Artificial Intelligence: Where Do I Start?
You and your data science team need to thoroughly understand your business before you can successfully target an AI project. Here's what you need to know.
- By Troy Hiltbrand
- June 16, 2017
When most people hear the term artificial intelligence (AI), visions of HAL 9000, Ultron, and Baymax come to mind. For years, popular science fiction has projected a society where AI is a common part of everyday life. The question is whether this level of artificial intelligence exists today and if not, how close are we to it.
Types of AI
Artificial intelligence is commonly classified in two broad categories: strong AI and weak AI.
Strong or general AI is characterized by independently thinking machines that can reason, plan, learn, communicate in natural language, and merge these skills to achieve their own targets. Although this is the goal of many artificial intelligence researchers, this technological singularity is not likely to be achieved for several years.
The second category is weak or narrow AI; it focuses on a single task. This is type of AI is much more common today and is everywhere, from Facebook (identifying people in pictures) to Amazon (predicting who will purchase what and when, thereby optimizing its supply chain to shorten the time between purchase and delivery) to Tesla (incorporating the technology in its self-driving car).
Weak or narrow AI is based on leveraging prebuilt and sometimes-evolving models to solve a single problem. This category of AI can also utilize a combination of multiple models, each solving separate problems that, when joined together, create a seemingly complex intelligent solution.
Creating the Model
The heart of artificial intelligence is the model on which it is built. These models and associated techniques are not new; they have been around for many years. What is new is the significant uptick in resources available to run these models. What used to take significant investment to get started can now be accomplished by almost anyone relatively inexpensively by leveraging shared resources in the cloud. This has expedited the development of AI and allowed the technology to be integrated into products and services that we use every day.
As this concept of model development and deployment is not new, there are established frameworks for how to run an analytics project. One of the most well-known techniques associated with AI model development is the Cross Industry Standard Process for Data Mining -- CRISP-DM for short. This model is composed of six steps:
- Business understanding
- Data understanding
- Data preparation
This framework was developed in the mid-90s as a methodology for data mining and analytics projects, and it can also be used as a repeatable process for developing and deploying models that form the basis of weak-AI systems.
Reaching Business Understanding
The first step of this process is arguably the most critical. It sets the tone and target for all the following steps. As you and your data science team begin your AI project, your first goal is to understand the business, define what you are trying to achieve, and describe project success. Typically, there are two types of business goals associated with an AI project. The first is to provide key inputs to the decision-making process. The second is to fully integrate into a business process to make the process faster, more reliable, and more scalable.
Regardless of the technology employed or the model development algorithms you use, AI that is not aligned with business objectives is doomed to failure or premature obsolescence. Teams that venture out on projects not aligned to business goals might achieve a state of personal fulfillment but will fail to deliver business value and will ultimately be deemed a waste of scarce organizational resources.
To be successful with weak AI, a project team needs to evaluate the end target by asking multiple questions.
- Will the AI model enhance our decision-making process? If so, who is the decision maker who will use the output?
- Will they adopt the knowledge gained from the model or continue to rely on intuition to make their decisions?
- Does the decision maker stand to profit from the integration of this model into the process?
- Will the decision maker be able to make better, faster, and more consistent decisions that will drive business objectives through the use of this model? Can this be quantifiably measured?
The answers to these questions can help you focus on how information can be used to enhance the decision-making process.
If the model will be integrated into a business process, there is a similar set of questions that must be answered.
- Will the model be fully autonomous and implement solutions based solely on the output or will the information be delivered to a downstream decision maker (who will perform successive steps in the process using this information)? What will be the impact on those downstream operators' jobs?
- Who is the ultimate consumer of the good or service whose process is now being optimized by applying AI?
- What will the impact be on this consumer? Faster delivery? Optimized pricing? Greater user satisfaction?
These questions help you define the business impact of implementing a model within a business process.
Articulating Business Value
Another key factor in understanding your business is to fully comprehend the monetization model surrounding the generated information. Does it solely have intrinsic value in making your business more competitive and more effective at achieving your existing goals, or does it open new lines of business where the information can be marketed as a distinct product line to be sold to consumers?
MailChimp, a mass email marketing platform, derives predicted gender and age demographics from its AI model. They provide these attributes as part of their premium service to customers that can utilize this information to personalize their individual marketing campaigns and segment their user base. This is an example of a company that has been able to monetize the output of their advanced analytics model.
Choosing the Right Question to Answer
Finally, once you understand the business benefit, you must face the final challenge: finding the question that the AI model will answer. To understand your data and determine if you have the appropriate data to achieve your business goals, you must fully understand what question you are trying to answer.
If you break down many of the successful advanced analytics models that make companies millions of dollars, you'll find that these firms have defined questions that can be answered at a high frequency and at a high level of reliability and insightfulness.
For example, Facebook's identification algorithm answers the fundamental question of what person is in a picture. For Amazon's proactive supply chain management, the fundamental question is if a specific customer will buy a specific product at a specific time. For Tesla, the fundamental question for their self-driving cars is whether to go, to stop, or to move in a specific direction.
For your project, what questions must be answered? How can the complex questions be further decomposed into more simple and discrete questions that can often be answered "yes" or "no"? This definition will allow your team to progress to the remaining five steps in the framework and ultimately achieve success.
Start Small for AI Success
After you have a complete understanding of your business objectives and can distinctly define the question or questions that you are trying to answer, you are then ready to proceed to building out the weak AI.
Strong AI is still years down the road, so focus today on small tasks that can be optimized with the implementation of focused and targeted AI. Digital business leaders across all industries are finding significant success by applying these fundamentals as the world waits for an eventual future where strong AI will become a reality.