Beyond the Bot

When it comes to innovation in customer interactions, the KLM chatbots on Messenger are often mentioned as a textbook example. Most companies, however, don’t do much—or anything at all—with dialogue-driven interfaces or the techniques for text recognition and artificial intelligence behind it. It’s a missed opportunity. So much is already possible, and there’s even more up for grabs if you think broader than a chatbot, and can bring technology, user needs and marketing together.

Many companies pick the low hanging fruit first, automating frequently occurring dialogues through question-and-answer services on their website, app or messaging platform. Such simple conversations are easy to script and bring fast profit, but as soon as it gets more complicated you stumble upon restrictions requiring different software from most of the standard bot software being used right now, with advanced forms of text and intent recognition, and complex dialogue structures.

From a bot to a Cognitive Services Platform

At the moment, about 85 percent of goal-oriented chatbot conversations are successful without needing human intervention. And if it goes wrong, because the chatbot doesn't understand the questions, an employee can take over the conversation. But that’s not where the possibilities end. Connecting your bot to systems with customer or other relevant data (CRM, PIM etc.) improves the success rate, simply because there’s more information available to provide the right answer. The bot can look up and interpret data, like product information, availability, or the status of an order or request. It’s also possible to link services carrying out certain intelligent tasks, like recommendations, predictions, sentiment analysis or image editing. That’s how a simple chatbot becomes a Cognitive Systems Platform.

Designing in tech

Through a combination of well known and new technologies, human and machine logic, and data and creativity, we can already realize countless innovative, reliable and customer-centric dialogue solutions. This requires a multidisciplinary team of specialists who think both from the perspective of the user and the technology. By way of example, here are some of the ingredients and possibilities of bot conversation using data and cognitive services.

There are a number of simple principles:

1. We see virtually all customer interactions as a dialogue. The bot is solely a means to conduct it.
2. Customer interaction always has a goal. Customers don’t get in touch to have a fun conversation, and neither do organizations.
3. As in every conversation, the aim is to gather information quickly (what does the customer want?) in order to provide answers based on this.
4. Answers are saved as data or content in systems, or as knowledge in the brains of people, inside or outside the organization.
5. AI can help with analyzing and formulating (smart) answers.

For the example dialogue you see below, we answer these two questions every time:

1. What is the question (goal) of the customer and how do we make it clear?
2. What is the best answer and where is it available?

The example:

You’re a tourist in Amsterdam and need a hotel room. You start a conversation with a travel agent (bot). Sequentially, the bot passes through certain steps and services.

Where are the biggest opportunities and challenges?

Smart Intent Classification and Entity Recognition

Intent classification is a technique that helps extract the customer’s intent from their question. This can be tricky, because customers communicate in different ways and can mention multiple things in one sentence. Natural Language Processing (AI) tools like RASA (open source), API.ai (Google) or LUIS (Microsoft) can be used to analyze complex sentence structures and translate them into one or more customer intents. They also extract factual information (entity recognition) from the text, such as the company name, destination, or shoe size. This enables the user to communicate information efficiently, something that would take much more time with a web form.

Dialogue Management

Most chatbot solutions work with scripted dialogues (‘if-then-else’). However, for more complex dialogues, this lacks flexibility and is hard to maintain. Luckily, the development of (partly) Machine Learning-based solutions is moving fast, and the quality and efficiency of dialogue development is also improving.

Automatic and supervised learning

The reinforcement learning technology of RASA improves the text recognition with every conversation. But to seriously improve the quality of conversations, it’s important that bots are ‘manually’ trained by people through supervision. Using conversations that the bot didn’t understand you can annotate what the user really meant, allowing y a concept or intent to be added, and enabling the bot to provide a better answer next time.

(Cognitive) APIs

We can use the information provided by the user, or the information we already have, to connect with external services for spelling (to correct input), image recognition (for reading the text on the credit card, or connecting a photo of a user to his name), recommendations (like restaurants), consulting availability, or carrying out bookings. Certain ‘cognitive’ APIs can even be used to recognize emotions, so users can be forwarded to a real human in case the conversation doesn’t run smoothly.

Seamless multichannel experience

The bot can send confirmation via Whatsapp of a booking made through the website, which can also be used for making changes later. Next time the same user makes a booking, the bot remembers their data (for example through the caller ID) making dialogue much easier.

Hybrid bot/human solutions

If the bot doesn’t handle the conversation too well, we obviously don’t want to annoy the user. In such cases, the conversation can be passed on to a human service agent. This could be proactive (see cognitive APIs) or when the user requests it. The information already available can then be summarized and forwarded to the agent, making it possible to pick up the conversation seamlessly.

Data-analysis and proactive alerts

Conversations can be started by the user as well as by the bot itself. For example, the bot could start a dialogue to give the user relevant insights or suggestions based on data analysis.

Content

A bot doesn’t necessarily have to communicate in text only. Different channels support the use of images, videos, or interactive functions, like offering buttons or actions.

So, it’s not just a bot

A ‘chatbot’ is a far too simplified name for supporting processes for (dialogue-driven) customer interactions. It’s basically all about building a platform that learns, communicates, and solves customers’ problems by using already—or very soon to be—available ‘smart’ services. These services can be used in newly designed and also existing processes, and all of this is made possible by the modular design of modern platforms with the available APIs.

Combining knowledge of smart technology with knowledge of service design results in a better customer experience. Human creativity and machine intelligence work seamlessly together. So, think in dialogues, and involve designers, technicians and customers, and magic is bound to happen.

This article is originally posted on Emerce (in Dutch).

Reason.ai helps to kick-off and scale

We and our partner Firmshift developed a platform for conversational commerce called Reason.ai. This is a modular AI-powered system that delivers a solid solution for conversational commerce and interfaces. Our goal with Reason.ai is to make it easier to set up ‘bots’ or conversational customer journeys.