3 Common AI Risks: How to Avoid Failure When Implementing Enterprise AI

By Dr. Homa Karimabadi

AI Risk vs. Reward

AI has proven to be a disruptive technology with Fortune 500 companies scrambling to implement AI technology. These firms must juggle the risk of competition from more nimble adopters, with the risks presented by AI projects themselves.

In a very short time, AI has made a measurable and significant impact across many industries. AI has also led to the emergence of new service and product lines such as smart home devices like Alexa, driverless vehicle technology, detection and classification products for images, and a myriad of other transformative advances.

This backdrop has created a sense of urgency among many companies to insert AI projects into their roadmap – but also many AI-related risks.

Different sectors and companies are in various stages of AI-readiness, ranging from those in exploratory/assessment mode to those that have their own AI division. The process of AI-enabling a company or a product is not to be taken lightly. The scarcity of skilled AI workers poses a challenge and is well documented.

But the ability to successfully recruit AI talent, either internally or through consultants, does not guarantee success. Unfortunately, there are common and costly mistakes that many companies make in their AI efforts, and large companies are not immune.

The 3 Most Common Risks For Enterprise AI Projects

Time to solution for AI projects can vary

  1. Companies often underestimate the complexity of developing AI solutions and overestimate their AI expertise, leading to:
  2. Poor decisions regarding the proper alignment of AI techniques to business objectives
  3. Ill-defined KPIs and unrealistic lift expectations
  4. Grossly inaccurate assessment of the level of effort, cost, and timelines for developing/adapting AI techniques to the point of having a positive impact on their bottom line
  5. Unreliable assessment of their AI-readiness (required talent, data, infrastructure) adds large amounts of risk.

Hiring AI talent can be tough

As the companies make the decision to pursue AI initiatives, the natural tendency is to repurpose existing personnel with a superficial knowledge of AI, to lead the effort. Another common approach is to hire someone with AI buzzwords in their resume but who lacks experience in algorithm development, and assign them a lofty title, such as Chief Innovation Officer.

The hope is that such people would be able to stand up an internal team of data scientists and/or lead the development of their AI-based roadmap. This hope is, however, misinformed.  Any such role is best filled with someone who has an intimate and in-depth working knowledge of the AI techniques and machine learning, in general.

A list of other common mistakes to avoid follows:

  1. Hiring decisions made by non-experts, precluding proper vetting of candidate AI engineers and scientists.
  2. Hiring a data scientist as the sole AI expert and injecting them into a team of engineers. Data scientists in such cases find themselves in a frustrating position of having to manage the uninformed directives from superiors while fighting a turf battle with engineers who may view AI as unnecessary or find it threatening. This leads to turnover and lack of ability to recruit top talent.
  3. Running a team of data scientists in the same way as an engineering team. The workflow processes and project management are very different in these two worlds.
  4. Letting the engineering team solve problems that are much better suited for AI-based solutions.
  5. Not knowing the strengths and weaknesses of the internal team and when and what tasks to outsource, to complement the team’s competencies.

Promising pilot projects do not gaurantee risk reduction in enterprise AI

The field of AI has advanced to the point that for some applications one can get seemingly promising results by using off-the-shelf systems, train them and produce a model. This has led to the misconception that AI is a magic bullet that can be easily adapted to solve any problem.  However, even in cases where one can produce reasonable results using off-the-shelf solutions, there is often a wide gulf between creating a model as proof of concept and developing a product which is subject to stringent requirements and specs.

This important distinction is not widely appreciated.

For example, a particular AI project may specify that a model should run in real-time, on edge devices, and with very high accuracy. Or the available labeled data used to generate the model may be very limited. In such commonly encountered scenarios, the ability to design and develop custom algorithms that go beyond off-the-shelf solutions is critical and exceeds the capabilities of most companies.

How to reduce risk in AI projects

Company leaders who are serious about developing a viable product that meets the stated requirements should avoid the following AI pitfalls:

  1. Treating the AI-initiative as a side project and expecting KPI-affecting results.
  2. Lacking a proper roadmap and/or commitment by the organization to dedicate the necessary resources to deliver a working product.

In conclusion, more companies are recognizing the importance of incorporating artificial intelligence into their business models and taking action. However, only one-third of such projects are successful, and implementation is perceived as slow. The common mistakes stated above are significant contributors to the high percentage of failed AI projects.

Avoiding these mistakes and ensuring a corporate-wide commitment to the initiative can help organizations overcome development and implementation barriers and ultimately lead them to success.

About the Author

Dr. Homa Karimabadi

Dr. Karimabadi has over 20 years of experience in intelligent algorithms, supercomputing techniques, and strategic conception and planning. He has founded/co-founded a number of startups where he led the development of intelligent algorithms for petascale computing, and knowledge discovery from Big Data with applications to healthcare analytics. Other ventures include deep learning algorithms for analysis of medical images and patient-physician interactive portals.

Dr. Karimabadi has given numerous invited talks in the U.S. and abroad on a variety of topics including the future of computing. He has published over 100 articles in scientific journals on a wide range of topics including numerical algorithms, machine learning, cosmology, space plasmas, among others.