Chatbot and AI Design Principles
8 likes1,831 views
This document provides guidance on designing successful chatbots and conversational AI. It discusses important considerations for understanding users, defining the problem, scoping the bot, designing conversations, and measuring success. Key points covered include establishing trust, handling unknown inputs, accessibility, and the importance of testing before going live. The overall message is that good UX design principles are needed to build bots that can engage users and successfully help them.
Chatbot and AI Design Principles
- 1. AWS Summit 2018 Version 2.0 Chatbot and AI Design Principles UX eye for a tech guy
- 2. Hola Soy humano Mauricio Perez UX/CX Service Design Strategist
- 3. AWS Summit 2018 Version 1.0 The story » Understanding your users » Identifying the problem » Defining the scope » Designing conversations » Measuring success
- 5. Customer personas What do we need to know about our customers to ensure the chatbot will be successful? Demographics Needs ● Age ● Devices ● Location ● Gender ● Language ● Occupation Behaviours What are their needs and emotions when it comes to interfacing with your product / services? Why? What daily behaviours do they have? What behaviours do they exhibit with your product / service and why?
- 6. Customer contexts What do we know about your customers contexts? Physical context Cognitive context Emotional context
- 8. How might we solve the problem of… Problem statement
- 9. Defining the scope of your bot 3
- 10. Chatbot objective The chatbot will try to solve the problem of… when people are… so that…
- 11. What kind of personality does your chatbot need to have in order to engage and service your customers in their contexts? Chatbot persona Goals __________________ __________________ __________________ Personality Understated Enthusiastic Playful Dry Formal Informal Protip: Its okay for your bot to only have 1 goal at launch!
- 13. Mapping a conversation flow So usually with manual automation, we start with a key as we start to map the architecture. Choose a tool that facilitates real-time collaboration
- 14. Conversation types There are two main types of conversations. The first is task-led, where the aim is to accomplish certain goals while going through the conversational funnel. The second will be a topic-led conversation; this type is different because it’s more about entertainment. Choose your technology wisely as some will restrict your flexibility when things go wrong. Task-led Funnels you to an end Topic-led Endless exploration
- 15. Hello human!The chatbot greeting What do we need to consider in our chatbot greeting given the wide variety of contexts and the way people tend to converse in text? ● I’m not human (avatar helps) ● What I can do… ● Establish trust Greetings earthling Hi I just need to you know… Yo sup homie! Ciao bella que cosa? Hi, I’m Paris… Human please I need… > : (
- 16. Known inputs Illustrate all possible entry points and any “known” categories the bot will recognise in dialogue. Depending on the tech you choose, this will be one of the most time consuming things to get the desired experience at the beginning.
- 17. Unknown inputs Prepare a good “I don’t know” fallback ● Vary responses with hints ● Confirm user input in dialogue ● Try humor and personality ● Allow for a safety net referral
- 18. ● Be responsive ● Be remorseful ● Allow user to try again ● Vary responses with hints ● Capture errors to learn ● Allow for escalation words Responding to unknown inputs Hmm, I am sorry but I am still learning and I'm not familiar with those words, Could you try again using different words? I am sorry but I still cannot understand. Would you prefer to speak to a real person?
- 19. Safety nets Do you have any safety nets in place when things go wrong? How will people behave when they want to bail out of a conversation or escalate an issue that they cannot get help with? HUMAAAN!!! My bad…
- 20. Escalation When referring a person to a support person, keep in mind that they will have to read through the entire transcript in order to understand the user’s needs in context in order to help them. By the time they get back to the user, they might have abandoned the chat altogether. If only we had a service designer…
- 21. Sad loops When things go wrong, what things need to be in place to allow us to learn from these scenarios?
- 22. Establishing trust ● Demonstrate stability and credibility ● Address privacy and security concerns ● Be helpful but not creepy ● Be transparent about intent and ability
- 23. Offensive context People will inevitably play with and even abuse AI in different ways. We can take an active stance to address any offences and respond to contexts appropriately. ● Offensive language ● Harassment and abuse ● Self harm ● Mental health ● Threats of violence
- 24. The idle time-out This will be a common scenario if the system needs to time-out the user session. 1. Confirm “Are you still there?” 2. Responses for user activity 3. Time out dialogue 4. Memory of session Are you still there?
- 25. Example of an idle conversation This will be a common scenario if the system needs to time-out the user session.
- 26. Yes they allow quick and easy interaction, look great and restrict input to avoid broken conversations but there are some things to consider… Interaction elements AI UIlessmore
- 27. Accessibility Google’s Allo app People with vision impairment using Voiceover may not notice buttons and even if they did, they are not interactive with Voiceover.
- 28. When we think about conversation design in our chats, all the rules we put in place might one day be used in a wonderful voice application in cars fridges etc… which is exciting! Until we have to decide how to deal with all those User Interface elements we put in there to begin with… #facepalm Think of the future Alexa, put on Game of Thrones! Sure, which episode? <image 1> <button>… Pancake recipes! Take me to…
- 30. ● Demographics ● Locations / Times of day / Devices ● Needs / Intent capture / Top utterances ● Error fallbacks ● Funnels mapping task completion / abandonment ● Session times ● Retention of new vs returning users Quantitative Analytics indicators
- 31. ● Transcript trawling (you poor soul) ● Text-based user ratings ● Sentiment capture ● Live observation ● Usability testing Qualitative Analytics indicators Thanks for chatting - how would you rate our conversation? 1 = poor, 5 = excellent.
- 32. DO NOT GO LIVE WITHOUT DOING THIS People, context and technology will always end up surprising you in conversations. I would recommend at least 3x rounds with 10 people each time to get an idea of how people will interact with your chatbot and how your bot will handle when things go wrong. Usability testing
- 33. Provided you have a good safety net for escalations, users can still have their needs met with a low success rate when a chatbot fails. Ideally, you should be aiming at an 80% success rate unaided from a chatbot before launch. Deciding when to go live
- 34. Fin Or it just the beginning? Come and talk at our booth! dius.com.au/…