Source: forbes.com
The buzz about the power of AI to disrupt industries and transform businesses is not hype; it’s real. Just look at what AI-powered FAANG companies have done in advertising, retail, entertainment and other industries. But most “non-digital-native” enterprises have yet to realize the benefits of AI and are facing increasing pressure to do so — and for good reason. In this series, I’ll try to point out common misconceptions about enterprise AI and share what I’ve learned about how successful organizations are dealing with them.
Harnessing the power of data has been a special focus of mine since my tenure as technical lead for the National Center for Data Mining at the University of Illinois at Chicago in the ’90s. Back then, big data was the hot topic, and our work in academia was adopted by thousands of companies who made massive investments to collect, aggregate and process data. Some organizations were very successful and managed to extract significant value from their data, while others saw little (or no) returns on their investments. The same thing is happening now with AI, only the stakes are much, much higher.
AI is far more powerful than data mining, which was primarily used to mine data for insights that can be applied asynchronously to the business. Enterprise AI is ultimately about automating real-time decisions, such as granting a loan or a credit line, blocking a fraudulent transaction, retaining customers who want to cancel their service, trading $1 billion in bonds or buying a tanker load of oil. The potential impacts on enterprise revenue, competitiveness, costs, risk exposure — and reputation — can be enormous.
As I’ve worked with global organizations on their AI journeys over the last several years, I’ve seen many that are struggling with the same issues. Here are five key misconceptions that I encounter frequently:
- Enterprise AI is primarily about the technology.
- Data science is the key to successful enterprise AI.
- Automated machine learning (AutoML) will unlock enterprise AI.
- Managing AI models is like managing software.
- Implementing Enterprise AI requires a massive, all-or-nothing project.
In the remainder of this post I’ll provide some context for the enterprise AI challenge, and then address each misconception in subsequent posts. First, some background and definitions:
What Is ‘Enterprise AI’?
Enterprise AI encompasses the end-to-end business processes by which organizations incorporate AI into 24-7 business functions that are accountable, manageable and governable at enterprise scale. Establishing and managing these processes is challenging both technically and organizationally.
As stated above, AI as applied to the enterprise is generally understood to refer to the use of data and computing to automate business decisions. The expectation is that AI will automatically generate optimized decisions that are at least as good as those produced by humans or by conventional software — and do so much faster, more efficiently and more accurately.
AI automates decisions by processing data through “models” that take in data as inputs and produce recommendations or predictions as outputs. The predictions then feed business applications. Different types of models are used for different use cases — for example, identifying fraudulent transactions, approving credit lines, trading stocks or bonds, spotting customers who are likely to churn, optimizing supply chains, etc. Of course, the use of models in enterprise automation is not new and will continue to drive many applications for years to come.
The models most commonly associated with AI are machine learning (ML) models, which have been shown to be extremely powerful in their ability to produce good predictions in real time. The ML models themselves are created from data, without having to explicitly program the underlying rules. And while ML models are executed in software, they are very different from conventional software. This has significant implications for how ML models are developed, deployed, monitored and governed, including the following:
• ML models are strongly influenced by the software and data used in their development. Even subtle differences between the software and data encountered in the production environments relative to the development environment can lead to unexpected behavior. As a result, it’s critical to keep track of the metadata that describes how each model was created and maintained throughout its life — including the data sets used to train it.
• Unlike conventional software, ML models go “stale” over time and have to be refreshed (i.e., retrained with new data) in order to continue delivering beneficial results.
• For many companies, and especially those in regulated industries, such as banking, finance and insurance, the risk and governance organizations must be able to explain how models make their predictions and to prove that they do so without discrimination or bias.
• Since models have a direct impact on business results, line-of-business managers require visibility into how models are performing against their KPIs.
• Over time, the number of models in use in an enterprise is expected to greatly exceed the number of business applications (that is, many models may be used in a single application).
Think for a moment now about large enterprises, with hundreds or thousands of applications, that depend on thousands (or tens of thousands) of models, each of which must be carefully developed, curated, monitored, governed and maintained, much more dynamically than the applications in which they’re used. This is the setting for the challenge of Enterprise AI.
In the next post, I’ll focus on the first and perhaps most important misconception, which is that enterprise AI is primarily about the technology.
Until then, look at use cases where adding AI to traditional analytics would make the greatest difference in business results, analyze how your organization is structured and consider what changes might be necessary to make greater use of AI. As we’ll see in upcoming posts, this goes beyond the choice of AI technologies and platforms and requires an enterprise-wide view of how to organize to address the challenges ahead.