Chatbots and AI have the potential to automate large amounts of customer service queries. Contact centres produce large amounts of structured and tagged data in a bespoken domain, which is a treasure grove for Artificial Intelligence.
Unfortunately, this treasure is only rarely harvested. Secondly, almost any Chatbot will get stuck in a conversation after several conversational turns. The predefined universe of a Chatbot can simply not match the real universe of human beings. There are attempts to do it altogether without human beings, such as the AirAsia customer with its Ava chatbot, which chose to eliminate all its contact centres.
Bots and Human Service Agents complement each other
At some point, a handoff to agents is required. The downside here is that human-to-agent chat is typically slower and less rapid than a human-to-agent phone call on a per-transaction basis, even if an agent can handle multiple chats at once. And these 70% of the remaining chat volume from a customer’s perspective is then human-to-human chat and not human to chatbot, thereby eliminating the gains of the chatbot.
What we suggest is to have Chatbots and AI in general work along contact centre agents, seamlessly blending in during a conversation, answering whenever the AI’s confidence score is high. The chatbot will therefore automate every portion of the conversation that he can handle, but an agent will work alongside with him, chipping in and out:
In this way, the human effort by an agent on one particular transaction can be reduced, and the customer’s requirements can still be met.
Human agent actions and answers should then be taken as tagged data. Especially if the human agent acts differently than what a Chatbot would have predicted, this can add a valuable data point to the Chatbot’s universe.
Thank you for reading this post, don't forget to subscribe to our AI NAVIGATOR!
We typically encourage companies not to automate everything, but to define an AI universe for certain very frequent conversational elements only. These can happen also later during a transaction during the fulfilment stage of a conversation, for example once it’s clear what a customer wants. Unfortunately, with the traditional approach with a handoff from a bot to a human agent, this simple transactional piece is done by a human. A hybrid human/ bot approach allows for automation even during later stages of a conversation.
Towards a single AI
A second improvement for AI in customer care is that the same language model should be used not for one domain only (Email response or Chatbot or Voice Suggestions). It is likely that the same language, product names, entities, intents are used across all media (voice, chat, email, written text). Many companies still have an approach where if they automate, they automate on one domain only. We think that It makes sense to extend a good bespoken NLU model across all media domains. Any learning in the NLU space will then profit both calls, chat and email analysis:
Currently all major AI vendors have a layered approach: An ASR engine for voice, then an NLU engine to extract intents and entities, followed by a conversational engine to predict actions. Written text is a bad representation of human voice – it doesn’t represent emphasis or emotions. We think that in the coming five years these layers will be merged into a single bespoken neural network across all media domains (voice or text). Such networks will learn across multiple media domains and will suffer less from signal loss and aliasing by having low-information signals (such as text) as forced intermediary layers, which is currently the case.
It’s the data, stupid!
With rapid advances and known weaknesses in the AI space, we think that a sensible approach is to keep a critical distance to algorithms, but instead to focus more on where which data is captured and how data is used. The most precious resource is customer data, not algorithms. We encourage companies to continuously improve, change and replace their algorithms, and to keep their primary resource – their data – intact and nurtured with continuous tagging. This is a shift away from a typical mono-vendor lock-in approach as driven by typical AI vendors. Investments should be done in data flow and data nurturing, as this will be useful independently of which AI is being used.
Human customer service agents will continue to play a vital role in customer care, providing a baseline of customer service. AI is a complement and not a replacement of human interactions. Ignoring AI will result in high cost, ignoring the human aspect will result in low customer happiness. Companies that master this human-AI blending will be providing good and cost-efficient customer care.
Copyright by Andreas Stuber, CEO, ExpertFlow
AI in CX Requires Humans and Data to Perform
Chatbots and AI have the potential to automate large amounts of customer service queries. Contact centres produce large amounts of structured and tagged data in a bespoken domain, which is a treasure grove for Artificial Intelligence.
Unfortunately, this treasure is only rarely harvested. Secondly, almost any Chatbot will get stuck in a conversation after several conversational turns. The predefined universe of a Chatbot can simply not match the real universe of human beings. There are attempts to do it altogether without human beings, such as the AirAsia customer with its Ava chatbot, which chose to eliminate all its contact centres.
Bots and Human Service Agents complement each other
At some point, a handoff to agents is required. The downside here is that human-to-agent chat is typically slower and less rapid than a human-to-agent phone call on a per-transaction basis, even if an agent can handle multiple chats at once. And these 70% of the remaining chat volume from a customer’s perspective is then human-to-human chat and not human to chatbot, thereby eliminating the gains of the chatbot.
What we suggest is to have Chatbots and AI in general work along contact centre agents, seamlessly blending in during a conversation, answering whenever the AI’s confidence score is high. The chatbot will therefore automate every portion of the conversation that he can handle, but an agent will work alongside with him, chipping in and out:
In this way, the human effort by an agent on one particular transaction can be reduced, and the customer’s requirements can still be met.
Human agent actions and answers should then be taken as tagged data. Especially if the human agent acts differently than what a Chatbot would have predicted, this can add a valuable data point to the Chatbot’s universe.
Thank you for reading this post, don't forget to subscribe to our AI NAVIGATOR!
We typically encourage companies not to automate everything, but to define an AI universe for certain very frequent conversational elements only. These can happen also later during a transaction during the fulfilment stage of a conversation, for example once it’s clear what a customer wants. Unfortunately, with the traditional approach with a handoff from a bot to a human agent, this simple transactional piece is done by a human. A hybrid human/ bot approach allows for automation even during later stages of a conversation.
Towards a single AI
A second improvement for AI in customer care is that the same language model should be used not for one domain only (Email response or Chatbot or Voice Suggestions). It is likely that the same language, product names, entities, intents are used across all media (voice, chat, email, written text). Many companies still have an approach where if they automate, they automate on one domain only. We think that It makes sense to extend a good bespoken NLU model across all media domains. Any learning in the NLU space will then profit both calls, chat and email analysis:
Currently all major AI vendors have a layered approach: An ASR engine for voice, then an NLU engine to extract intents and entities, followed by a conversational engine to predict actions. Written text is a bad representation of human voice – it doesn’t represent emphasis or emotions. We think that in the coming five years these layers will be merged into a single bespoken neural network across all media domains (voice or text). Such networks will learn across multiple media domains and will suffer less from signal loss and aliasing by having low-information signals (such as text) as forced intermediary layers, which is currently the case.
It’s the data, stupid!
With rapid advances and known weaknesses in the AI space, we think that a sensible approach is to keep a critical distance to algorithms, but instead to focus more on where which data is captured and how data is used. The most precious resource is customer data, not algorithms. We encourage companies to continuously improve, change and replace their algorithms, and to keep their primary resource – their data – intact and nurtured with continuous tagging. This is a shift away from a typical mono-vendor lock-in approach as driven by typical AI vendors. Investments should be done in data flow and data nurturing, as this will be useful independently of which AI is being used.
Human customer service agents will continue to play a vital role in customer care, providing a baseline of customer service. AI is a complement and not a replacement of human interactions. Ignoring AI will result in high cost, ignoring the human aspect will result in low customer happiness. Companies that master this human-AI blending will be providing good and cost-efficient customer care.
Copyright by Andreas Stuber, CEO, ExpertFlow
Andreas Stuber
CEO
Expertflow
Virtual Global AI Conference
Co-Hosted by AI Capital & SwissCognitive
“Beware of AI lock-in.”
Andreas Stuber, CEO, Expertflow
Remarks from SwissCognitive: Andreas Stuber was one of the global speakers at SwissCognitive’s first Virtual AI Conference, co-organised with AI Capital on 31st March and 1st April. The conference gave an intensive overview from various industry-perspectives on how AI helps us to overcome challenges caused by the Coronavirus, and how this technology is going to provide us with new ways of processes and functioning after the crisis. The Virtual AI Conference was attended by 500 attendees, calling-in from 20 countries, and its content was spread through SwissCognitive’s social media channels, reaching 400k followers in the AI eco-system.
Share this: