Framing the AI Applicability Problem in Solutions — 4 Steps to Determine AI’s Use within a Solution Set

Original Post on Medium: Framing the AI Applicability Problem in Solutions — 4 Steps to Determine AI’s Use within a Solution Set | by Sam Bobo | May, 2023 | Medium

Contrary to popular belief, Artificial Intelligence can not solve all of the world’s problems despite the recent surge in popularity of Generative AI capabilities. Too often, with the?extreme rise in hype?of new technological capabilities, organizations blindly rush to ride the hype-cycle wave only to find that, in reality, the technology in question simply does not fit the intended use case.

Public Displays of Intelligence — Where AI Hype Fails

Truth of the matter is that the dawn Artificial Intelligence acceleration has arrived and the question asked should not be “how can I incorporate Artificial Intelligence as quickly as possible?” rather, “what use cases within my organization can Artificial Intelligence create a competitive advantage and augment the capabilities brought to market.”

Early in my career at IBM Watson, I consulted Entreprenurial CEOs and CTOs from detailed product design and implementation through revenue maximization to identify, prioritize, and resolve their clients’ most daunting problems through the use of Conversational AI capabilities powered by IBM Watson (as an aside, this Entreprenurial spirit sparked my career in Product Management).?My consultative work spanned Education Technology, Internet of Things, and Analytics start-ups but received exposure to broader industry implementations of Conversational AI.

What became abundantly apparent through my interactions with these leaders and ongoing involvement within the industry was that a framework was required to assess a presented problem for the applicability of AI capabilities within the proposed solution. After speaking with a colleague of mine on the matter, I realized that laying out this framework within a blog post was warranted.

1. Applicability of Artificial Intelligence

Artificial Intelligence, being a subset of the Machine Learning practice, centers around the concept of learned intelligence. Whether supervised models — models that are annotated with the resultant outcome and learned by a machine — or unsupervised models — those that deduce inherent patterns within the training data, AI problems are data driven problems and typically those in which a machine is required to learn from a plethora of experiences.

2. Select the AI Modality

There are two types of machine learning problems:

a. Structured?— with structured data, data is presented in a tabular format. Take a table in an excel spreadsheet or a SQL database. Structured data is not constrained to tables, rather can take the form of mark-up languages or structured JSON objects, however, typically get converted into the former. With structured data, the intended outcome is some recommendation based on an underlying algorithm. Some of the algorithms include k-nearest-neighbors (KNN’s), random forests, and multilinear regression.

Take, for example, Netflix. In simplistic terms (and generalizing rather than delving into what are likely proprietary details about its algorithm), Netflix categorizes media content — movies, TV shows, documentaries, etc — with a series of tags that describe the content, from genres, actors/actresses, producers, etc as well as more meticulous learned data from the media itself uncategorizable by a human. Upon favoriting a piece of media, watching it (and for what duration), pace of watching (binging versus a single episode), etc, that data gets converted into a classification, often turned into a quantifiable or mapped number, and run through an algorithm to furnish a list of recommendations for what to watch next. The recommendations are clear and interaction is on the UI by general behavior, not communicative.

b. Unstructured?— conversely, there exists unstructured data — spanning audio, natural language, images, video, and more. Dubbed early as “cognitive computing” these systems rely on natural language processing capabilities to interface with the words uttered and written by humans.?For purposes of simplicity, I will focus solely on the NLP aspects over image recognition.?Conversational AI / cognitive computing capabilities are employed in situations where the interface with the end customer requires conversation or interactions. For example, chatbots, interactive voice response applications, command and control (i.e Assistant or Alexa). The system generally needs to interpret human words, determine the underlying intent, and produce a response — either pre-determined or generative.

Given my expertise and experience in Conversational AI capabilities, the remainder of the blog shall assume the user has (1) confirmed that AI is applicable and (2) selected the modality of interaction being conversational via unstructured data

3. Employ Conversational Design and Understand the Desired Outcome

Start by asking yourself the following questions: (1) What is the end user trying to accomplish (single action or list of possible actions)? (2) What information is required to collect from the user to accomplish the end goal(s)? (3) How do I prompt the user for that information in the most elegant way possible?

All of the aforementioned questions are critical to be asked when employing conversational design. Empathizing with the end user (as we’ve?all?been in that situation), design principals inform us that the user seeks to be efficient with his/her/their time when interacting with the system and accomplish the goal as quickly as possible. Therefore, the diction employed within prompts and resulting text should aim to make the interaction pleasant and as quickly as possible. Should the bot fail, how can it fail gracefully and maintain a positive sentiment with the end user.

As a quick note for the above, the conversational design should be cognizant of any data look-ups or modifications to data (accounts etc) and the sensitivity of said interaction computationally.

I would be remiss if I did not mention that the user should, at all times, understand that interaction is within an automated system and its intended use and limitations. This is the primary definition and scope of Responsible AI

Microsoft Responsible AI Standard v2 General Requirements

4. Select Conversational AI Capabilities

3 Generations of Conversational AI capabilities exist in the world today: (1) Intent Intelligence (Rules Based), (2) Conversational Intelligence, and (3) Generative Intelligence

The Three Eras of Conversational AI — a framework for explanation

Depending on the type of response, one employes different conversational AI capabilities which can span one or more “generations” of AI capabilities. For example:

1. Intent Intelligence?— employ intent intelligence (or rules based intelligence) when the AI system needs to maintain extreme precision in its operations either for data collection or a distinct set of information that needs to be gathered by the system.
2. Conversational Intelligence— employ conversational intelligence capabilities when the interaction with an end user can be dynamic in nature within the confines of a pre-described flow. For example, the system might dynamically shift its speaking style based on the sentiment or might have different information to play back to the user using queried information. Furthermore, these systems are typically intent based to map within the confines of the application logic using semantic based training to understand the breadth of information to be captured by the user.
3. Generative Intelligence?— employ generative intelligence when the nature of the information being provided back to the user is net new or should be summarized in a manner easily consumed by the user. For example, a common paradigm is to utilize generative AI to provide a failback option when a system is asked an out-of-bounds question from Conversational Intelligence (with guardrails) or should be summarizing a large text of information such as summary from a document that may not be prescribed within an FAQ or common set of answers for a bot.

Artificial Intelligence is an immensely powerful tool in the modern era of computing. Arguably, many problems will be solved with AI capabilities, but it does not mean that AI should be the only solution within the larger solution set. Think carefully about the approach to solve a problem. More use cases for AI are arising daily and I am excited to see the future of this practice!