AI Chatbot Development Methodology
- 1. AI Chatbot Development Methodology 19th October 2018 Lviv
- 2. +38 096 448 80 46igor.luzhansky igor@chatbots.studio About me • Co-founder of http://Chatbots.Studio • Co-author of course “Chatbots for marketers” • Owner of MBA with exceptional recognition • Bot evangelist https://medium.com/@chatbots_studio/
- 3. +38 096 448 80 46igor.luzhansky igor@chatbots.studio Content • Scope of methodology • Team overview • Project stages • Envisioning • Planning • Developing • Stabilizing • Deployment / piloting • Conclusion
- 4. +38 096 448 80 46igor.luzhansky igor@chatbots.studio We cover how to manage the AI bot delivery 1 • Inherits waterfall methodology ( I love MSF) 2 • Distinct project activities per each project stage 3 • Criteria's of completion 4 • Key decisions per stage 5 • New team roles
- 5. +38 096 448 80 46igor.luzhansky igor@chatbots.studio What will NOT be in the methodology ? 1 • How to program any bot or its component 2 • How to market, earn money on chatbot 3 • What is the chatbot and how to use it my business 4 • How to run a software development project. You should know that already
- 6. +38 096 448 80 46igor.luzhansky igor@chatbots.studio Content • Scope of methodology • Team overview • Project stages • Envisioning • Planning • Developing • Stabilizing • Conclusion
- 7. +38 096 448 80 46igor.luzhansky igor@chatbots.studio Each problem of the project should have owner
- 8. +38 096 448 80 46igor.luzhansky igor@chatbots.studio AI Chatbot project requires 2 new roles under user experience Role Problem Responsibility Conversational expert • Conversation is not engaging, It is not fun to talk with bot • Bot does not understand me • Bot does loses the context of conversation • It requires me to do a lot of actions to get what I need • Define the conversation of flow • Monitor the real conversations and optimize them • Maintain NLP engine, resolve fallouts and add synonyms Live user support • Users are leaving a conversation • Users are complaining about bots • Service is becoming lower • Resolve user queries • Contribute to conversation improvements
- 9. +38 096 448 80 46igor.luzhansky igor@chatbots.studio Content • Scope of methodology • Team overview • Project stages • Envisioning • Planning • Developing • Stabilizing • Conclusion
- 10. +38 096 448 80 46igor.luzhansky igor@chatbots.studio MSF is rigid but clear
- 11. +38 096 448 80 46igor.luzhansky igor@chatbots.studio Content • Scope of methodology • Team overview • Project stages • Envisioning • Planning • Developing • Stabilizing • Conclusion
- 12. +38 096 448 80 46igor.luzhansky igor@chatbots.studio Vision for AI chatbot project should include 1 • Objectives and metrics for AI Chatbot deployment 2 • Interfaces of a Chatbot •3 • Approach for AI
- 13. +38 096 448 80 46igor.luzhansky igor@chatbots.studio Each project should bring tangible business benefits Increase revenue Develop new sales channel Increase secondary sales Reduce cost Optimize information management Automate manual work Reduce cost of advertisement companies
- 14. +38 096 448 80 46igor.luzhansky igor@chatbots.studio Be clear at the very beginning where you bot will be working
- 15. +38 096 448 80 46igor.luzhansky igor@chatbots.studio When you definitely need AI in your Chatbot No Controls in Interface • SMS • WhatsApp • Voice • E-mail A lot of data records • FAQ • Search Assistance Non structured nature of information • Dreams • Complains • Reviews
- 16. +38 096 448 80 46igor.luzhansky igor@chatbots.studio When you definitely do NOT need AI in your Chatbot Rare AI language • Lack of NLP libs • No support from Big players • Ukrainian is in the list Customer has a distinct task • Buy ticket • Setup an appointment • Make a request No enough data to train NLP bot • less than 100 records
- 17. +38 096 448 80 46igor.luzhansky igor@chatbots.studio Content • Scope of methodology • Team overview • Project stages • Envisioning • Planning • Developing • Stabilizing • Conclusion
- 18. +38 096 448 80 46igor.luzhansky igor@chatbots.studio Should be considered during AI Chatbot planning stage 1 • Defined a detailed conversation flow 2 • Character of the bot 3 • Approach for AI implementation 4 • Define strategy for bot training 5 • Gather corpus of data 6 • Select development tools •7 • Define requirements for integration 8 • Select live agent platform
- 19. +38 096 448 80 46igor.luzhansky igor@chatbots.studio Watch dialog structure, it should be careful
- 20. +38 096 448 80 46igor.luzhansky igor@chatbots.studio Consider the full technical landscape Component Role Possible Solutions Conversation Engine • Manage the conversation logic • Maintain the session • Route the conversation between interfaces • Botkit • Botpress • Use own by using the plain programming NLP Engine • Identifies intent (NLU) • Holds the context of conversation • Speach2text2Speach • LUIS • DialogFlow • Amazon Lex • Rasa.ai • Wit.ai Live Agent • Takeover the conversation with a user • Live Chat • Intercom Bot analytics • Google analytics for bots • Chatbase UI framework • Presentation layer for web bot and admin part • Angular • React DBMS • Storage of the data • MySQL, MS SQL, Postgres and etc.
- 21. +38 096 448 80 46igor.luzhansky igor@chatbots.studio Bot is not a human, but has a character
- 22. +38 096 448 80 46igor.luzhansky igor@chatbots.studio Plan to find a data for training in advance Company sources • Contact center logs • Existing support scripts • FAQ Dedicated sources • Semantic DBs • NERs Scrapping • What ever you can find and get from internet
- 23. +38 096 448 80 46igor.luzhansky igor@chatbots.studio NLP Engine vs Own AI - Cheap and quick development - Not need for Data Scientist to train - Great results in standard cases - Has APIs for integration - Does not support the rare languages - Have issues with a lot intents - Could miss the context of conversation - Limited with NLU mostly NLPEngine - Could implement practically any NLP function - Covers NLG as well - Could support rare language - Requires a lot of Data to train - Long delivery - Requires specific knowledge OwnAI
- 24. +38 096 448 80 46igor.luzhansky igor@chatbots.studio Content • Scope of methodology • Team overview • Project stages • Envisioning • Planning • Developing • Stabilizing • Conclusion
- 25. +38 096 448 80 46igor.luzhansky igor@chatbots.studio Developing of stage is not only about programming 1 • Define client onboarding strategy 2 • Create unit tests for external components 3 • Define test cases for NLP
- 26. +38 096 448 80 46igor.luzhansky igor@chatbots.studio Plan in advance how the clients will come to your bot From own sources • Site • FB page • Call center phone • Messenger Search Offline • Posters • Outlets Partners • Part of other business process
- 27. +38 096 448 80 46igor.luzhansky igor@chatbots.studio What conversational AI test cases should cover ? • Miss spellings • Synonyms, abbreviations, professional jargon • Big quantity of similar intents (NLP engines could not recognize the difference) • Logical (and, or) combinations • Rejections of previous messages
- 28. +38 096 448 80 46igor.luzhansky igor@chatbots.studio Content • Scope of methodology • Team overview • Project stages • Envisioning • Planning • Developing • Stabilizing • Conclusion
- 29. +38 096 448 80 46igor.luzhansky igor@chatbots.studio Specific activities for stabilizing stage 1 • Define the launch approach 2 • Issue specific environments for NLP engines 3 • Train live agents
- 30. +38 096 448 80 46igor.luzhansky igor@chatbots.studio Define your launch strategy before the launch. Start with small Target audience • Limited • Loyal • Use the full scenario Scope of launch • What will be used • What will be omitted How to resolve issues • Redirect to agent • Set the trial mode Treat unhappy clients • Do not over promise • Think about incentives
- 31. +38 096 448 80 46igor.luzhansky igor@chatbots.studio Content • Scope of methodology • Team overview • Project stages • Envisioning • Planning • Developing • Stabilizing • Conclusion
- 32. +38 096 448 80 46igor.luzhansky igor@chatbots.studio Take offs • AI Chatbot is NOT a toy • AI Chatbot delivery requires a systematic set decisions • You will need new roles on your team
Editor's Notes
- #8: https://www.slideshare.net/drdenizkilinc/msf-microsoft-solution-framework
- #9: https://www.slideshare.net/drdenizkilinc/msf-microsoft-solution-framework
- #14: Get a slide of my value added presentation
- #15: Add a slide with possible bot interfaces, split them at AI based and not
- #17: Language is not supported by NLP engines Users have a distinct task You do not have enough data to train your AI
- #22: Screens with the bot characters Find a post with that ?
- #24: NLP engine – support your language, able to recognize the intents correctly - Does not meet your requirements AI + you have enough data to train, your have a DS in your team, you need NLG, functions of NLP engine is not enough - You have not done that before
- #26: Определение стратегии тестирования и багов дополнить
- #27: From site, FB page From offline QR From advertisement Other will force a user to use it
- #28: miss spelling Many similar content Synonyms Jargon and similar words
- #31: What is the target audience for start What is the scope for launch How you resolve issues How do you treat unhappy clients