Archived content

Archive date: 2024-02-09

This content is no longer being updated or maintained. The content is provided “as is.” Given the rapid evolution of technology, some content, steps, or illustrations may have changed.

Today, we no longer question the use of artificial intelligence (AI) to reach or accelerate a digital transformation strategy in an enterprise. According to the IBM Global AI Adoption Index 2022, companies find AI value in various industries and domains, such as IT operations, security, and marketing. In fact, the same study reports that 35% of companies are already using AI in their business and 42% are exploring it.

A first step in the adoption of this technology is using pre-trained AI models, which removes the need to have deep skills, specialization, and data scientists. For many companies, a low-hanging fruit is the use of virtual assistants (chatbots) with natural language processing (NLP) models that cover specific domain or industry content. However, based on our field experience in the AI market, many companies that adopted chatbots became disillusioned. Most likely, you already faced a virtual assistant that doesn't do what you expected from it.

So, what is happening with chatbots? Why are companies and users disenchanted? During our experiences of implementing enterprise virtual assistants, we saw many bad practices, lack of principles, and other adoption gaps that are the root of this problem. In addition to exposing these challenges, we provide recommendations, best practices, and tools to make the adoption of virtual assistants in your company a success.

Anti-principles of the virtual conversational experience

In this section, we present the bad practices, or anti-principles, that normally guide virtual assistance projects to failure and propose ways to mitigate these problems.

Virtual assistant project is assumed to be a traditional development project

Every virtual assistant uses, or should use, natural language processing functions, which include AI or data science mechanisms. Although it is no longer necessary to have a data scientist create mathematical models that support human language recognition, you still must consider the development principles and model tests those specialists use. Specifically, AI models.

We recommend that you leverage today's best practices in data science and not consider your chatbot to be a traditional, process-centric application. Your conversational engineer should take sufficient ground-truth data (in this case, phrases) to train the chatbot's NLP models and reserve phrases that are not used. They must continuously observe the accuracy of the models for potential overlap between user requests. Even as the tools to build a chatbot become simpler, don't forget that machine learning models are at its core, which require training data, validation data, and test data sets to perform well.

Chatbot is not designed for omnichannel operations

In a digital transformation, all front-office and back-office functions are reoriented to meet client needs in the shortest possible time. If a chatbot can't access a conversation initiated in a different channel (omnichannel operations), users face a bad experience, and must repeat their request. The way to solve this problem is by creating, storing, and maintaining a personalized conversational context of all your channels to use for defining and programming the conversational flows that you deliver to your clients. In fact, this information can help you improve the interaction experience and open up scenarios such as cross-selling and proactive attention. Obviously, your virtual assistant must be implemented with technologies that allow simple integration with the different channels (web, social networks, telephone, short message service (SMS), and such)

Chatbot is considered to be a replica of an existing channel instead of a conversational experience

In most failed virtual assistance projects, companies adopt small solutions that are sadly sold as a replica of a web portal channel, which responds to the word "found" in the conversation. For example, the chatbot offers menus that are a copy of the interactive voice response (IVR) or top-menu of a website. Nothing can be more damaging than a conversational experience with that rigidity. To have a successful conversational project, it is important to have a conversation with your customer care owners about the personality of the assistant and how the bot should conduct conversations beyond menu prompts. The virtual assistant is a "super employee" that can be trained to respond in approachable ways, and perform tasks that go beyond what can be done in another channel.

Organizations see the chatbot as a one-time effort

A common problem is that companies see virtual assistance projects as a project that runs once. ("We build it and we are finished.") In fact, the use of AI in business contexts should be seen as a capability where organizational knowledge is continuously added. Unlike traditional IT systems like software applications, AI systems like chatbots increase their business value over time. Therefore, you must train them with what your clients continuously request. Or, to use other (data science) words: to increase adoption of the chatbot model, it is necessary to continuously monitor the model. This not only helps you to avoid a natural degradation of the model, but it also helps you to eliminate possible biases inside the training. When your organization is able to do this, you eventually gain a super employee in your ranks.

Chatbot isn't ready to scale

When you analyze the possible channel interaction formats, some assistants can accommodate the demand for one channel (like mobile channels), but not for another one (like voice channels). In a digital transformation context, platforms must be able to accommodate demand and seasonal spikes across all channels. By using cloud-native resources and multi-channel design, you can help reduce this risk.

Chatbot is designed for a question-answer interaction

The conversational experiences that humans expect from any virtual or real party are not single-response only. The assistant must be able to conduct complex conversations, such as recognizing new conversational paths, pausing the current chat line, and then reengaging until the customer is satisfied. Creating these advanced interactions should be simple and not require advanced programming skills.

Other considerations for success

Adopt a hybrid cloud approach. In the past, organizations faced regulatory restrictions in regard to sensitive information, when they used public cloud solutions in their chatbot. With the advent of hybrid cloud platforms, it is now possible to use the same public cloud technology on premises.
Create flexibility for integrations. The success of a virtual assistant is also measured by its integrations with business systems that help an organization meet customer needs. For example, integrations with customer relationship management (CRM), service management, or order management systems.
Use information repositories. Low-reach chatbots give up with an "I'm not yet trained for this question" response when they face low scores in their language model training. The bot might not know all of the answers immediately, but it should quickly explore other unstructured content repositories to offer a response that is closer to what the customer is seeking. For example, PDFs, document repositories, and existing documentation.
Provide regional language support. To start quickly with a virtual chat project, particularly for voice-based projects, it is critical that your platform already includes training in the languages that your clients use.
Focus on the interactions with highest volume. The success of your assistant is also measured in savings that are generated for your business. Chatbots should prioritize the interactions that occur the most.
Don't forget your back office. Going further than virtual assistance, your organization should foster a culture of AI integration in all processes, not just in the front-office. Employees, agents, or executives in a branch should also benefit from conversational experiences to deliver results faster and to standardize business knowledge, which is often dispersed in large organizations.
Set up an AI center of excellence. To facilitate the previous points, we recommend that you turn your AI implementation into a business practice, with functional areas that cultivate conversational specialization, continuous model training, and customer care with AI.

Conclusion

For many companies, AI adoption starts with a chatbot. However, these systems are not successful when they are mistakenly conceptualized and best practices are not applied. In this article, we reviewed several ways to avoid failure, such as using a multichannel approach, conceptualizing your project as an AI project instead of a traditional, process-centric application, and focusing on scalability. We also commented on the necessity of continuously maintaining your virtual assistant models and not considering it as a "build and done” effort. From a technology perspective, you must pay attention to your unstructured content repositories to extend the corporeal domain of your assistant, and search for platforms that facilitate integration with your existing services. Finally, it's important to also consider the experiences that your employees require and to create an AI center of excellence.

Learn how to get started with the virtual assistance and knowledge management platforms of IBM watsonx Assistant and IBM Watson Discovery to create unique conversational experiences with little effort. They are designed for hybrid cloud environments and are ready to help you achieve multichannel interactions at scale.

Why virtual assistants fail and how to avoid it

Why virtual assistants fail and how to avoid it