07/08/2022

Why AI Projects Usually Fail And How To Fix It?

Insights

5 min remaining

Introduction

Many projects are led astray by the recurring belief that artificial intelligence (AI) is magical and can create anything from nothing.

This is why the 2019 Price Waterhouse CEO Survey found that less than half of US companies are pursuing strategic AI initiatives. The risk of failure is high. This series examines the most common reasons AI projects fail in companies at the start of their AI journey.

These failures and ways to prevent them are important for your AI projects.

Strategic Failures

Let’s review the lifecycle of an AI project before we discuss how to fail. Download the full-page PDF of our Lifecycle Instant Insight and follow the instructions.

Failure to meet business requirements

Failure to meet business requirements is the most common problem in AI projects. More often than not, business stakeholders ask for AI not because it solves a problem well-suited to the capabilities of AI, but simply because they want to be able to say, “Our product/service/company uses AI!”. Technology for its own sake is almost a sure path to disaster.

How many AI projects are currently meeting the business requirements when we look at the types of problems AI can solve? IBM defines six main use cases of AI:

  • Accelerate research, discovery
  • Enrich your interactions
  • Prevent and anticipate disruptions
  • Recommend with confidence
  • Scale your expertise and learn
  • Reduce risk and detect liabilities

These are simplified into three larger buckets.

  • Save time (reduce inefficiencies)
  • You can save money by putting your talent to use on more difficult problems.
  • You can make money faster and get better results.

We must ask ourselves: Is AI the right set of technologies to solve the problem?

You don’t have to gather requirements in a huge undertaking. For small projects or proofs of concepts, you can often just document the results based on these types of questions.

  • Who is the intended end user for the project?
  • What is the purpose of the project? What are the objectives of the users?
  • Why should the company undertake the project?
  • The project will take place where? This can include a mix of on-premises, private, and cloud systems in AI.
  • What are the best times to demonstrate business results/impact

The lack of business requirements is a major reason AI projects fail. What are your methods for gathering requirements? How do you gather requirements for other types of projects? AI is simply software. If your organization is proficient at gathering requirements for traditional software projects, then there will be a very little learning curve for AI.

Analytical Approach Failures

To deploy AI and machine learning effectively, it is important to fully understand the problem you are solving. Machine learning projects can be divided into two types of problems and two types of data.

These are the two types of problems that machine learning solves for:

  • Classification – We must bring order to chaos by tagging, sorting, and categorizing. This type of machine learning is often called unsupervised learning.
  • Prediction – We must understand what causes something to happen, and then build a model to predict what will happen. This type of machine learning is commonly called supervised learning.

These are the two types of data that machine learning can solve using:

  • Continuous Data: Number data.
  • Categorical data is anything that’s not a number.

This is where AI projects can go wrong. Although you may believe you have an innumerable problem (such as reading and analyzing written content), it could be a numeric one once the data has been extracted and cleaned. This category includes problems such as prediction and sentiment analysis. You often transform categorical data into continuous information, solving a continuous data problem.

Another common error is to assume that a problem can be solved by one type of machine learning, when in fact it could involve multiple steps and an ensemble approach. Let’s return to the sentiment analysis example. We need to convert a bunch of tweets into predictions about which tweets will get the most engagement. Although we think we are solving the prediction problem, that is only one step of the problem. Before we can solve the problem of what makes a tweet engage, we must first solve the problem of turning text into numbers. This is a classification problem.

Machine learning is often used to solve ensemble problems in business. Sometimes, there are many iterations and stages. You will only learn to evaluate the various techniques that are stacked together to solve the problem you are facing through experience and time.

About the author

Kobe Digital is a unified team of performance marketing, design, and video production experts. Our mastery of these disciplines is what makes us effective. Our ability to integrate them seamlessly is what makes us unique.