Why Your AI Projects Won’t Scale – and it’s Not the Technical Talent Gap
AI is a relatively new technology and proper training and education takes time, so be sure that you have a strategy in place to onboard your entire team to unlock the benefits of AI.
For the last few years, companies have been increasingly experimenting with artificial intelligence (AI), advancing their computing and data analysis capabilities, investing in new technologies, and trying to reap the benefits of all that AI has to offer. We can expect interest and adoption of AI to become even more pervasive in the coming years, as the barriers to entry lower and more AI talent emerges. While many tools are at the disposal of enterprise organizations today, it’s important for businesses to accelerate efforts across their products and processes to ensure they don’t lose a competitive edge.
It sounds simple enough—invest in the right technology and talent and let AI do the work. But having the right infrastructure and people in place is only half the battle. In fact, despite increased interest in, and adoption of AI in the enterprise, 85 percent of AI projects fail to deliver on their intended promises, according to research from Pactera Technologies. So, while AI is poised to become a normal part of all business operations, many businesses still struggle to implement and scale their AI projects.
The real reason AI strategies won’t work or scale, as expected, all comes down to talent—but it’s not just data science talent, as many may think. While the AI skills gap or lack of resources to fund highly-priced AI talent is often at the center of the conversation, it’s more often than not the product, design, and business talent that stalls AI projects from being successful. As important as technical talent is, understanding how AI will work within a product and how it translates to better customer experience and new revenue is just as critical.
In healthcare, for example, we have algorithms that can read an X-ray as accurately as a human can—but integrating them into the clinical workflow is a real challenge. Being able to train and deploy accurate AI models doesn’t address the question of how to most effectively use them to help your customers. Doing this requires educating all organizational disciplines—sales, marketing, product, design, legal, customer success —on why this is useful and how it will impact their job function.
When done well, new capabilities unlocked by AI enable product teams to completely rethink the user experience. It’s the difference between adding Netflix or Spotify recommendations as a side feature, versus designing the user interface around content discovery. Or, more aspirationally, the difference between adding a lane departure alert to your new car versus building a self-driving vehicle that doesn’t have pedals or wheels.
A real instance of the challenges organizations face when implementing and scaling AI projects comes from a recent Google Research paper about a new deep-learning model used to detect diabetic retinopathy from images of patients’ eyes. Diabetic retinopathy, when untreated, causes blindness, but if detected early, can often be prevented. As a response, scientists trained a deep learning model model to identify early stages of diabetic retinopathy in patients from pictures of corneas from eye exams over the past 2-3 years.
While in theory, the trained model was at least as accurate as human specialists, this wasn’t the case when applied to clinics in rural Thailand. There, the quality of the machines were not as advanced as the kind Google had access to for model training. In some cases, there were not rooms in the clinic to perform the exams that were completely dark, as the trained model assumed. Some patients refused to take the test because of trust issues—the nurses weren’t trained to explain why this new test was necessary or patients were scared that a bad result would require them to spend another day going to a hospital for follow-up treatment. The lack of not only infrastructure, but cohesive education for employees, and understanding of practical limitations is a great example of the major gap between data science success and business success.
Successful AI products and services require applied skill in three layers. First, data scientists must be available, productively tooled, and have domain expertise and access to relevant data. While the technology is becoming well understood, from bias prevention, explainability, concept drift and similar issues, many teams are still struggling with this first layer of technical issues. Second, organizations must learn how to deploy and operate AI models in production. This requires DevOps, SecOps, and newly emerging “AI Ops” tools and processes to be put in place so models continue working accurately in production over time. Third, product managers and business leaders must be involved from the get-go, in order to redesign new technical capabilities and how they will be applied to make customers and end users successful.
While there’s been tremendous progress in education and tooling over the past five years on the first layer (education and tooling for data scientists), we’re very early on in tackling the second layer (operating AI models in production). The third layer (design and product management) is far behind and becoming the most common barrier to AI success. Fortunately, these problems can be addressed and corrected with a few easy steps.
Tightening the business-wide AI talent gap comes down to investing in hands-on education. Outside of the classroom and conference halls, professionals from all across an organization must get experience actually working on AI projects and understanding what they can do, and how the technology can push a business forward. AI is a relatively new technology and proper training and education takes time, so be patient, and be sure that you have a strategy in place to onboard your entire team to unlock the benefits of AI for your customers and business.