It’s Time to Consider, or Reconsider, Implementing Enterprise AI
Beth Partridge
CEO/Founder/Chief Data Scientist - We bridge the gap between data science and business.
Words of encouragement from a former Intel executive turned data scientist.
In 2012 Harvard Business Review declared Data Scientist “the sexiest job of the 21st century.” I was skeptical, having spent nearly 20 years in the naturally data-driven world of semiconductors at Intel. But the more I learned the more excited I got. There’s an inflection point in the data science learning curve when you realize that this technology really can do things that intuitively seem impossible. I started seeing game-changing possibilities for AI in every functional area I’d ever worked in. I decided to pursue a master’s degree in data science, with dreams of assuming a CDO role somewhere and bringing big data to a big business.
When I eventually turned my attention to the status of data science adoption and the job market, my excitement waned. By 2017 most larger businesses had tried and failed in their initial attempts at applying Enterprise AI. Business leaders were soured by what seemed like empty promises, and data scientists were frustrated by leadership that, from their perspective, was squandering all their hard work.
Having spent half my career in technical roles and half in management, I could see both sides of the problem. I could also see that breaking down the barriers between these two “camps" would take more than desire and some workshops. Success with Enterprise AI would require exceptional levels of ongoing communication, coordination and collaboration between business and technology teams throughout an organization. So I changed tack. Instead of seeking a role focused on a single business, I founded milk+honey, a consultancy committed to helping bridge the gap between business and data science, and accelerating successful adoption in as many businesses as possible.
Now here we are, five years and a global pandemic later. Adoption is still stalled at 25 to 30 percent, with most success stories focused on the so-called “data-natives” that were conceived on a foundation of data-driven products and business processes.??
There has, however, been significant progress in recent years. The technology itself has matured, and early data science talent constraints have been alleviated by a healthy ecosystem of online, industry and university data science training and education. Many companies have now staffed data science teams and successfully created at least one machine learning model. That said, only about half of those models end up being implemented. And therein lies the challenge. To effectively implement Enterprise AI, leaders must make a fundamental, wholesale shift in the way they approach and run their businesses.
Of course, before any leader is going to make that kind of commitment, they’ll need to understand both what they’re getting into and why they should bother. They need a functional understanding of the underlying technology, and a no-looking-back appreciation of its power.
I’ve spent the last few years developing learning and training content to meet this need. I’m going to present it here in LinkedIn in five installments. Everything will be framed in terms of business processes and outcomes that any business professional can understand. I pledge not to use the a-word (algorithm) even once, except, obviously, in this sentence. Here’s how the series breaks down:
1. What is Enterprise AI? The remainder of this post will focus on clarifying what Enterprise AI actually is.
2. How does it work? Machine learning models (the engines that drive this technology) can be incredibly complex and tricky to develop. But business leaders need only understand them from a functional perspective: what goes in and what comes out. At this level they are shockingly simple. This post will demystify machine learning and give you the foundation required to actively and confidently play your part in the definition, development, and implementation of ML-based tools for running your business.
3. Why is Enterprise AI so powerful? The full power of ML is not immediately apparent without a baseline understanding of how it works. When appropriately staffed and implemented, ML projects naturally spark an ongoing, self-perpetuating cycle of innovation that will transform your business if you let it. I’ll demonstrate this power by comparing ML and non-ML based solutions for a few common business processes. At some point while reading this post, you will reach a critical inflection point: an “ah-ha” moment when you fully comprehend why successful implementation is a business imperative. It will seem crazy to do anything without ML moving forward.
4. Why is it so hard to do? Many of the early implementation hurdles are similar to those faced with previous technology adoptions and transformational change initiatives. But this is not a typical next-generation advancement. ML-based analytics exist in a different operating dimension. The rest of the company must move into the new dimension along with it to be successful. Overcoming that final, most difficult roadblock will require you to make a fundamental shift in the way you approach and run your business. I’ll use a four-stage adoption model to explain the typical phases companies go through on their adoption journeys to enable you to pass through them more quickly, or bypass them altogether.
5. How is it done successfully? Every company is starting from a slightly different place so there is not one universal solution for success. This transition will impact every single employee in one way or another. There are four distinct elements that must be addressed in lockstep for it to all come together. We’ll go through each of these areas in sufficient detail for you to come away with a clear understanding what you have to do, the skillsets you will need to do it,?and where to start.
I welcome any and all feedback, comments and especially suggestions for improving this content. DM on this platform, or email at [email protected]. Onward!
Part 1: What is Enterprise AI?
Summary
Your degree of confusion with data science-related terminology will largely depend on when you were first introduced to the topic. I’ll briefly run through the evolution of terminology in the business press from the beginning, then use a simple analogy to clarify meanings. You will once and for all have a useful understanding of the key terms and how to interpret them for your purposes going forward.?
Most of the business world’s introduction to data science was that infamous Oct 2012 issue of the Harvard Business Review called “Getting Control of Big Data”. The term Big Data was commercially coined by O’Reilly Media in 2005. Big Data debuted on Gartner’s Hype Cycle of Emerging Technologies in 2011.
The promise of “invaluable insights” waiting to be discovered in piles of data was one of the hottest topics in business for a couple of years. Data was declared to be the new oil, the new gold, and my personal favorite from IBM, the new bacon. Many large companies rushed out to have their data analyzed.
Most of the early attempts were abject failures. Not because the technology itself didn’t work, but because of the way the projects were defined. It is nonsensical to look for insights in the absence of a question or problem that you are trying to solve. It would be a bit like giving someone keys to the library of congress and asking them to share the top three insights. Big Data went from hot topic to hot potato overnight. It didn’t even make it down to the Trough of Disillusionment and just sort of disappeared from the Hype Cycle after 2014. ??
But the data revolution roared forward. It resurfaced on the Hype Cycle and in the press shortly thereafter as Machine Learning, which carried its own challenges and confusion. Then after a breakthrough in a particular type of Machine Learning called Deep Learning enabled machines to “see” and “speak” for the first time, our fascination with robots led to widespread use of the more familiar term Artificial Intelligence (AI), where we are today. The term AI has more baggage than all the others combined, and initially created even greater confusion. But at least we seem to have settled on it for the long haul.
I briefly considered including formal definitions here, but decided they were not helpful for communicating the concepts. The most useful way I have found for explaining them is with an analogy.?
Consider for a moment the conceptual function of any engine. An engine takes in some sort of fuel and converts it to a form of energy that is used to power another object. In this analogy, Machine Learning is the engine, Big Data is the fuel, and Artificial Intelligence is the output. Machine Learning is how you process Big Data to derive Artificial Intelligence.
There is one aspect of the term “artificial” that is worth exploring a bit further. Note that the power output from an engine could be described as artificial in that it is used to accomplish tasks that previously required a person or live animal. We have grown so accustomed to engines everywhere in our day-to-day lives that we don’t even think of it in that context. Such is and will continue to be the case with Artificial Intelligence. Once implemented in a real-world application, we tend to stop thinking of it as AI. We are already there with Alexa/Siri and Google Maps for example. Most people would not identify these as applications of AI. They are just tools we use in our daily lives.
It would be useful at this point to dispel most of your preconceptions about AI and not get hung up on the definitions. From a businessperson’s perspective, these terms are used interchangeably to refer to the same thing: a new class of capabilities that can transform your business.
领英推荐
Two Types of Applications
There are broadly two categories of applications within this new class of Enterprise AI capabilities: Automation and Decision Support. ?These two groups of applications differ not only function, but more importantly in terms of difficulty of implementation and overall long-term business value.
Automation
Automation applications are designed to optimize a business or manufacturing process in terms of cost, performance, effectiveness, or some combination of those. Common current applications in this category include chatbots, biometric identification, Robotic Process Automation (RPA) and the plethora of emerging Natural Language Processing (NLP) tools that do everything from translate to summarize and now even generate text.
I initially resisted calling this first group of applications Automation, in part because that word carries so much negative baggage, but also because very few of the applications in this category are purely drop-in replacements for humans. In most cases, in the process of automation there is also significant improvement in the fundamental capability. An RPA application performs complex data entry functions faster and with fewer errors. An ML model that reads x-rays can detect artifacts not visible to the human eye. Even when the application does simply replace a human in terms of quality, the increased speed and reduced cost make possible applications that were impractical with humans, like Twitter sentiment detection for example.
The early controversy about automation and fears of robots taking over the workforce seem to be subsiding. My one caution here is that you should avoid replacing humans with poorly thought-out customer service apps that contain endless f#%^ing loops at all costs.
Implementation Considerations
Many of the applications in this category became possible when a breakthrough in a particular kind of ML called Deep Learning enabled machines to interact with the natural world for the first time. At the highest level, machines can now “see”, “hear” and “speak” and navigate their surroundings. While the many of the new applications are indeed groundbreaking, the underlying DL models themselves are still in their infancy and could more accurately be described as brute force. Many require huge amounts of data, can take weeks, months or even years to development, and offer the least value in terms of interpretability and insights as compared to other types of models.
The good news is that many automation applications can be developed by modifying, extending or simply accessing existing, publicly available models. Many automation applications can be implemented with little to no in-house data science expertise. Best of all, successful adoption of automation applications is NOT subject to that last daunting roadblock that is the primary topic of the remainder of this series. There is no reason to hold off on pursuing these types of applications if there is sufficient return on investment.?
While the business gains made possible by automation applications can be significant, they will largely be one-time improvements. The real pot of gold is in the second group of new applications for Decision Support.
Decision Support Applications
Broadly speaking, decision support models provide information for making decisions about responses or actions to take. The Map app on your phone is a great example of a consumer decision support tool. It calculates all the possible route times and recommends the fastest one for you to take on your way home. The recommendation models behind most online shopping, content streaming and social media platforms calculate the probabilities of getting the desired response for each possible product/service/connection and recommend the ones with the highest scores for the application to display. Anomaly detection models monitor live transaction streams and classify suspicious transactions for the banks to disposition.
These are just a few examples of applications that we are all familiar with. The same route time optimization approach used by our Map apps could be directly applied to supply chain management, for example. The recommendation models that determine content for your News and social media feeds would work for employee communications. ?Transaction fraud detection has many similarities to cybersecurity.
Without an understanding of how ML models work it’s a little bit more of a stretch to see that in fact, the same general modeling approach used in Map and supply chain optimization models could be applied to any business or manufacturing process where there is a network of nodes with many potential interconnecting paths to choose from. ?Recommendation engines are useful to any business decision that involves connecting elements between two large sets of options, for example customers & products, questions & answers, legal arguments & judges, leads & salespeople. Classifiers like fraud and network intrusion models can flag anomalies, stratify outcomes and bucket populations within groups of anything really – the possibilities endless. ?
The important takeaway at this point is this: Enterprise AI is not a collection of specific, predefined applications and tools. Rather, it is a fundamental capability that can be used to aide in decision making anywhere there is sufficient data to create a model.
Implementation Considerations
In contrast with the automation models described above, decision support models typically require less data, have shorter development times and are much easier to interpret to glean those illusive hidden gems and insights. The difficulty with decision support models comes with implementation.
Take the supply chain management optimization tool for example. It doesn’t do any good for the tool to recommend an alternate route or carrier if the operator is not authorized to make a change on the fly or there are no other carriers qualified for that route. An ad personalization app does no good if the content creation team isn’t providing differentiated content for display. A customer service operator cannot stave off the departure of a customer without some sort of deal or incentive to offer and the authorization to do so.
ML-based Enterprise AI tools are useless without corresponding changes in decision making processes, authority and supporting functions. Many of the required changes challenge traditional operating paradigms and hierarchical power structures. We’ll go into much more detail about these challenges in subsequent posts.
Anatomy of an ML model
Just as there are vast differences between a two-stroke engine on a lawnmower and a jet engine on a fighter plane, ML models come in all shapes and sizes. Some are as simple as a single formula, while others are incredibly complex and require huge amounts of storage and computing power to function. All ML models are implemented as software, most commonly today using the Python programming language. At the highest level, you can think of them like other computer programs.
In practice, ML models must be integrated into the associated business process or application to be used. How that is accomplished – where they are physically located - is determined by the same storage, performance and connectivity constraints as any other application. Some models can be fully embedded in the main application on the end client device, while others are so large they must be housed on a server somewhere and accessed via the internet.
Despite these strong similarities, there are a couple of significant differences between ML projects and other software projects. Traditional software development is usually done in a development computing environment that is isolated from actual production environment in use to protect it from accidental disruption. The development environment is designed to simulate the end production environment as closely as possible. ML model development is a two-step process where the basic model structure is first defined in code, then calibrated or trained using large numbers of historical data points. The training process is iterative in nature, which means that each and every data point is accessed and processed repeatedly until the model reaches the maximum accuracy. This training process is best-case wildly impractical, and often impossible without the computing and storage power provided by many computers working together in what are referred to as clusters.
This computer power is only necessary during one stage of the development process and is typically not required to use the finished models at all. It would be cost prohibitive and wasteful for every company to purchase and maintain its own computer clusters. Luckily for us, four of the earliest adopters of ML technology recognized this need, and the business opportunity, and began offering access to their clusters of computers on a lease basis.
Each of the major cloud suppliers- Amazon Web Services (AWS), Google Cloud Platform (GCP), Microsoft Azure and IBM Cloud- offer highly configurable, relatively affordable leased access to all the computing resources necessary for ML development. It’s worth noting that all the software and development tools required for ML are open source - freely available for most commercial uses.
The second significant difference between ML and other software-based tools is that ML models require ongoing, regular maintenance - even if there are no bugs, changes or further improvements required. Google learned this the hard way with one of their earliest high profile ML projects.?In 2008 the Google Flu Trends utility made headlines when it accurately predicted the spread of the seasonal flu virus a full two weeks ahead of the CDC, giving especially at-risk people time to take potentially life-saving precautions. Google left the utility running but stopped doing regular maintenance on the underlying model. Fast forward to 2013 when there were again headlines, but this time not good ones. The symptoms of the flu that season were slightly different from the previous years. These subtle differences were enough to throw the model by a whopping 130%, with potentially life-threatening consequences.
I included the details of that cautionary tale to make another important point. ML is powerful tool, period, full stop. While most ML models outside of the healthcare domain will not have potentially life-threatening impact, many will have the potential to be life-altering for your employees and/or customers. By design, if you’re choosing the right projects, they should all have business-altering potential. ?As with any other decision in the company you will bear the ultimate responsibility for outcomes, good or bad, intended or not. Early adopters will have rooted out many of the unforeseen gotchas but certainly not all. You will be the one making go/no-go release decisions, risk assessments and decisions about thresholds and trigger points in the models themselves. It’s impossible to make informed decisions without a functional understanding of how ML works and how it has been implemented in your models specifically.?
Coming Soon:
If it were possible to go straight from “what it is” to “why it’s so powerful”, I would. Unfortunately, it’s impossible without a top-level, “what goes in and what comes out” understanding of how ML models work. Watch for the second post of this series, “How does it work?”, in the coming weeks.