Originally published on IT-Online on 5 September 2022
By Ryan Falkenberg, Co-CEO of CLEVVA Pty. Ltd.
When conversational-AI powered chatbots started entering the mainstream a few years ago, many businesses thought they would be a panacea for their customer service needs.
These digital assistants, the thinking went, would greatly reduce the load on customer service agents. By resolving more of the known queries, chatbots would free human agents to focus more of their efforts on having richer, more meaningful and impactful conversations that really matter. The reality has turned out to be very different.
Far from improving the customer experience, chatbots are frequently cited as a primary cause for customer dissatisfaction. Customers often complain that they cannot get their queries resolved by the chatbot, and end up being connected to a human agent anyway.
So, what went wrong? Why have chatbots proven to be such poor ‘virtual agents’ and what can be done to change this?
A major part of the problem is the way that chatbots learn. Most chatbots learn in two ways: they learn off content provided by knowledge bases and systems and they learn off customer responses. The problem with the first is that the content within most company knowledge bases is very generic.
This makes it very difficult to give a very specific answer to a very specific question. Also, learning from your mistakes and promising to get better next time sounds reasonable, unless you are the customer who has to endure the learner chatbot.
While Google has succeeded in fine-tuning its predictive accuracy by learning off billions of patient users, most chatbots don’t have this luxury. They are expected to get it right the first time. If not, customers tend to ask to speak to a live agent who can.
It’s a bit like going up to the counter at an ice cream store and asking for a recommendation on their nut-free options. The assistant looks at you and answers ‘We have lots of ice-cream flavours. Would you like a list of everything we offer?’ You would think to yourself ‘Am I in a Monty Python video? Did he hear what I said? I asked specifically about nut-free flavours, not every ice-cream they have!’ Then you try again, only for the assistant to then offer a taste of the new cherry flavour. It won’t take long before you either ask to speak to the manager, or simply leave and look for an ice-cream shop that knows what it is doing.
Understanding chatbot limitations
A chatbot is only as smart as the data it has access to and the patterns it recognises. It is not intelligent simply because it can understand natural language.
Many businesses have been duped into believing that a technology claiming to be powered by artificial intelligence is by default intelligent. The reality is that it may have the potential to learn, but like a human child it needs to learn the right stuff and have the chance to practice it enough times so it learns the right patterns.
Very few people can perform well if they are given a generic book to read, and then expected to perform immediately on the job. They need to learn through trial and error. That is why experts take years before they have the experience to adapt to most situations. They need to learn the patterns.
It’s the same for chatbots, except that within a regulated company like a bank, you can’t afford your chatbot to take years before they start getting it right. And you also can’t afford the mistakes along the way.
AI’s most fervent evangelists have made it seem like chatbots are all powerful and can perform like human agents. The truth is they can’t. Chatbots are simple question and answer machines. You ask the question and they try to give you the answer. They are designed primarily for unstructured conversations, not the structured, rule-bound yet highly contextual conversations that human agents are required to have when resolving a customer query.
Beyond chatbots – virtual agents that can do the full job
To automate the resolution of service queries in a consistent, compliant and yet hyper-personalised way, a far more capable virtual agent is required. One that will clarify, analyse, assess and diagnose before looking to answer or action. A virtual agent that performs like an expert, not an assistant. That starts out intelligent, and rather learns to adjust to different contexts not the basics of the job.
Fortunately these virtual agents are now available and in production with many large and often regulated businesses. This is good news for customers, and for the bottom line.
View the original article here.