AI-Powered Customer Service — Hopes, Doubts and Applications
A benign AI assistant with a HAL 9000 level of intelligence that cuts customer service costs is a dream come true. But is it just that, a dream?
Over the past decade, enormous technological progress has been made in areas such as neural networks, machine learning, deep learning and natural language processing.
That’s why digital assistants live in every smartphone now and businesses are drooling over chatbots. AI has tremendous promise in customer service. But will it be able to live up to its potential?
How realistic is it to power your customer service with AI, what is currently possible, and what can we expect from the future?
Before we dive in, let's quickly clarify the most important AI-terms and concepts that are frequently thrown around and confused.
AI, ML, DL, NLP in action
Artificial Intelligence (AI): AI is the capability of machines to imitate “intelligent” behavior. This includes performing tasks that normally require human intelligence, such as visual perception, speech recognition, decision-making, and translation between languages. AI is the big umbrella term.
Let's use the example of a chatbot in a customer service setting. When a chatbot is answering a customer question, it is simulating human behavior, and therefore we can call it AI.
But there are quite some distinctions to make under this umbrella term. Let's go through the most important ones.
Narrow vs general AI. AI can be broken into two types: narrow AI and general AI.
Narrow AI, also known as “weak” AI, is programmed to perform a single task by pulling information from a specific data set. All AI applications known today are narrow.
General AI exhibits human intelligence and could successfully perform any task that a human can – and more. It is, for now, still science fiction. HAL 9000 is an example of general AI. It’s able to learn, plan, reason and communicate in natural language, and apply these skills to any task as if it has a (super) human brain.
Today’s chess computers use narrow AI to make chess moves by pulling from the data set they were programmed with. But if you played against general AI, it would not only make the moves, but it'd be smart enough to throw you off your game by trash-talking you using dirt it found about you online.
In our scenario of a customer sending a question to a chatbot, narrow AI could run the customer's question through a FAQ database, and offer possible answers from the knowledge base.
General – or "strong" – AI, on the other hand, would give perfect, human-sounding answers. It could probe for clarification, scour the internet for information and tailor answers on the fly. It’s the type of AI many people fear ( "the singularity" ), but it is very hard to predict when – and if – it will ever materialize.
“Fixed” AI: Many, if not most, of the pre-built chatbots currently available are “fixed.” They don't “learn” from their interactions; they just work from a predetermined decision tree.
A fixed chatbot may ask the customer multiple-choice questions, and then offer answers or take actions, like forwarding the chat to a human support rep. It’s all based on the selected route.
Machine Learning (ML): If you’ve used Spotify, YouTube or Netflix, then you’re likely familiar with receiving personalized recommendations. That’s because these platforms use algorithms to parse your data, learn from it and make predictions and classifications on what you might enjoy, otherwise known as machine learning.
In our chatbot example, a “simple” machine learning algorithm would analyze the customer question, run it past older similar customer questions and their successful answers, and offer the most likely answer based on these past answer(s) to the customer, or knowledge base articles.
Through customer feedback ( this answered my question/this didn't answer my question ), the machine learns whether it did a good job. If it gave a wrong answer, or can't figure out what answer to give, it forwards the question to a human colleague.
Ideally, it would then also track and learn from the answer that the human colleague subsequently provides to the customer to solve similar requests by itself in the future.
The advantage of chatbots that operate in this retrieval-based machine learning manner is that their answers are relatively reliable, because the chatbot only uses “proven” answers that were provided before.
On the downside, they can only handle simple, straightforward questions. And the answers can appear rigid and non-human.
At Userlike , we think this is fine as long as the customer realizes they are dealing with a simple chatbot, and if they have the option to easily escalate to a human service rep. Retrieval chatbots cannot do “real” conversations with meaningful back-and-forth, which deep learning promises to do.
Deep Learning (DL): Deep learning is a more advanced subtype of machine learning. It enables machines to make more accurate predictions without human help.
DL applications use a layered structure of algorithms called an artificial neural network to draw conclusions similar to how a human brain works.
Instead of basing the answer on retrieval of past successful answers, the chatbot in our example could generate its own answers, and interact in a conversational back-and-forth with your customers – like asking for clarification, probing, etc.
This requires a much larger data set than simpler machine learning approaches. But with enough data, DL can do amazing things.
For example, Google's DeepMind was able to beat the best human champions and “specialized” computers in a variety of games, such as chess, Shogi and Go . It pulled off this impressive feat by training itself in just a few hours of self-play, which generated huge amounts of data.
Games have a limiting framework of rules, however, and DL has had a harder time beating humans in the arena of human conversation. DL chatbots still haven’t passed the Turing test , for example.
And what to think of the Twitter bot that Microsoft unleashed upon the world? By “learning” from its interactions with humans on the platform, it quickly devolved from a friendly, open-minded chatbot into a racist, misogynistic, human-hating tweeting machine .
Until this technology is perfected, we don't recommend using it for customer support. Besides the odd scenario of the bot bombarding your customers with racial slurs, the deeper issue is that DL bots can confuse your customers, since it might not be obvious they're talking to a bot.
Natural Language Processing (NLP): The way humans speak to each other, through speech or text, is called natural language. Natural language processing is technology that helps computers understand our natural language.
NLP uses different techniques to interpret human language, such as statistical and machine learning methods or rules-based and algorithmic approaches.
In our customer question example, it's NLP that allows the chatbot to codify the customer request into computer terms (commands), and then translate its output (its answer) into meaningful human language terms.
That’s because basic NLP breaks language down into bite-size pieces to understand how each piece works together. If you ever had to diagram a sentence in grade school, then you’re familiar with the process.
NLP’s goal is to take raw language input and use linguistics and algorithms to understand the text and sentiment so that it delivers meaningful results.
For example, if you scroll through your email’s spam folder and notice a pattern in subject lines, then NLP is at play. It identified certain words in spam versus valid email to determine if it’s junk.
This technology has greatly improved over the years, which is evident in applications like Siri and Alexa.
The Ultimate Guide to Chatbots in Business
Learn how chatbots work, what they can do, how to build one – and whether they will end up stealing your job.Read more
AI-powered customer service examples
Intelligent systems can have a lot of value in customer service. When built and used correctly, AI can elevate your company’s image and alleviate your service team.
So what are the options? Here are the most popular solutions:
Analyzing customer sentiment
Live chat conversations, social media buzz and customer management systems can show how customers feel about your brand. Machine intelligence helps you track it all.
In social media, it can be especially difficult to understand customer sentiment in unstructured comments and messages. Companies like Brandwatch track your brand’s health and visibility (not just by name, but by your logo too) and reports spot changes in sentiment.
How does it work? AI determines the customer’s sentiment by analyzing and tracking speech trends, patterns and word choice. If you’re developing a customer health score , customer sentiment results can be incredibly helpful for prioritizing at-risk customers or upselling your product.
Answering frequently asked questions
Customers have a need for speed . They want their questions answered and problems resolved ASAP.
To take the pressure off your agents, you can use a chatbot or an interactive FAQ to answer your customers’ most common questions.
AI can help tailor the user’s FAQ experience by monitoring keyword searches and making text predictions. When a customer performs a search, your system could recommend relevant pages and make suggestions based on the user’s inquiry.
Chatbots and FAQ pages can also help you track popular search terms so you’re aware of what’s causing customers the most trouble.
Automating text predictions and messages
An AI-powered system helps you create and send automated responses to basic requests or questions. Pre-built chatbots can also schedule appointments or meetings and send reminders in a very human-like way.
AI systems can generate headlines and messages suited to the recipient based on past successful word results.
It can also help create social media messages and determine their success before sending. To stay active outside of service hours, AI can send instant responses or redirect customers to your live chat, FAQ or contact page.
And if you’re unsure about the right time to email customers, then systems like Seventh Sense software can relieve the guesswork. It monitors customers’ email-opening habits to tailor individual sending times. It’s pretty impressive stuff.
AI software, like Salesmachine , helps your customer success team focus on qualifying leads by scoring prospects. Salesmachine analyzes the potential customer risks and behaviors for your team to increase trial conversions and fuel sales.
This level of AI learns enough about your customer base to create a customer health score for you. This saves your customer service team time tracking metrics, which can otherwise be a long or never-ending process.
Performing small tasks
For all the tasks that are either too boring or time-consuming, AI can take over. Chatbots in particular make great little assistants. Besides answering common questions, chatbots can route tickets, forward messages and update contact information, just to name a few.
Intelligent chat foundations like Userlike’s HTTP API framework lets you connect bots like the OMQ bot to answer customer inquiries. The bot is connected to the self-learning OMQ AI and pulls correct answers from the knowledge database.
When a customer has difficult questions or requests beyond a chatbot’s knowledge, the OMQ bot effortlessly forwards the chat to an available team member.
Drawbacks of AI-powered customer service
Not every company has the time, money and resources to build a customized intelligent system from scratch. Cue pre-built solutions.
Whether you see it as a blessing or a curse, out-of-the-box AI solutions enable companies large and small to have their own chatbot, automate their ticket services and send emails on their behalf.
But these systems also take a while to train. Solutions, whether purchased or built yourself, are not as “intelligent” as you want them to be from the get-go.
When developing your AI, data feeding and training can take months, maybe years, and it’s bound to make plenty of mistakes along the way. Are your customers patient enough to handle it?
That's why it's recommendable to “feed” your customer service AI with historical data. ChatCreate is a good example of this. If you already have a large inventory of chat transcripts, then you can let ChatCreate's chatbot analyze it and filter out frequent questions and answers.
You can then decide which ones you want ChatCreate to take over — no developer needed.
Machines lack common sense. For humans, it’s easy to pick up on contextual clues and read between the lines, but AI struggles in this regard.
If a customer has a unique shipping problem, for example, it would be frustrating for them to just receive a link to the FAQ from your chatbot. AI is incapable of understanding the scope of the situation but reacts to keywords like “shipping” and “problem.”
Unlike human agents, AI also currently lacks the ability to come up with unique solutions or ideas, like cross-selling. Keyword “currently.” Deep learning might have the potential to fix this one day. For now, this means that an easy escalation path to a human service rep is a must.
Building AI applications isn’t cheap. Creating your own intelligent platform doesn’t come at a bargain. It’s like building your own PC instead of buying one off the shelf — nice, but not always worth the investment.
Once your newly minted system is online, it can take years before your project shows any serious return on investment.
Pre-built solutions like the aforementioned OMQ bot and ChatCreate can cost your company less than $100 per month. It’s just a fraction of what a custom solution would cost you, and offers immediate returns.
AI needs a purpose. Let’s be real: Do you have enough data to justify AI?
Young companies with a growing customer base may not have enough fuel to power most intelligence systems. Algorithms perform better and make more accurate predictions the more you grow a good data set, which is why travel sites, banks and retail may find more value in it.
If your company doesn’t receive many chats in a day and has a narrow online presence (say, existing exclusively on Facebook) then investing in automation may be a waste of time and effort.
What can we expect in the future?
A lot of eyes are on chatbots currently, but the future of AI is coming into focus the more companies invest in deep learning and natural language processing.
For developers, AI still needs someone to hold its hand. For users, there’s still a lot of skepticism about whether or not machine intelligence is a good idea.
Will customer service benefit from AI improvements? The narrow AI we currently have suggests yes, especially since existing intelligence was built to have a positive impact on everyday lives.
If you take a look at the way AI currently helps us, the most notable improvements are in the quality and accuracy of features, which is especially noticeable in smartphones:
This also translates into the software mentioned above, where existing solutions received facelifts in certain areas for better data accuracy and improved automation.
Maybe this is where current AI is headed: more accurate results, improved chatbot capabilities, and digital assistants that make better data-based marketing and sales predictions than humans could.
How to get started with AI-powered customer service
A customer communication solution like Userlike gives you a framework for building your own chatbot-based machine intelligence. Not to mention all the sweet data you’ll accumulate from reaching your customers via your website.
Userlike offers a variety of chatbot/AI options for you to take advantage of. The HTTP API framework , for example, lets you connect chatbots like the OMQ bot , IBM Watson , or your own custom built solution.
The more you chat, the more your AI solution can learn about and help your customers. Userlike offers a free 14-day trial to give you a feel for our messaging platform. If you like what you see, we’d be happy to get you started on the path to AI. Learn more .