The One Practice That Is Separating The AI ​​Successes From The Failures

Anyone who has been following the news on AI in 2022 knows of the high rate of AI project failures. Somewhere between 60-80% of AI projects are failing according to different news sources, analysts, experts, and pundits. However, hidden among all that doom and gloom are the organizations that are succeeding. What are those 20%+ of organizations doing that are setting themselves apart from the failures, leading their projects to success?

Surprisingly, it has nothing to do with the people they hire or the technology or products they use. Indeed, many of the successful AI companies are using the same products and services from the same vendors that the companies with AI project failures are using. Likewise, the organizations with high rates of AI success don’t have some magical team of data science or machine learning unicorns that somehow possess mysterious skills. Many of these successful AI organizations have the same skill sets that the average organization has. So if it’s not team and technology, what could it be?

Stop Treating your AI Projects like App Dev Projects

One of the biggest insights from these AI successes is that they don’t see AI projects as application development or functionality-driven projects. Rather, they see them as data projects, or sometimes even data products. A data project doesn’t start with an idea of ​​what the functionality needs to be, but rather focuses on what insights or actions need to be gleaned from the data in whatever current shape it’s in.

It might seem somewhat obvious to many that AI projects are data projects, but perhaps the AI ​​failures need to understand this at a greater level of detail. What makes an AI system work isn’t specific code, but rather the data. The same algorithms with the same code can be used to generate text, recognize images, or have conversations – the functionality is determined by the training data and the configuration of the system. Therefore, to achieve the desired outcome of an AI project requires focus on data iteration and data-centric methods versus coding-centric methods.

More specifically, the code for a facial recognition application doesn’t actually do facial recognition, but rather the code just sets up the data to train the model and then executes the model once it’s trained. The data determines the functionality when it comes to AI and ML projects. So, if you are supposed to run AI projects as data projects, why are people still making the mistake of throwing developer-focused methods and approaches at what clearly doesn’t have much to do with development?

Agile is Dead. Long live Agile.

The most popular methodology for application development is Agile, which focuses on short, iterative sprints tied to the immediate needs of the business user versus long development cycles. However, Agile falls flat when dealing with AI because it doesn’t tell you how to deal with data, the core asset of an AI system.

Another approach is the Cross-Industry Standard Process for Data Mining (CRISP-DM), a decades old method that guides data mining efforts. However, it’s particularly focused on data projects and lacks some critical elements needed for AI projects, and hasn’t been updated in over two decades. There have been other data-centric approaches, but they don’t provide detail on how to run and manage data-centric projects, but they don’t tell you how to deal with the specific requirements of AI model training and iteration, and they haven’t been built for Agile. This leads AI project managers to struggle with the right approach to run an AI project. No wonder so many AI organizations are making up their own approaches and failing so much at it.

CPMAI Methodology updates CRISP-DM with Agile and AI-specific details

The alternative to Agile is the waterfall methodology that has been around for decades. Like a waterfall, you can begin your project by designing your application, building your application, testing to make sure the application meets the criteria you built for and then deploying the application. The problem with waterfall is, especially for large and continuously changing projects, this process can take a very long time – sometimes upwards of 18-24 months. During this time the project requirements may change, new technology is created, or business needs may evolve past the original scope. This reality of waterfall is what led to the development of Agile as a more iterative approach. However, while Agile has been very successful for software development projects, if you try to use agile alone on data projects you’re going to run into problems.

Take an AI-enabled chatbot for example. With each iteration the functionality doesn’t change, it’s still a chatbot. The new iterations might change the number of words it can understand, add the ability to converse in new languages ​​or increase the accuracy of the model but the functionality of the chatbot remains the same? it is still just a chatbot. Unlike with software projects where it may take twenty iterations to even get to the first functional iteration of your model, AI projects have that functionality from the beginning. Therefore, you also need a data centric methodology to apply.

Taking the right approach to AI project Management

So if Agile doesn’t work well, but we can’t apply waterfall, and if CRISP-DM doesn’t have what we need, what approaches are successful AI practitioners using? A hybrid of these approaches, of course! Agile and data centric methodologies do not compete, but rather they run on different timelines. The agile and data centric methodologies focus on different iterations. The data centric methodology focuses on the data, the agile methodology focuses on the functionality and they run together.

Agile doesn’t tell you how to do things like data preparation, how to understand the data you have or need, how to build a model, retrain a model, and other critical functions of AI projects. This is why having a data centric methodology, going through specific steps in the correct order, and asking these questions in the beginning is essential. Approaches such as the Cognitive Project Management for AI (CPMAI) methodology are blending data-centric approaches with agile methods to produce methods that are more optimized for the highly data-centric, variable nature of AI projects.

Other project management methods have been tested in the space such as Microsoft’s Team Data Science Process (TDSP) and IBM’s iteration on CRISP-DM. However, many organizations have been reluctant to adopt vendor-originated methodologies and turned to vendor-neutral approaches. Regardless of the approach used, whether it’s CPMAI, CRISP-DM with agile enhancements, TDSP, or others, what is setting apart the successes from the failures is, as Louis Armstrong used to say, “it’s not what you do, it’s the way that you do it”. Perhaps as these successes see more publicity, we’ll see a resurgence in interest in methodology to drive AI projects forward with success.

.

By ll07v

Leave a Reply

Your email address will not be published.