The term Enterprise AI (EAI) typically refers to AI that is used to run a business. It does not include AI that’s been incorporated into the products and services a business sells. There are many EAI implementation options, ranging from out-of-the-box tools to full custom models.
Business leaders need to understand the differences between these options before they can determine what level of data science expertise they require, what the resulting competitive advantage may be, and how an investment in EAI will impact the bottom line.
Out-of-box Tools: The easiest way for a business to leverage EAI is to implement fully functional, stand-alone tools and applications that are already embedded with AI. Language translation, speech recognition and map applications are examples of this type of AI technology. Some of these third-party tools are customizable in the same way legacy enterprise software systems can be customized, for example chat bots and virtual assistants. Integration support is typically provided or facilitated by the vendor. There is no shortage of these products, many of which tend to focus on particular industry segments, but they no longer provide real competitive advantage because they’re available to everyone.
Universal Models: Several new developments in Deep Learning – a family of machine learning (ML) algorithms – have led to two of the most exciting accomplishments in data science: 1) machines can now recognize people, places, things and activities in the world around them (computer vision) and 2) machines can communicate in natural language. While these advancements are significant, the current universal models use brute force to accomplish their results. They require huge amounts of training data, and can take weeks and months to train. The data sets used to create these models are considered “universal,” because regardless of what an end-user application is targeted to accomplish the data sets are always the same. For example, a ball is a ball, and “adios” always translates to “goodbye”.
All of the major cloud platform service providers offer inexpensive access to their computer vision and natural language models. A software developer can use a simple interface (API) without any knowledge of machine learning to retrieve results. An image file sent to the model will return the list of objects it contains, and a block of text will return a translation, sentiment, or some other requested aspect of the content.
Some applications require recognition of additional, different or more specific sub-types of objects or phrases. A technique called Transfer Learning has emerged to fill this gap. For example, a car manufacturer might need to differentiate between specific models of cars; a biologist might care about the small differences between related species of birds. With Transfer Learning, end-users are able to append their own unique layers of detail to the existing models. This transfer gives the model the ability to recognize additional objects or details without retraining the entire baseline model.
ML as a Service (MLaaS) Models: There is a lot of disagreement in the data science community about the value of shared machine learning models. At a recent VC event, a panel of AI experts was asked if the MLaaS approach was effective. Two of the panelists, both with many years of experience and graduate degrees in ML, were on opposite sides of the question. One said it was a waste of time, while the other said he had just invested heavily in an MLaaS company.
Since MLaaS models are shared by multiple companies, the ability to optimize them for a given company’s data is limited, resulting in compromised accuracy. Due to this limitation, the only candidates for MLaaS models are the subset of applications that are either very similar between companies or do not require high levels of accuracy. The majority of potential EAI applications within a company do not meet these constraints, and therefore are out of reach for a business that relies exclusively on shared models. The ability to fine-tune and evolve a model over time is similarly limited with shared models.
Despite their drawbacks, MLaaS model services may have a place in the early stages of EAI adoption. In some cases it’s better than nothing, and in others MLaaS can provide valuable insights that speed development of custom models. These services can also help to further cultural acceptance of AI within a company, but their use as a viable long-term solution is limited.
Outsourced Custom Models: Custom models that have been developed by third parties overcome the accuracy and application limitations of MLaaS models. But ML models are never really “done,” and as a result outsourced custom models can stagnate. They can’t be optimized and improved over time because there’s no in-house data scientist that’s intimately familiar with the model and the company, someone who can identify, implement and even anticipate valuable changes. Outsourced models also do little to further development of ML skills within an organization, the absence of which can seriously hinder broader EAI adoption.
In-house Custom Models: Custom models developed in-house achieve the greatest levels of accuracy, and generate the most effective results. The applications of ML are limited only by the availability of enough of the “right” data to feed them. By definition, ML models “learn” and get more accurate over time. Not only do they improve at the original task, but many can also be mined for valuable insights. These insights seed ideas for additional business process improvements, and drive improvements to the ML model itself. Healthy ML projects are constantly evolving and improving. Without in-house expertise to evolve and mine the model, a company can miss out on what is often the most valuable ROI from ML.
Companies that grow data science expertise in-house and establish EAI as a core competency develop the most fluent and accurate ML models. With in-house expertise, ML applications with the greatest ROI for the individual company are the ones selected for implementation. Maybe most importantly, data science is infused into the company culture, accelerating broader adoption throughout the organization.
The prospect of developing in-house data science expertise can be daunting. However, when the understanding gap is bridged, a strategic plan created, and the first win achieved, the benefits become clear and the thrill of witnessing them becomes contagious. Sometimes it’s like having a crystal ball. From the c-suite to the end-users, everyone gets to play. Successful adoption of EAI not only transforms the bottom line, it invigorates the entire organization.