Customer Service & Artificial Intelligence:

Transcription

1 Guide Customer Service & Artificial Intelligence: A Roadmap of Value

2 Artificial Intelligence In Human Communication In January 1984, Steve Jobs of Apple was presenting at a keynote event. On stage, he took their newest personal computer out of a bag. Seconds later, in a tinny robotic voice, the computer said, Hello, I am Macintosh. It sure is great to get out of that bag. Hello, I am Macintosh. It sure is great to get out of that bag. It was the first time most people present had heard a computer speak, and Apple s demo became the industry s first brush with a real-life, natural language user-interface in commercial software. Thirty-three years later, in 2018, a newer company - Google - emerged with another keynote demo. It showcased their Google Duplex voice software interacting with unwitting humans to reserve restaurant tables and book hair appointments. Similar to Apple s demo, most people had never experienced such a seamless interaction between humans and a natural language user-interface. What Apple s demo told the world was of the possibilities of a new consumer technology. What Google s demo exposed was that this technology is now ready to communicate with us in natural language. Of course, widely used products like Alexa and Google Home have already confirmed this in the commercial market. To some, it might seem like the barriers to truly intelligent software have fallen, and that general artificial intelligence is now a reality. It is easy to get carried away with the newscycle. But those of us working at the cutting edge of the technology know that very important problems remain to be solved. There are things AI can do today, and there are things it still cannot. 2

3 What AI Can Do To understand what problems in natural language software are solved and commercially viable today, two concepts are useful: Natural language user-interface (NLUI) Natural language reasoning (NLR) A natural language user-interface is any UI component that allows people to interact with computer software via natural language. This might be written language, spoken language, or signed language. For example, a computer programme that transcribes spoken words to text, or gives voice to written words, could be used as components in an NLUI. Most of the hard problems of NLUI have been solved. Voice-to-text, text-to-voice, and textto-text (e.g. translation) software have achieved human or near-human competence in the last decade. When layered on top of rules-based software, NLUIs are now commercially successful and scalable. What AI Can t Do On the other hand, natural language reasoning is the ability to draw meaning from, reason with, or derive new knowledge or information from human language. For example, a piece of software that takes news articles as inputs, and outputs fresh documents drawing conclusions about news events, would be doing so by reasoning with natural language. It is important to note is that although the hard problems of NLUIs are generally solved, many problems in NLR have a long way to go. Some NLR is commercially scalable, such as sentiment prediction or simple question-answering. But we do not yet know how to train machine learning software to reason with natural language with anything more than minimal complexity. Where This Leaves Us For these reasons, at DigitalGenius, we have learned to harness the best of both these paradigms. Years of experience have taught us what machine learning can do for us in the real world and what it cannot. We have learned to use the best available solutions in NLR with cutting edge NLUIs to automate repetitive human problems in the real world. This approach is at the core of the promise of value our technology creates for our customers. 3

4 Artificial Intelligence at DigitalGenius Our History with Artificial Intelligence Our first experience in building artificial intelligence at DigitalGenius was with scripted and rules-based chatbots, as far back as We deployed these systems to multinational clients, including HSBC and Coca Cola. But we soon learned that the future of NLUI and NLR software would involve machine learning, and this is what we turned to. We went to machine learning conferences around the world and met some of the biggest names in the field - people like Jurgen Schmidhuber, David Silver and Andrew Ng. We looked to UCL, Cambridge and Imperial College to hire some of the best machine learning and computational linguistics talent available. We communicated with the developers of leading deep learning libraries like Tensorflow, Keras and Pytorch, sometimes making our own contributions to these. And all this brought us to where we are today: a veteran start-up in the automation industry, with an obsession with and expertise in a single, focused problem: how to automatically solve customer tickets in the customer support industry. An obsession with and expertise in a single, focused problem: how to automatically solve customer tickets in the customer support industry. 4

5 Our Approach to AI for Customer Support Our product lines are broken into two central use cases: CoPilot and AutoPilot. CoPilot: Helping Customer Support Work Smarter Our CoPilot product aims to help customer service professionals work faster, smarter and more enjoyably. We automatically import trawls of historical data from contact center logs. Then we train our neural networks to predict fields, tags, language used, and other forms of metadata. We also rank repetitive customer messages with a smart neural search engine that matches customer queries to canned FAQ responses. Over time, as we collect more data from agent users, we begin to automate these responses, deflecting tickets, imitating the past behavior of human agents. We train our neural networks to predict fields, tags, language used, and other forms of metadata AutoPilot: Full End-to-End Ticket Resolution Our AutoPilot product aims to automate all aspects of customer service ticket resolution. We connect our clients CRM environment to a mixture of curated AI models, decision tree flows, and external APIs to guide end-users through the most repetitive queries. We aim to automate large fractions of our clients ticket base, leaving agents free to resolve the more complex, nuanced issues that arrive at their desk. These product lines are the result of years of research and development from our engineering and machine learning teams. There were things learned along the way that you can only discover by experiencing them. 5

6 Our Philosophy That journey also helped us develop a core set of research and development principles when improving and expanding our offering. Our engineering philosophy is based on a few guiding principles: Industrial-Grade NLUIs The best tools in the market for stepping between the natural language of users and the logical components of our software stack. NLR Where It s Needed, and Where It Works We use natural language-reasoning technologies where they work and add value. Deep learning is great where simple reasoning is needed: what is the intent of the user? Is their mood positive or negative? What language are they using? What new data do we need from them? For everything else, rules-based procedures can add more value and scale more reliably. Data is Paramount Machine learning programs learn the data they are fed indiscriminately, but the automation industry is far from reaching an AI-friendly standard in data storage and presentation. At DigitalGenius we treat data curation, preparation and monitoring as a fundamental piece in our product pipeline. Talent is Non-Negotiable To stay on the cutting edge, we need people at the cutting edge. At DigitalGenius, we aspire to hire the hungriest and brightest engineers in the field of machine learning. No Divide Between AI Research and Development We decided early on to merge the research and production AI teams. We found separating them (as many in the industry do) makes little sense - conducting research without considering development means few projects make it into the real world. All our AI engineers rotate on projects, sometimes working on green-field research, at other times working with customers to diagnose issues with their models. Everyone on the team now has a wider understanding of our platform, and this informs how we strategize for the future and identify relevant research projects. At DigitalGenius, we treat data curation, preparation and monitoring as a fundamental piece in our product pipeline. 6

7 Where We re Going We have used the best data management technology, our market-leading access to machine learning talent, and our years of experience to build the best customer service ticket automation product. It s producing real value for our customers every day. Yet we also have our eyes on the future. As we become better at building software that can reason with language in complex ways, new innovations will become ready for commercial scaling, and those with the deepest understanding of them will be positioned to lead. So our artificial intelligence engineers are always reviewing and monitoring the latest deep learning research, attending conferences (sometimes publishing in them too), to make sure DigitalGenius stays at the front-line of the customer service automation industry. 7

8 Scaling Artificial Intelligence In Production We Were First We have been doing this for longer than most in the industry, and we have a history of arriving at key innovations before, or alongside, the broader market. Many of the larger tech companies (Google, Facebook etc.) have a large budget to publish and promote their technologies. In many cases, we have seen our in-house research repeated by these companies. Key discoveries include: Scaling Deep Learning in Production In the last few years, Google, Amazon, Facebook and Microsoft have worked on cloud platforms for scaling machine learning models in production. Before all this, at DigitalGenius we built and scaled a platform for training, monitoring and serving deep learning models in the cloud. Developing Our Fast Training and Inference System A few years ago, deep learning models were seen as having huge potential, but the amount of time and compute needed to train and store many of its architectures made it difficult to scale. Our engineers spent a long time optimizing our software and practices to make deep learning viable in live production systems. Some of our discoveries and innovations were repeated later in Facebook Research s Fasttext system. Beating IBM Watson IBM Watson is at the forefront of the AI industry, but in the niche of customer service and simple question-answering, we have proven our mettle. We took one of IBM s published benchmarks, where they had state-of-the-art results in an academic question-answering problem - and we beat their results with our own technology. 8

9 An Overview of Our System There are three broad areas our production pipeline focuses on: Dataset Curation and Storage Model Training Model Inference Dataset Curation and Storage When things go wrong with automation in customer service, it is often because technology users have lost oversight of the datasets their models are learning from. We treat dataset analysis and curation as an integral part of our product. We help our customers understand their data from the point of view of ticket automation, and help them build automationoptimized data models for value creation in the real world. Model Training We have a production-ready platform for training and monitoring deep-learning models. In 2017, it was home to about 100 neural network models; today that number is closer to 800, and growing. Our machine learning engineers use this platform as a basis to build skills. We have dedicated neural network architectures for selecting answers to customer queries, for predicting field values, for predicting the language of a customer query, and for predicting the resolution time of a case. We have built our system to make it easy and scalable for our artificial intelligence engineers to add new skills to the repertoire as they appear on our product roadmap. Model Inference Inference is the process of getting new predictions from live models. All our models in production are serving predictions in real-time to customer service agents and users all over the world. We continuously monitor the latency of our software to iterate the experience of our users. We issue real-time predictions with an average latency of 0.37ms, significantly faster than other NLUI products such as Amazon Lex and Google Home. 9

10 The Value We Create All this work has allowed us to globally scale value into the customer service industry. We are producing tens of gigabytes of customer service metadata that was previously the product of human labor. We are predicting the likely resolution time of new customer service tickets, allowing human agents to prioritize their work. We are saving hundreds of hours of people s time by making their processes quicker via our machine learning products. And we are automatically solving tens of thousands of customer tickets throughout the world, in twenty different languages. We are at the forefront of a technological paradigm shift that is reshaping the global labor force, releasing new value into people s lives. Yet our ambition doesn t stop there. We believe that any industry involving human communication will be transformed by artificial intelligence. We want to play a key role in those transformations. To do so, we will continue to maintain our veterancy, expertise and passion for artificial intelligence technology. We want to carry on the journey that Apple started with their demo in And on the way, we will delight in the moments of discovery and magic that artificial intelligence increasingly creates in the world. 10

11 About DigitalGenius DigitalGenius is the AI platform that puts your customer support on autopilot by understanding conversations, automating repetitive processes and delighting your customers. The platform is powered by deep learning that understands your customers objectives, then drives automated resolutions through APIs that connect seamlessly to your own backend systems. This is the practical application of AI that delivers the concrete ROI you ve been waiting for. The DigitalGenius AI platform is used by KLM Royal Dutch Airlines, The Perfume Shop, Air France and other forward-looking businesses to drive conversational process automation through the use of deep learning. Learn more about how it works at digitalgenius.com. UK 1 Canada Square, London, UK US 180 Sansome St., San Francisco, CA