Have you been the victim of chatbot incompetence recently? It typically starts with a specific query that you need help with. You don’t have the time to listen to the contact centre’s hold music, so you turn to the company’s much vaunted chatbot.
It seems fairly straight forward. You type in your question and press Enter. The chatbot comes back with a list of completely unrelated content links, and asks if any of these solve your problem. ‘No’ you say. You retype your question, hoping this time it’s a little clearer. Again the chatbot cheerfully responds with a new list of possible ‘helpful’ articles and FAQ links, and mentions that it is busy learning and is grateful for your help. It will get more accurate the more people engage with it.
Oh, so your diabolical customer experience is all for a good cause – to train their chatbot! The cheek of it. After a third attempt, you notice a link that may be relevant. You click on the link and are taken to a 3 page document providing generic product information. The assumption is that you will take the time to read this document and then work out the answer yourself. With raised blood pressure, you click on the ‘Connect me to an agent’ link and hope that possibly they may have the knowledge needed to solve your query. Sometimes they do, sometimes they don’t. That’s how it goes with so many omni-channel customer journeys these days.
I must confess I expected more. I envisaged that by now we could engage with chatbots that are capable of diagnosing my specific issue, and then offering me a relevant response that results in a relevant action. In other words, a chatbot that is not simply an over-hyped digital assistant that can execute basic instructions or offer me links to possible content matches. I had in mind an digital advisor that could operate at the level of an expert – one whose intelligence is defined as much by the relevance of the questions it asks as the answers it finally offers.
If you talk to most AI companies, their chatbots already perform like digital experts. They will refer to their amazing natural language understanding and incredibly intelligent algorithms that are powered by ‘machine learning’, ‘deep learning’ and ‘neural nets’. They will give you the sense that all you need to do is point their technology in the direction of your knowledge base and the digital advisor will magically onboard all your product, policy and procedural expertise. And then, with just a little bit of guidance, you can soon have trained your chatbot into a digital Einstein that can change your customer service offering forever.
When you ask them to show you a working example, they will probably show you one of their canned demo’s – built off a scenario where the source data is in rich supply, the use case is clearly defined, and the user script can be carefully followed. As a result, their chatbot’s conversation will feel so intelligent, so human-like, that you will feel you simply have to have one.
Just don’t ask them mid-way to type in something unscripted and to upset their crafted story line! I am certain that you will be quickly informed that they have not managed to train this chatbot to cover all contexts, and that this is simply used for illustration purposes. The real reason is that it is all really a digital mirage. It looks so achievable until you shift your eyes down to your current position, and suddenly the mirage vanishes. There are a number of reasons for this. Companies seldom have the quality of data needed to accurately train a customer facing chatbot.
Most companies operate in a world of legacy systems, limited integrations, poor quality data, and poorly documented internal policies and procedures – the very things that cognitive systems depend on to build their engagement accuracy. A customer support chatbot is powered more by prescriptive than predictive logic.
To understand the difference, ask Siri or Alexa for an answer based on available information, and they can usually give it to you. For example, if you ask: ‘What is the weather looking like tomorrow in London?’, you will be amazed how accurate the answer is. That is because the information exists, and thousands of people have already asked the same or a similar question. The patterns are thus established and the correct answer can be predicted.
However, try ask a question that requires more context before answering. Say: ‘What is the best home loan for me?’. You will probably notice that the response will be to offer you possible links to companies offering loans. It won’t begin by understanding your needs. This is because a financial need analysis is driven off a diagnostic set of prescribed logic. There is no answer yet – the problem still needs to be understood.
In regulated environments, you need to be able to prove your chatbot asked the right questions and offered the right advice. Where a chatbot is powered by predictive logic – the logic you need to train and that keeps ‘learning’ based on multiple engagements – you will find it will struggle in a regulated environment. This is because the logic is designed to change and adapt, based on user engagement. It is also hard to prove how a decision was reached, as each recommendation is made in what is often referred to as a ‘black box’. This is hugely problematic when you are offering customers advisory support in a regulated environment, such as banking and finance.
Context matters. And the way most companies capture prescriptive logic lacks context. Prescriptive logic is typically captured using documents (knowledge bases) or decision trees (process flows). It’s how we have trained employees brains for decades and it’s how we are trying to train our chatbots. So just like giving staff exercises to learn how to apply the formula to different situations, we get teams to train the chatbot, telling them when they are right and when they are wrong. The problem is that documents and decision trees are not able to capture all the possible scenarios. They can only describe a few. And as a result, the more variables you need to consider in order to offer a customer accurate, relevant advice, the harder it becomes to achieve.
The good news is that there are now digital platforms available that allow you to achieve the holy grail – a chatbot capable of asking me context relevant questions that then lead to relevant answers and actions. These platforms have been built off data-powered, prescriptive logic that can ensure your customers are offered a consistent, compliant and context-relevant digital engagement – one that leads to a successful customer service outcome – every time.
These platforms have acknowledged that not all logic should be predicted, and that for customer support chatbots the foundation of the logic has to prescribed. The trick is ensuring it is also contextual, and these platforms have now managed to do this in a way that can be maintained effectively. The dawn of chatbots capable of offering customers consistent, compliant and yet highly context-relevant customer engagements is upon us. And it is about time, too.
Ryan Falkenberg is co-founder and co-CEO of CLEVVA, a technology that allows companies to build a Digital Expert capable of powering all staff-assisted and digital channels.