With the AI-powered application, 3M integrates and unifies data from disparate enterprise systems—order, customer, demand, manufacturing, inventory, customer service—to predict expected delivery dates for individual orders and provide accurate promises to customers when they place orders. This type of AI-enabled process makes business-to-business procurement feel more like shopping on Amazon, dramatically enhancing 3M’s customer service by ensuring orders arrive when they are promised. It also has the significant secondary benefit of revealing bottlenecks in the supply and logistics network, enabling further network optimization. Authorized users in 3M can access relevant KPIs across every location globally, which are updated in real time, via an intuitive user interface. They can also use these insights to make important operating decisions, such as ordering additional stock, redirecting orders, and sending out service alerts to potentially affected customers.
A second use case focuses on the “order invoice escalation” process, with the objective of dramatically reducing the number of invoice-related customer complaints. Given the complexity of many purchase orders—numerous invoice lines, cross-border deliveries, multiple applicable discounts and taxes—a significant number of customer complaints are related to invoicing issues. 3M developed an AI-powered application to predict invoices that may cause customer complaints, enabling 3M specialists to review and amend those invoices before they are sent. 3M’s invoicing system previously relied solely on a rules-based approach to flag potential issues and lacked AI-enabled predictive capabilities. The new application employs sophisticated AI methods along with the existing rules and achieves far greater accuracy in identifying problematic invoices. For a company of 3M’s scale and global reach, the resulting benefits—in terms of both reduced costs of investigating invoice-related issues and increased customer satisfaction—are significant.
By 2020, 3M expects the impact of its business transformation initiative will result in $500 million to $700 million in annual operational savings and another $500 million reduction in working capital. That’s over $1 billion—three percent of the company’s 2018 revenue.
United States Air Force: Predictive Maintenance
While many industrial organizations have embraced AI predictive analytics to optimize inventory, protect revenue, improve customer relations, and more, the advantages of digital transformation are not limited to private enterprise.
The U.S. military spends a third of its annual budget on maintenance. Any reduction in that number has profound implications for military readiness—not to mention impacts on available resources, morale, and more. Starting in mid-2017, several internal groups in the U.S. Air Force began to consider whether applying AI to aircraft maintenance would alleviate unplanned failures, increase aircraft availability, and improve the regularity of maintenance schedules.
The Air Force maintains a fleet of almost 5,600 aircraft that are on average 28 years old. The USAF relies on 59 air bases in the U.S. and over 100 airfields overseas. Planes are flown by 17,000 pilots and maintained by thousands of different personnel in many different groups. The factors that contribute to an aircraft needing maintenance, or to the failure of any one of its six main systems (engine, flight instruments, environmental control, hydraulic-pneumatic instrumentation, fuel, and electrical) or subsystems, can range wildly. Base temperatures and humidity, maintenance crew behaviors, flight conditions and duration, and of course, equipment condition and age can all impact maintenance needs.
To solve this challenge, the Silicon Valley outpost of the Defense Innovation Unit (DIU) that is chartered with accelerating commercial technologies for military use worked with several USAF divisions to develop a project for AI-based predictive maintenance. The teams started with the E-3 Sentry (Airborne Warning and Control System, or AWACS) and compiled seven years of all relevant structured and unstructured data: sorties, workforce experience, subsystem interdependencies, external weather, maintenance text logs, oil samples, and even pilot notes.
In three weeks, the teams aggregated these operational data from 11 sources with 2,000 data points for a single E-3 Sentry subsystem to build a prototype. After a 12-week effort, the teams delivered an initial application that used 20 AI machine learning algorithms to calculate the probability of failure of high-priority aircraft subsystems, so maintenance could be done on those systems just prior to failure.
The predictive maintenance application also provided capabilities to optimize maintenance schedules to align with usage and risk, prioritize maintenance across equipment, directly initiate activities through existing work order management systems, identify root causes of potential failures, and recommend technical actions to an operator. The USAF team can now analyze equipment health at any level of aggregation, including systems and subcomponents, risk profiles, operational status, geographic location, and deployment.
FIGURE 9.1
Overall, the initial project improved aircraft availability by 40 percent. Because the initial project was successful, the USAF expanded the project to the C-5 Galaxy and the F-16 Fighting Falcon and plans to make the application for predictive maintenance available to any USAF group or aircraft platform. Once the application is widely deployed, the USAF expects it will improve overall USAF readiness by 40 percent.
As these case studies demonstrate, digital transformation can have a potentially revolutionary impact on large enterprises and public sector entities. The changes go deep into process and culture.
But success in these efforts is contingent on two fundamental pillars: a technology infrastructure able to support these new classes of AI and IoT applications, and direct leadership from the CEO. In the next chapter, I describe the new “technology stack” that digital transformation requires. And in the final chapter, I outline how CEOs can take concrete action and move forward on their digital transformation journey.
Chapter 10
A New Technology Stack
In the four decades that I have participated in the information technology revolution, the industry has grown from the order of $50 billion to about $4 trillion annually. That growth is accelerating today.
I have seen the transition from mainframe computing to minicomputers, to personal computing, to internet computing, and to handheld computing. The software industry has transitioned from custom application software based on MVS, VSAM, and ISAM, to applications developed on a relational database foundation, to enterprise application software, to SaaS, to handheld computing, and now to the AI-enabled enterprise. I have seen the internet and the iPhone change everything. Each of these transitions represented a replacement market for its predecessor. Each delivered dramatic benefits in productivity. Each offered organizations the opportunity to gain sustainable competitive advantage.
Companies that failed to take advantage of each new generation of technology ceased to be competitive. Imagine trying to close the books at a major global corporation without an ERP system or to run your business solely on mainframe computers. It is unimaginable.
A New Technology Stack
The current step function in information technology that I have been discussing has a number of unique requirements that create the need for an entirely new software technology stack. The requirements of this stack to develop and operate an effective enterprise AI or IoT application are daunting.
To develop an effective enterprise AI or IoT application, it is necessary to aggregate data from across thousands of enterprise information systems, suppliers, distributors, markets, products in customer use, and sensor networks, to provide a near-real-time view of the extended enterprise.
Data velocities in this new digital world are quite dramatic, requiring the ability to ingest and aggregate data from hundreds of millions of endpoints at very high frequency, sometimes exceeding 1,000 Hz cycles (1,000 times per second).
The data need to be processed at the rate they arrive, in a highly secure and resilient system that addresses persistence, event processing, machine learning, and visualization. Thi
s requires massively horizontally scalable elastic distributed processing capability offered only by modern cloud platforms and supercomputer systems.
The resultant data persistence requirements are staggering in both scale and form. These data sets rapidly aggregate into hundreds of petabytes, even exabytes, and each data type needs to be stored in an appropriate database capable of handling these massive volumes at high frequency. Relational databases, key-value stores, graph databases, distributed file systems, blobs—none is sufficient, and all are necessary, requiring the data to be organized and linked across these divergent technologies.
Do It Yourself
In the 1980s, when I was at Oracle Corporation, we introduced relational database management system (RDBMS) software to the market. There was a high level of market interest. RDBMS technology offered dramatic cost economies and productivity gains in application development and maintenance. It proved an enabling technology for the next generation of enterprise applications that followed, including material requirements planning (MRP), enterprise resource planning (ERP), customer relationship management (CRM), manufacturing automation, and others.
The early competitors in the RDBMS market included Oracle, IBM (DB2), Relational Technology (Ingres), and Sperry (Mapper). But the primary competitor to Oracle, the one that became the world’s leading provider, was not any of these companies. It was in many cases the CIO. He or she was going to build the organization’s own RDBMS with IT personnel, offshore personnel, or the help of a systems integrator. No one succeeded. After a few years and tens to hundreds of millions of dollars invested, the CIO would be replaced and we would come back and install a commercial RDBMS.
When we introduced enterprise application software to the market, including ERP and CRM in the 1990s, the primary software competitors included Oracle, SAP, and Siebel Systems. But in reality, the primary obstacle to adoption was the CIO. Many CIOs believed they had the knowledge, the experience, and the skills to develop these complex enterprise applications internally. Hundreds of person-years and hundreds of millions of dollars were expended on these wasteful projects. A few years later, there would be a new CIO and we would return to install a working system.
I remember even some of the most technologically astute companies—including Hewlett-Packard, IBM, Microsoft, and Compaq—repeatedly failed at internally developed CRM projects. And after multiple efforts, all ultimately became large and highly successful Siebel Systems CRM customers. If they couldn’t do it, what were the chances a telecommunications company, bank, or pharmaceutical company could succeed? Many tried. None succeeded.
Reference AI Software Platform
The problems that have to be addressed to solve the AI or IoT computing problem are nontrivial. Massively parallel elastic computing and storage capacity are prerequisite. These services are being provided today at increasingly low cost by Microsoft Azure, AWS, IBM, and others. This is a huge breakthrough in computing. The elastic cloud changes everything.
In addition to the cloud, there is a multiplicity of data services necessary to develop, provision, and operate applications of this nature.
The software utilities shown in Figure 10.1 are those necessary for this application domain. You can think of each as a development problem on the order of magnitude of a relatively simple enterprise software application like CRM. This assembly of software techniques necessary to address the complexity of the AI and IoT enterprise problem approximates the union of the commercially viable software development methodologies invented in the past 50 years. This is not a trivial problem.
Let’s take a look at some of these requirements.
FIGURE 10.1
Data Integration: This problem has haunted the computing industry for decades. Prerequisite to machine learning and AI at industrial scale is the availability of a unified, federated image of all the data contained in the multitude of (1) enterprise information systems—ERP, CRM, SCADA, HR, MRP—typically thousands of systems in each large enterprise; (2) sensor IoT networks—SIM chips, smart meters, programmable logic arrays, machine telemetry, bioinformatics; and (3) relevant extraprise data—weather, terrain, satellite imagery, social media, biometrics, trade data, pricing, market data, etc.
Data Persistence: The data aggregated and processed in these systems includes every type of structured and unstructured data imaginable. Personally identifiable information, census data, images, text, video, telemetry, voice, network topologies. There is no “one size fits all” database that is optimized for all these data types. This results in the need for a multiplicity of database technologies including but not limited to relational, NoSQL, key-value stores, distributed file systems, graph databases, and blobs.
Platform Services: A myriad of sophisticated platform services are necessary for any enterprise AI or IoT application. Examples include access control, data encryption in motion, encryption at rest, ETL, queuing, pipeline management, autoscaling, multitenancy, authentication, authorization, cybersecurity, time-series services, normalization, data privacy, GDPR privacy compliance, NERC-CIP compliance, and SOC2 compliance.
Analytics Processing: The volumes and velocity of data acquisition in such systems are blinding and the types of data and analytics requirements are highly divergent, requiring a range of analytics processing services. These include continuous analytics processing, MapReduce, batch processing, stream processing, and recursive processing.
Machine Learning Services: The whole point of these systems is to enable data scientists to develop and deploy machine learning models. There is a range of tools necessary to enable that, including Jupyter Notebooks, Python, DIGITS, R, and Scala. Increasingly important is an extensible curation of machine learning libraries such as TensorFlow, Caffe, Torch, Amazon Machine Learning, and AzureML. Your platform needs to support them all.
Data Visualization Tools: Any viable AI architecture needs to enable a rich and varied set of data visualization tools including Excel, Tableau, Qlik, Spotfire, Oracle BI, Business Objects, Domo, Alteryx, and others.
Developer Tools and UI Frameworks: Your IT development and data science community—in most cases, your IT development and data science communities—each have adopted and become comfortable with a set of application development frameworks and user interface (UI) development tools. If your AI platform does not support all of these tools—including, for example, the Eclipse IDE, VI, Visual Studio, React, Angular, R Studio, and Jupyter—it will be rejected as unusable by your development teams.
Open, Extensible, Future-Proof: It is difficult to describe the blinding pace of software and algorithm innovation in the systems described above. All the techniques used today will be obsolete in 5 to 10 years. Your architecture needs to provide the capability to replace any components with their next-generation improvements, and it needs to enable the incorporation of any new open source or proprietary software innovations without adversely affecting the functionality or performance of any of your existing applications. This is a level-zero requirement.
Awash in “AI Platforms”
As discussed, the technology vectors that enable digital transformation include elastic cloud computing, big data, AI, and IoT. Industry analysts estimate this software market will exceed $250 billion by 2025. McKinsey estimates that companies will generate more than $20 trillion annually in added value from the use of these new technologies. This is the fastest-growing enterprise software market in history and represents an entire replacement market for enterprise application software.
Digital transformation requires an entirely new technology stack incorporating all the capabilities described above. This is not about using structured programming and 3,000 programmers in Bangalore or your favorite systems integrator to develop and install yet another enterprise application.
The market is awash in open source “AI Platforms” that appear to the layperson to be solutions sufficient to design, develop, provision, and operate enterprise AI and IoT applications. In this era of AI hype, there ar
e literally hundreds of these in the market—and the number increases every day—that present themselves as comprehensive “AI Platforms.”
Examples include Cassandra, Cloudera, DataStax, Databricks, AWS IoT, and Hadoop. AWS, Azure, IBM, and Google each offer an elastic cloud computing platform. In addition, each offers an increasingly innovative library of microservices that can be used for data aggregation, ETL, queuing, data streaming, MapReduce, continuous analytics processing, machine learning services, data visualization, etc.
If you visit their web sites or sit through their sales presentations, they all appear to do the same thing and they all appear to provide a complete solution to your AI needs.
While many of these products are useful, the simple fact is that none offers the scope of utility necessary and sufficient to develop and operate an enterprise AI application.
Take Cassandra. It is a key-value data store, a special-purpose database that is particularly useful for storing and retrieving longitudinal data, like telemetry. And for that purpose, it is a great product. But that functionality represents perhaps 1 percent of the solution you require. Or take HDFS, a distributed file system, useful for storing unstructured data. Or TensorFlow, a set of math libraries published by Google, useful in enabling certain types of machine learning models. Databricks enables data virtualization, allowing data scientists or application developers to manipulate very large data sets across a cluster of computers. AWS IoT is a utility for gathering data from machine-readable IoT sensors. Again, these are all useful, but by no means sufficient. Each addresses a small part of the problem required to develop and deploy an IoT or AI application.
Digital Transformation Page 20