Why do 85% of AI projects fail… and how can you make yours have a positive ROI?

Even though connectivity has been talked about for many, many years now, it is still a major concern for most electronics manufacturing factories around the globe. In fact, our Business Development Director Michael Ho wrote an article on this subject last November after his trip to Asia. Today, the question is more: “Why does having the wrong connectivity solution doom your AI project?”

Well, it’s not only about getting data out of your machines/processes anymore. For your AI project to have a positive ROI, your data needs to be normalized, structured and contextualized as explained by our Product Manager, Julie Cliché-Dubois, during her interview with EMS Now. Cogiscan’s Co-NECT platform is one of the most widely adopted solutions in the industry, not only because of the largest library of adapters, but also because of the domain expertise embedded in the platform to make sense out of the factory-wide “jungle of data”, as Julie explained to Eric Miscoll. In fact, many equipment vendors, including ASM, Mycronic and Juki have chosen Co-NECT as their connectivity solution.

The second reason contributing to failed AI projects is most companies start the project with simply compiling the data as-is without considering the organization, or data structure. Data is often stored in multiple places, and is often unlabelled, rendering it totally unusable from a Machine Learning perspective.

In fact, there’s already a term for this problem: “data swamp” which refers to a “data lake containing unstructured, ungoverned data that has gotten out of hand.” In other words, there is no AI project without data, and most certainly no AI project without structure. It doesn’t matter how long you’ve been collecting data, without structure and intent from the beginning, you’ll likely have to start over.

With aggressive cycle times in our industry, many manufacturers can’t risk the latency involved in running AI from the cloud. While some industries find the cloud method feasible, electronics manufacturers utilize a hybrid solution – leveraging the “brain” of the cloud for training, and then embedding on-premise to solve latency concerns.

This hybrid approach is what iTAC had in mind when developing the IIoT.Edge platform. Their approach is simple – collect all required data within a flat data structure. Then, the platform uploads this flat data structure to the cloud leveraging its “collective intelligence brain” to develop and train algorithms. When the algorithm is trained, it is downloaded to the edge (or shop floor) and can regularly stream data back to the cloud. This hybrid model allows the algorithms to be executed on-premise to keep data flow timely and cycle times short.

The final critical element for your AI project to drive meaningful change on your operation is TIME. As mentioned in a Forbes’ article: “Developing and deploying ML models and ensuring that they drive real actions require a considerable time investment and are an ongoing process. It involves multiple stages and many associated challenges,” including the availability of high-quality data for training.

We’ve taken all these considerations into account while creating our AI use cases. Our data scientists have poured hours into build our models and we’ve trained them in real life factories across the globe.