimum. The final outputs were consistently good enough satisfy all our alpha release criteria:
Listening intently, and adapting tone
Using customer query data provided by easyJet, it also didn’t take long to identify a principle set of system intents. These can be created manually within Voiceflow which accommodates them by attaching each one to a brace of keywords. These can then that edited and refined to ensure the system is always able to capably detect a customers intentions. For example, if the user enters ‘Hi’, ‘Hello’, ‘Hey there’, ‘Bonjour!’ etc, this can all be configured to trigger a ‘greeting’ intent — which the chatbot can field with a friendly introduction and a few pleasantries.
We understood that our chatbot would need to accommodate a few use cases in which the customer was explicitly unhappy. These 'negative' intents would categorised as being more serious and therefore handled by the system in a way that was more sympathetic and understanding. Generally speaking, I believe that customer comms should try to avoid being overly apologetic (because saying 'sorry' for trivial errors or temporary service failure runs the risk of reminding the user that they’re having a bad experience). But, in the case of airline services, problems are often service critical and have the potential to completely erode customer satisfaction. We’re talking lost luggage, delayed or even delayed flights.
In these situations behavioural traits such as optimism needed to be toned right down. Using a more bespoke prompt applied to an AI response node connected to these negative intents, we were able to quickly engineer a behaviour that would instead pivot to being compensatory — acknowledging lacklustre service and working to improve matters with QR codes that could be directly redeemed for free complimentary drinks at outlets within the airport.
Enriched conversations with media assets
Integrating visual elements into the easyJet chatbot interaction significantly elevates the user experience, transforming routine transactional exchanges into engaging travel narratives. By pairing flight details with evocative images of Geneva’s mountainous landscapes and London’s iconic cityscapes, the chatbot not only enhances the information’s clarity but also stirs excitement and anticipation, leveraging the human preference for visual learning.
These strategic visual cues break up text monotony, making the user interface more dynamic and memorable. Additionally, visually appealing call-to-action buttons prompt immediate responses, seamlessly integrating decision-making into the conversational flow. This adept use of multimedia not only boosts user engagement but also reinforces easyJet’s commitment to innovative and customer-friendly service.
An MVP ready for take-off
This easyJet chatbot project served as an exemplary proof of concept, effectively demonstrating the capabilities of a robust and polished virtual assistant delivered rapidly as an alpha release. The integration of comprehensive flight details with interactive and visually stimulating elements showcased the chatbot’s ability to handle complex interactions with ease. This alpha release not only met but exceeded expectations in terms of functionality and user engagement, proving the concept’s viability and effectiveness.
The success of this initial deployment paved the way for a rapid development of a Minimum Viable Product (MVP). By leveraging the insights and positive feedback from the alpha testing, the project team was able to quickly iterate and refine the chatbot, focusing on scalability and additional features that could enhance user interaction further. This swift and strategic response underscored the project’s capacity for rapid development and deployment, setting a strong foundation for subsequent enhancements and broader implementation.