top of page
  • CCAI

Conversational AI: Three Keys to Getting the Most Out of Your Investment

Updated: Sep 14, 2020

In this post, AI means 'Specific AI', a collection of useful tools and methods such as machine learning (ML) algorithms and the abilities to use massive amounts of data for analysis and prediction. 'General AI', aims to enable a human-like variety of complex tasks using sophisticated processing. For now, only the former is a reliable business tool.

At this point it would be hard to have missed the rise of Conversational Artificial Intelligence (AI) assistants, like Amazon's Alexa and Google Assistant, and specialized chatbots such as BofA’s Erica and Replika. You might not know, though, that all share common roots in service automation for enterprise contact centers from over 30 years ago. More to the point, Natural Language Processing (NLP)—a subset of AI—has been used for speech and text recognition business applications at scale since the mid-1990's. Since then, other tech advancements, for example massive data aggregation and high-speed internet bandwidth, mean NLP plus other AI tools and methods are more widely available.

Graph on a computer screen

Today, due to the popularity of conversational consumer services and the growing awareness of AI, many large or enterprise business leaders have either brought in one of these AI systems or are looking at them. The market for Conversational AI is predicted to grow five-fold in the next five years. The number of providers of these systems—both platforms and professional services—has grown markedly. From stalwarts such as Nuance Communications and Cisco to a plethora of startups, scores of possibilities and considerations fill the industry. As a business leader for your organization, how will you ensure success? (Hint: Vendors and their tech matters. The project process matters more.)

Regrettably, current enterprise-level investment in Conversational AI does not reliably result in the kind of revenue enhancement businesses hoped for. Whether your business needs income growth or cost suppression, a clear return on a six or seven figure price tag just isn’t guaranteed given present practices. What's more, an important promise of AI implementation is ongoing learning from the collection of and insight gleaned from large data sets. As with revenue, the failure to correctly capture and analyze data leads to case studies in disappointment rather than success.

Is success possible?

Return on investment for any enterprise product or service that interacts with consumers starts with two key pillars: clarity of purpose and the customer value proposition. Both of these emerge from answering "why" and "what" questions for the business and its customers. Clarity of purpose combines the reasons to take a certain market approach with specific desired outcomes. Value proposition means offering something that people will trade for their money, time, or data.

Building the first iteration well should prove value and purpose. Customer usage data can then augment next iterations, adapting the product and even the business based on how customers behave before, during, and after the value exchange. Consistent refinement of both purpose and proposition creates a momentum and growth.

Idea > Magic > Profit?

With that understanding, we need to ask, "Why aren’t these AI technology implementations doing better?" There are 3 reason for that.

1 - Specific AI is not yet a well-understood and well-practiced business tool. Companies can use Specific AI in a variety of ways, and the particular approaches, data needs, and outcomes differ with those contextual elements. Implementations often require specialists and intensive analysis that many enterprises don't employ prior to these projects.

2 - Reaching clarity of purpose becomes more challenging when the implementation details and concrete outcomes are difficult to determine. We don't yet have generally available industry metrics about how using AI pays off. Without that, and lacking deep upfront analysis of the opportunity, setting expectations is nearly impossible. Sometimes this leads simplistic and flawed attempts at implementing Interactive AI. We make incorrect assumptions and experiment haphazardly. We frame questions poorly and get misleading answers, all the while thinking we’re getting good signals about our customers.

3 - The value proposition gets fuzzy when how the technology is used isn't clear. Conversational AI systems can't be viewed in the same way as web or mobile applications. We throw around the term “natural” for the interface, but getting to natural is much harder than it might seem. Simply put, creating conversational AI systems is harder precisely because interacting in a conversational way is so easy for us to imagine, but actually quite hard to build well. We don't have to learn conversation as people, but most of us have been using it for decades. We are experts at conversation as a tool because we've been using it many times a day since we were born. We're experts and it feels nearly effortless. The conundrum is, when thinking about creating a Conversational AI, our brain says it should be easy, and it's actually the opposite. We're trying to teach an application to handle a sophisticated, robust interaction method that we do so well, and we've forgotten how long it took us to master.

If it's not magic, what can we do?

Here are the 3 keys to getting the most out of your investment in Conversational AI.

  1. Get Robust Data - This can be the most challenging part of current AI implementations, and the hardest to recover from when done incorrectly. When conversational AI projects start with limited, inappropriate, or contaminated data, or a combination of those, it's like building a house on a faulty foundation. Whether dealing with small data sets, using an inappropriate medium as source, or ignoring bias and poor proportions, no project reaches real success based on bad data. Yet, out of all the problem projects we've observed, bad data is a leading cause of trouble. To overcome this, it is imperative to ruthlessly pursue correct and clean data in the right amounts. Fortunately, today there are more ways than ever to do this. The alternative usually means starting over. ---

  2. Do Effective Design - From defining business goals to crafting experience, design combines the business, technical, and human elements of a project. Done well, the result is a strong and flexible plan for creating a product or service that will perform well on goals now and later. Most frequently, ineffective design fails to set clear and concrete project goals and underestimates the skill and effort needed for good conversational experience. Many projects are started with vague mentions of generating leads or cost savings. At a high level, those are okay as discussion points. They provide unclear and insufficient guidance for making design choices and evaluating performance. Goals should be specific and contextual: 'Reduce the call handling cost by 6.25%.' 'Increase appropriate offers by 24%.' Specificity directly informs design choices and provides performance targets. As to the experience, human language is a system that supplies tools for efficient negotiation of and movement toward purpose. We create vast variety of meaning with a small set of core categories: nouns, verbs, adverbs, connections, and the spice of language, adjectives. To communicate effectively, it is necessary that the combinations are crafted well by a skilled conversation designer. --_----------------------------

  3. Optimize Regularly and Rigorously - "No plan of operations extends with any certainty beyond the first contact," Moltke the Elder tells us. After the data is certified and the design validated, it's tempting to think success is a foregone conclusion. That's almost never true. Most conversational AI projects go into production and immediately encounter situations requiring attention. This manifests often as transfers or escalation to human agents, or as customers abandoning the interaction. From unanticipated user phrasing to inadequate robust conversation repair techniques, even well-designed and engineered apps need refinement and polish to consistently achieve performance goals. The two primary levers to adjust are the language model based on the ML data and the conversational flow. For both, the appropriate project team members analyze how users are responding and where the app fails to handle interactions well. Common changes include: adding to or improving the model, tweaking the order and wording guiding users through the interaction, and better structuring information that the system provides. The methods to do this can be tedious and time-consuming, but they do pay off.

Benefits of The Work

Conversational apps using Specific AI update the aging contact management app paradigm for enterprises. They can also automate lead generation and even sales. To create applications that perform well and provide a good customer experience, you need a rich approach. Apps and systems that do not perform as desired suffer from one or more deficiencies, while strong methods produce high-yield offerings. Today's enterprise Conversational AI requires a virtuous cycle of learning from rich interaction leading to high-quality data going back into learning. From the results of that cycle come the desired customer engagement, trust, and loyalty that will power your business.


Looking to dive deeper with your team on being successful with voice, chat, or multimodal apps? CCAI offers a 3.5 hour workshop for teams looking to accelerate conversational AI.

bottom of page