AI in Your Contact Center: Understanding the BasicsAI in Your Contact Center: Understanding the Basics

At Enterprise Connect 2018 last month, I attended as many sessions as possible regarding artificial intelligence (AI). What I observed is that while some companies represented in the sessions can provide partial AI implementations within your contact center, none can support the entire breath of what's needed within the customer engagement environment.

In this article, I'll provide a framework of the various AI-based applications so you can better understand which vendors are able to support your contact center AI automation needs.

What Is AI?
Let's start by defining AI. You'll find many different definitions for AI, but the consensus is on a mimicry of human behavior.

For many people, machine learning (ML) is necessary for true AI. ML is the ability of the machine to identify content, actions, behaviors, or intents that aren't clearly understood. You could use statistical models to calculate a probability; however, human guidance or governance in an enterprise is the best way to proceed. If you're considering AI in your enterprise, be sure to understand how you will manage or govern the ML process.

Learning is a human trait, but in some situations ML can be problematic. Consider Tay, the general-purpose chatbot Microsoft pointed at Twitter. Within 24 hours, Microsoft had to pull Tay off line because a few Twitter users taught it some very bad, racist habits. As I mentioned above, be sure you know how the ML process works for AI tools you might select.

The customer engagement corollary is the use of conversation recordings to train a Web chatbot or voice assistant. It would be great if agents handled every customer interactions precisely in the correct way, but this is never the case. When delivering a conversational AI interface with the use of previously recorded interactions be sure you're able to edit out bad behavior and incorrect answers.

When exploring AI, you'll also need to understand cognitive processing and scale.

Cognitive processor refers to the computing platform that supports the AI application. This platform can comprise a single processor or multiple processors, or even a neural network of thousands of processors. Any of these will work fine for contact center use cases, and are able to run any of the applications discussed in this article -- however, scale is an issue.

Regarding scale, a best practice is to start with a limited production implementation -- in the contact center, this means limiting the number of concurrent users your bot can support. You might start with 10 or 50 concurrent users, but not hundreds. The reason is that every use case is different, and you'll have to measure bot performance during the first couple months in order to understand how best to add processing capacity to meet your enterprise's scale.

Further, ML will tend to decrease processor load over time. Running a limited implementation for a couple months will allow processor requirements per interaction to stabilize enough so you can calculate production scale. The use of cloud-based, "leased" processors on a month-to-month or year-to-year basis will mitigate the risk of buying too much processor power.

Useful AI tools within the customer engagement environment break down into the following categories:

Click below to continue to next page

Continued from previous page

Automatic Speech Recognition and Text-To-Speech
ASR/TTS, which provides the ability for the machine to translate speech into text and, conversely, translate text into speech, has been in common use for decades. For example, I implemented my first ASR/TTS solution in 1995, for Johnson & Johnson. For that use case, we trained a machine to recognize 60,000 different component numbers to provide the company better accuracy on phone orders. The bulk of the ML we had to facilitate went into addressing duplicate component numbers and component labels that used sound-alike letters and numbers (B, 3, G, Z...). Within weeks, we achieved 99% accuracy in identifying specific components.

Nuance Communications' Dragon Naturally Speaking speech recognition software, an example of a commercial ASR/TTS tool, has been around since 1982. This article is partially written -- i.e., dictated -- using it. To start, the software uses ML in a concept it calls "training." Basically, it provides you with some documents to read and it learns how your voice sounds as you form words that it already knows. Further, it learns every time you use it, and has the ability to read your emails and stored documents to learn how you write. If you have back or neck trouble from leaning over a keyboard all day, I highly recommend trying it out. But a word of caution: It spells every word correctly, but that doesn't mean it's always the correct word.

Robotic Process Automation
RPA is the use of computer scripts either to move data or process transactions that require multiple computer systems or to create functions within a computer system that aren't native to that system. Enterprises started spending money on RPA a couple years ago. It differs from scripting in several respects, most notably that it uses ML; has a centralized repository of business rules; and, in most cases, has the ability to use NLP to improve its utility.

RPA can be particularly useful in addressing CRM-related work, or extra work, in contact centers. A perfect example is using RPA to eliminate the manual copying and pasting of case notes between CRM and ordering systems. RPA is less prone to error, and allows contact center agents to focus on more rewarding work.

With ML, a robot has the opportunity to improve itself by recognizing new products and processes introduced into a customer engagement environment. When a new product shows up on an electronic purchase order, for example, the robot might not know what it is; however, it will know that the content of the field is part of a taxonomy that identifies it as a product, not a name or address. This makes the ML governance process easier. Further, under the right conditions, the robot with ML can process the purchase order for a new product without requiring its rewrite as a script.

A repository of business rules is useful when multiple RPA processes handle similar data and move this data to similar or identical information system platforms. For example, knowing that a known entity like an address always goes in the same place can ease deployment of new RPA solutions.

NLP comes into play on a regular basis with RPA. For a simple example, consider that electronic purchase order again. Just one typo means there's a good chance that a simple script will fail and the order will require manual processing. NLP has the ability to use "stock" or programmed language elements to mitigate the impact of a typo. Further, you could use NLP to summarize notes or exceptions described in plain English on the purchase order. A bot can accommodate shipping terms, delivery requirements, and special requests, but a scripting tool will only be able to flag an exception for manual processing.

You can find RPA tools from hundreds of vendors, including Blue Prism, IBM, Microsoft, Pegasystems, and UiPath.

Click below to continue to next page

Continued from previous page

AI-Based Analytics
Contact centers have several ways to use AI-based analytics. In the contact center, AI-based analytics usually uses metadata about an interaction, not the interaction itself. On the other hand, NLP-based analytics apply to the content of the interaction itself. In order to set context, NLP comprises two components: natural language understanding (NLU) and natural language generation (NLG). NLU can do the following, but is not limited to: discern intent, disambiguation, terminology extraction, translation, parsing, stemming, named entity extraction, topic segmentation, discern sentiment (emotion), summarize content, and tag content and taxonomy (classification). NLG simply generates language-based responses. (NLP will be the subject of another article.)

The operative decision criteria regarding which to use is: Do you care about what the customer has done or what the customer is doing right now?"

In my opinion, knowing what a customer is doing right now is more important than knowing what the customer did last week, or even just a moment ago. An example is when a bank customer uses an IVR to transfer money, and then presses "0" to reach an agent. In most cases, the IVR will transfer the customer to the call queue dealing with money transfers. However, the reality is that the money-transfer has most likely completed and the customer probably wants to do something else, like apply for a mortgage. Metadata, in this case, has its limitations. An NLU-based routing solution would ask the customer what he or she would like to do next and then route to the appropriate call queue -- or maybe even invoke NLP to process the loan application.

What surprises most of my clients is that they can perform analytics with minimal lag time (some might call this real-time analytics, but it's better described as near-real-time, I think).

Metadata analytics, such as possible with the Altocloud platform (now Genesys), can be performed with sub-second delay. Analyzing text can take a little longer because the machine may not understand the real intent until a recitation is complete. Amazon Lex, Gridspace Sift, and IBM Watson are examples of solutions that can perform NLP analytics.

Sentiment analysis is also of interest to most of my clients, Typically, these analysis are implemented once a conversational interface is establish. The only place I see customer implementing NLP analytics in advance of conversational interfaces is for the study of compliance. An example of this would be for analyzing agents' performance related to statutorily required recitations in the financial business.

Click below to continue to next page

Continued from previous page

Conversational AI Interfaces
The use of NLP to conduct a conversation with a customer is, by far, the most impactful use of AI in a contact center. We're seeing major shifts in labor utilization, with anywhere from 10% to 70% decreases in contact center labor requirements depending on the use case. Decreased customer effort typically accompanies this given that a bot is faster than a human.

Often, these implementations engage all of the technologies discussed above in a seamless, noninvasive way (as long as you don't consider the "crowding out" of your IVR invasive). What I mean by noninvasive is that the bot, Web chatbot, or voice assistant becomes a user on your Web chat, voice telephony, and information infrastructure. Nothing has to break in order to implement either interface.

By integrating the bot as a user on these systems, you provide a seamless method for transferring interactions to the human agents when necessary. The bot knows what it doesn't know, and the bot can detect negative sentiment faster than a human. Under these, and some other conditions, the bot can transfer a customer interaction to a human for resolution.

A cognitive processing platform can use both Web chat and voice interfaces, as well as the new class of voice computing interfaces -- Amazon Alexa, Apple iHome, and Google Home, for example. In general, the interfaces are abstracted from the cognitive processor, but some are limited by the media they use. For instance, you can't send a graphics file to a voice interface. However, you can orchestrate the use of different media in what some call visual IVR. Nuance's implementation for American Airline is a perfect example of this. In this use case, the interaction may start in voice, but the platform can push a URL to a mobile device or email interface to support seat selection. In the American implementation, the URL brings up a seating map for seat selection.

Web chat integrations are usually based on Web Services integration. The only trick is getting the transfer function to work. Most of the larger players have prebuilt transfer functions into their APIs; however, if you use Web chat software with a lesser-known NLP solution, then you'll be writing some code.

Voice interfaces are a little trickier than Web chat. You should be able to reuse your IVR for ASR/TTS, but doing so will typically require some re-engineering and re-configuration -- all the way back into the PSTN. High-quality sound nets high-quality results. If you're using a high-compression codec, then you're asking for a low-performing speech recognition solution. In cases of IVR reuse, the ASR/TTS processor sits behind the legacy IVR. This way you can migrate one new voice NLP automated process at a time with minimal risk.

Alexa-type implementations can provide the same voice computing interface; however, they're architecturally different. In these cases, the device itself, and not the IVR, may host the ASR/TTS. Further, to transfer interactions to a legacy voice system will require use of compatible codecs. These new devices favor the open-standard Opus codec, which is also available on some voice systems, with WebRTC, and in many session border controllers.

By the way, Alexa use isn't limited to Amazon devices. Amazon introduced Alexa for Windows 10 last year, and any place you can run the Amazon app you can run Alexa. Further, you don't have to say "Alexa" all the time. You can tune the interface to respond to your company's name with little difficulty. For instance, you can say: "Open ABC Company product ordering" after you press the Alexa button and/or register the "skill" on your device.

You can use omnichannel and conversational AI together, but planning the dialog flow is very important. You must pay special attention to the failure-case situations that the bot will encounter. Visual IVR also creates some failure-case challenges that you'll need to plan out well in advance of deployment.

All of the major contact center CRM and telephony providers are offering or have a plan for making AI-based solutions available to their contact center customers. Further, more than 1,000 RPA, conversational AI, and NLP analytics platforms are on the market today.

In many ways contact center operators have never had it so good. AI-based automation tools are reducing labor requirements and customer effort. The tools are out there. If you make a good plan and execute you will succeed on many levels and discover things about your business and your customers that will lead to greater success.