Chatbot at Kiwi.com: Dreams vs. reality

Talk about how we implemented chatbot for customer service at Kiwi.com - and failed.

Transcript

1. Chatbot at Kiwi.com: Dreams vs. Reality
2. Zdravo! My name is Jaroslav Follow me on @kubajzHK
3. in numbers... × 1500+ agents in CS × 10% as chat support
4. It doesn’t end with selling the ticket... × Delayed flights, strikes at/around airports, bad weather, lost baggage
5. Why we could not afford chatbot... × Tired man at the unfamiliar airport, who just missed the flight, needs proper care => we can’t fail our customers “I don’t understand your question.” is not an answer, it’s an insult.
6. CHatbot needs to work 100 % time … but we are not able to predict all possible scenarios
7. Hello Hi Bob, how can I help you? I need to add a baggage Ok, Can I ask you for your booking ID? Is it that number from e-mail?
9. So how does it works?
12. What’s behind?
13. Tree is never the answer.
14. Which type of coffee do you prefer? Big or small? What? I like Moka pot. #@Y! Ok, big one. I like big shot too. WAT?! I want Moka pot! Sorry, I can’t understand. How can I help you?
16. Catch the intent Fill in the slots Process the action on fulfillment
17. DynamoDB
18. Take conversation state + new message Dispatch action Update state Store state + send response
19. © Nicolaus Graf: React + Redux Introduction
20. export async function query(event, context, callback) { const body = JSON.parse(event.body); const stateStorage = getStateStorage(); const store = createInitializedStore(stateStorage, ..., ...); if (body.userSays) { store.dispatch( analyzeMessage( await store.getState(), apiAI, body.userSays, body.sessionId), ); } const state = await store.getState(); await stateStorage.storeState(body.sessionId, state); callback(null, prepareResponse(messages)); }
21. So What’s the result?
23. Challenges for chatbot × Requirements for messages are changing × Every agent has his own taste × It’s not instant - takes time to suggest
24. We Also discovered that... ...agents are Lacking more basic needs “It takes time to find info for specific flight.”
26. Some results × About 20 % of messages from extension × We saved ~ 2 mins per conversation => 7 mandays per day
27. What we have learned × Chatbot might be future, but it’s not present × It will take some time to evolve × Yet we can leverage it in specific scenarios today × Solve problems, not build sand castles
28. Today future is about cooperation between human & bot, not in full replacement
29. Additional resources... × How we failed with chatbot. The success story http://bit.ly/2ysC95p × Serverless Chatbots with Amazon Lex & AWS Lambda https://www.youtube.com/watch?v=TlKtGGoMpF0
30. hvala! Any questions? You can find me at @kubajzHK code.kiwi.com