How to implement a successful digital customer service strategy using NLP
This post is aimed at those of you trying to solve a problem which appears, on its face, to be unsolvable; “How can I allow my business to scale without drowning under escalating contact centre costs?”
Let’s consider this problem. If you assigned an agent to each and every user request you received, you’d never get out of the red, and yet if you bury your phone number so deep within the website that no one can find it, the next thing you know people are leaving you one-star reviews on TrustPilot. So, what can be done?
Before jumping in to the practical steps of implementing an efficient online CS strategy, I need to digress briefly and talk about a burgeoning discipline known as NLP (or sometimes NLU, depending on who you speak to). What is NLP? Well, Natural Language Processing, put as simply as possible, is a subset of a wider discipline with which most of you will be familiar; AI.
If the broad-strokes goal of AI is to create machines that look, act and think like humans, the goal of NLP can be similarly rendered; to make machines that can communicate with humans in normal, everyday, natural language.
This is easier said than done. Whilst machines are extremely effective at performing simple tasks, codifying something as complex, and as ever-changing as human language is a tough proposition. The way in which we communicate with each other on a daily basis is filled with nuance, is open to interpretation, and may entirely depend upon the relationship between the locutor and the interlocutor.
Take the following example: If I were to say to you “please close the window,” you would have a very clear understanding of my intent, and would maybe (depending our our relationship) get up and close the window for me. It’s a very clear and unambiguous request. But, if i were to say “It’s a bit cold in here,” you might still get up and close the window, even though the message’s intent has been entirely repackaged. Or you might not. Or maybe we’re in a room without any windows at all and neither elocution makes any sense… but, as a human there is a very good chance that you would be able to instantly decode the intent of the message unconsciously, and quickly decide how you want to react to this information. The same cannot be said for machines, for whom this task is far more complex. This is what NLP is all about, and this is what those who practice it — a vocation known as Computational Linguistics — are trying to achieve; to apply laws and logic and order to the inherently chaotic phenomenon which we call language.
NLP-based applications have been around for some time now, and some have been proven, beyond any reasonable doubt, to be effective cost reducers for businesses both large and small. The online Help Centre, for example, should be an absolute staple for your business regardless of the vertical. Get your support topics organised and prioritised and expose them to your users as quickly as possible. There are some really great products out there with good features: rolling banners, fixed contents for recurring issues, dynamically powered Popular Topics…
Other, more recent evolutions have emerged in the last few years within the NLP space, and include applications such as chatbots and intent-based process automation. Again, like the help centre, this software can drastically reduce the number of contact tickets you are currently receiving in to the call centre, and can free up your agents to deal with the more complex tier-2 and -3 type topics. They will also create an uptick in the overall UX of the site, since you will now be surfacing the best and most relevant content to your users faster than ever before, meaning that they no longer need to sit in a call-queue, or wait 3 days for an email response from one of your long-suffering agents.
Build better voice apps. Get more articles & interviews from voice technology experts at voicetechpodcast.com
However, with so many vendors, each of whom purports to hold the magic wand, how can you be sure which of the above mentioned solutions is right for your business?
First, let’s figure out the problem you are trying to solve. Too many emails? Too many live chats? Agents always dealing with the same topic over and over again? It’s crucial to first diagnose the problem before you can treat it. Align with your analytics teams. Is there a certain point in the user journey where you are experiencing a high drop-off rate? Speak to your MI teams and find out where. Are people calling you asking about the same 5–10 topics day after day? Align with your contact centre teams and find out what those problems are.
Secondly, you need to select the right supplier. This sounds obvious, but this is easier said than done. Chatbots vendors number approximately 700 in London alone. There is an awful lot of noise around this topic at the moment and it’s difficult to navigate one’s way through the din. You need to select a vendor with good, out of the box NLP. If you have a customer support knowledge base which is made up of 400–500 intents, you cannot afford the time necessary to train the system by inputting “intent variables”, or “utterances” for each of those use cases. If we conservatively assume the you would need 20–30 utterances per intent, this proposal quickly becomes unmanageable. You need to select a vendor with advanced AI and NLP technology so that this type of intelligence is available from day one. Without this, as a best case scenario, you are consigning some unlucky intern to months, or possibly years, of unimaginable tedium before you’re even ready to launch.
Thirdly, don’t hesitate to launch a PoC or an AB test. Or maybe as a phase one of the launch you can only expose the new technology to 5–10% of your users as a way of dipping your toe in the water. Set KPI’s. Get references from other projects from similar verticals. You should know what to expect before it happens.
And finally, leveraging key learnings needs to be underscored as fundamental to the success of this project. This is not something that you can implement and forget about. If you’ve launched a Help Centre application, where are the clicks coming from? Which part of the site is the user navigating to after a successful resolution? Is the user leaving positive of negative feedback? If negative, then why? If you’ve launched a chatbot application, which questions has the user been asking? More crucially, which questions has the user been asking to which the chatbot has no response? In cases where the user was unable to find an answer, this is the moment that they will look to abandon the digital channel and look for a human point of contact.
All of this user behaviour needs to be collected, collated, and reacted to. With the right software, this is not an arduous process, but it does require some effort; again, this is not something you can deploy and forget about… In the case of a Help Centre, you need to assign a content manager, in the case of a chatbot, a botmaster, somebody who can run the necessary reports and close the gaps in the knowledge base in order to move forward and optimise performance.
As a final thought, I wanted to comment on a commonly held fear that AI and automation represent an existential threat to the workforce. I do not believe that this is true (at least not in the short term). In the next 5–10 years the future of this technology is — to my mind — very much aligned with a hybrid model; a seamless combination of AI and human. The AI will handle the more mundane and repetitive tasks; “Where is my nearest branch?”, “How can I reset my password?”, “How do I switch on the Sat-Nav?”, and all the while a human agent is available to take over if it becomes necessary.
