Recently, I wrote a blog entitled “The only question to ask when evaluating Chatbots and Virtual Cognitive Agents” but as always, there is more to the evaluation story, so I decided to dig in a bit deeper into the AI technologies which make all the difference during live customer interactions.

What AI technologies are required for my particular use-case?

The catalyst for writing this blog entry is a recent Forrester report which compares Virtual Agents and Chatbots solutions across ten criteria. The ten criteria loosely fall into the following categories:

  • 4 Intelligence features: Machine learning, intent engine, dialog management, and natural language understanding.
  • 3 Administrative features: Reporting and analytics, security and administration, and multichannel.
  • 3 Product investments: Roadmap, vision, and revenue.

In each criteria, a vendor could be rated “Differentiated”, “On Par” or “Needs Improvement”.

While administrative features and product investments are certainly important when choosing any solution, they are usually not the main selection criteria. So, I am going to focus on the intelligence features.

Besides getting the most Differentiated ratings of any vendor in the report, Amelia is also the only solution to get a Differentiated rating in all four intelligence features.

Two other vendors got two Differentiated ratings and four vendors got a single Differentiated rating, leaving 3 vendors with no differentiated rating in the intelligence category. I’m not actually trying to diminish the other solutions which all have their unique advantages, but again it goes back to the basic question, i.e. what problem are you trying to solve for the customers. And Forrester actually explains this well for Amelia:

IPsoft’s robust NLU and dialogue management, married to enterprise-grade security, make it a top pick for companies looking to do more than just answer questions.

… “more than just answering questions” – that is the key difference.

Depending on the what problems other vendors are trying to solve, they may not need to excel in all or any of the intelligence categories. However, let me explain why each of these functionalities are key for us in our quest to build a Cognitive Platform capable of working as an Enhanced Human Agent servicing customers through natural conversations.

  • Natural language understanding (NLU): This is a core feature for any virtual agent engaging a user in a conversation. NLU is all about understanding what the user means, not just what they say. This is where disambiguation, co-referencing, negation, etc. all play an important role so that the context of a statement can be derived accurately. Getting the context correct is paramount as it impacts how the virtual agent drives the dialog with the user more appropriately through dynamic natural language generation of clarifying questions, which are not scripted.
  • Intent engine: Asking Siri or Alexa to play a song is relatively simple, all the system needs to do is identify the word “play” and then recognize the name of the song and artist. However, intent engines can’t use simple keyword matching when trying to identify more complex requests, especially when customers are not well versed on the subject. Intent engines are very important since they are often the starting point for contextualized dialogs and processes, and if a virtual agent gets off to a bad start, it’ll have a difficult time recovering the trust of customers as it reminds them of IVRs.
  • Dialog management: Maintaining an intelligent dialog with a customer can be an “art” even if you are a human service agent, so this has always been a critical research challenge for us with Amelia. Making sure that the entire conversation or even past conversations are kept in context because vital information could have been provided in earlier exchanges, which may still be relevant when completing a customer’s request or resolving their issue. Or maybe the customer changes their mind or asks a clarifying question before they can make their decision, so contextual switching within the dialogue is essential to a smooth and pleasant conversation.
  • Machine learning: Smart use of machine learning makes everything better. Continuous improvement via retraining different machine learning models is fundamental to obtain higher levels of accuracy as interactions continue to evolve over time. Machine learning can improve intent engines, natural language understanding, dialog management and other key components of the virtual agent’s ability to engage the customer. It is simply not possible to continue to maintain these systems through human analysis.

These were the AI criteria which Forrester had evaluated. However, for completeness, I would add Natural Language Generation (NLG) and Sentiment Analysis to the list of criteria needed to properly evaluate virtual agents as those technologies are imperative to strong dialog management.

What should be abundantly clear is that this area is complex, multi-faceted and evolving. Here at IPsoft, we will continue to push the limits by setting the right aspirational goals about which problems to solve for our customers. At our recent Digital Workforce Summit here in New York several of them presented their journeys and you can see some highlights here.