Top Ten Obstacles to Successful Enterprise AI
Source: Mark Montgomery on LinkedIn
“In order to remain competitive, companies need to rapidly transition from ML projects to refined enterprise AI systems.” - Mark Montgomery
Click to enlarge. Image: KYield. Figure 1: Examples of EAI productivity trajectories.
Despite the results in a small group of startups and superstars that are achieving an attractive return on investment (ROI) from AI, delivering double-digit productivity growth, and rapidly increasing market power, the super majority of companies are still reporting low levels of ROI.
The purpose of this article is to provide clarity on obstacles to ROI in enterprise AI (EAI) based on a quarter of a century in R&D and engagement with hundreds of leading organizations over the past decade.
A recent McKinsey Global Institute survey revealed that corporate managers are confident 3 percent productivity growth is now sustainable for the broader economy, which we believe is certainly achievable and much needed. The actual productivity growth for each organization depends on many factors, including industry, adoption of automation, regulation, and market health, but our research and early adopters provide clear indications that a carefully designed EAI system has the potential to sustain compounded productivity increases for the foreseeable future.
Unfortunately, many obstacles exist in organizations that reduce or eliminate ROI in AI investments, or more broadly successful digital transformation. The following ten obstacles have been encountered on many of our engagements with large organizations.
1) Strong executive champions are rare
In this context, I mean a combination of leadership ability, willpower, and technical understanding that results in effective championship, particularly by the CEO. AI requires a steep learning curve, but increasingly represents the new competitive bar. Our experience has left me with little doubt that EAI will be increasingly necessary to survive for the majority of companies across all industries moving forward, so the leadership challenge must be overcome.
During our early engagements with senior management teams ten to fifteen years ago, none had considered AI augmentation for the entire enterprise until we presented the concept to them. It was unsurprising then as it wasn’t yet operational. At one point technical fellows in a leading tech company reported to their CEO that EAI was impossible, which in hindsight may have cost the company its leadership position and hundreds of billions of dollars in market cap. EAI has since been proven and should be among the highest priorities.
Since enterprise-wide AI impacts every aspect of the organization, the initiative should be led by the CEO. The full management team needs to be engaged, but the executive chair and/or CEO are typically the only individuals charged with the breadth and depth of responsibility across the entire organization. While execution of EAI systems must be delegated, and should be monitored closely, decision making and leadership cannot be.
2) Culture still eats strategy
Often has been the case in our interactions when the famous Peter Drucker quote has proven prescient – “culture eats strategy for breakfast”.
An enormous effort has been made to educate society on the science of AI, and most cultures have evolved significantly since wide-spread fear seemed to dominate five years ago. Many hard lessons have been learned due to poor preparation, bad design, persistent myths, and overhyped brands. However, most organizations have nonetheless become aware that failure to adopt competitive AI systems across the enterprise is the greatest risk from the technology.
Learning to partner in an objective manner free from pre-conceived bias is critical in becoming a leader in AI. Even the largest and most successful technology companies have learned to partner with small specialty vendors, which is one of the key differentiating factors between leaders and laggards.
3) Confusing ML/DL projects with AI systems
Most companies followed common advice and started small with machine learning (ML) to learn as they go with small projects, and perform many small experiments. While ML projects can be profitable, they are inefficient and severely limited unless run on top of a refined AI system.
The leaders in AI systems who are rapidly increasing productivity, revenue, and market share did precisely the opposite, building out aggressive systems their enterprises increasingly run on. In every case, the efficiency and accuracy of AI systems are dependent upon data quality (see DMS).
Our research reveals that enterprise-wide AI systems can deliver significantly higher ROI than robotics by augmenting productivity across the workforce. Moreover, the installation process can result in rapid metamorphic transformation for the entire organization, unleashing many other benefits. However, in order to achieve a sustainable trajectory in this new competitive environment, it will require refined AI systems that can provide a competitive advantage at an affordable level. Currently, this goal cannot be achieved by the super majority of organizations with either custom EAI or commoditized infrastructure.
4) Assuming big is better
Perhaps no myth in AI is more damaging to customers and the economy than big is always better, whether relating to scale of data or size of vendor.
While it’s true the big data approach is highly relevant for consumer tech firms like Facebook, the enterprise environment needs deeply structured data and precision data management, which requires vision, elegant system design, strong engineering skills, and specialized knowledge. Well-designed EAI systems are by definition scalable with small experienced teams of top-tier engineers.
Commoditized IT typically costs less than custom technology, but provides no competitive advantage, meaning other factors like market power become more important. In addition, custom technology efforts developed on top of ubiquitous technology can be very expensive. I call this problem the IT commoditization paradox, which favors entrenched incumbents in the near term while threatening the economy in the long-term. Compounding the challenge in the networked economy is that ubiquity has been proven to be susceptible to existential risk. Fortunately, EAI also provides opportunity to automate customization, and good EAI design can sharply reduce risk.
5) Failure to operationalize
Surveys from McKinsey and IBM are consistent with what we found in direct discussions and observations–only a small fraction of AI systems have been operationalized and scaled. Furthermore, only about 25% of companies have adopted AI across one or more business units. Compare these results with leaders who are applying AI across most of the enterprise and it becomes obvious why the majority are experiencing low ROI. However, this trend is not necessarily negative as many companies were holding off while learning, becoming more aware of obstacles, understanding their limitations, and seeking better alternatives emerging from specialists.
The goal of EAI: A Continuously Adaptive Learning Organization
The goal of our systems work in EAI is to achieve a Continuously Adaptive Learning Organization (CALO), providing augmentation and rapid learning in a continuous loop tailored to the needs of each entity. In order to succeed, the EAI system must prompt action based on the most important lessons learned in a semiautonomous manner, thus enhancing competitiveness with specific functions that have the capacity to achieve missions and objectives at significantly reduced risk, and critically important for achieving the highest ROI possible–preventing major crises.
CALO Competitiveness & Sustainability
Click to enlarge. Image: KYield. Figure 2: Simple CALO.
6) Redundancy is prevalent
One of the most common problems in enterprise AI was also common in earlier eras of computing. Redundancy is pervasive, driven by many factors not necessarily aligned with the mission of the organization. A significant number of costly custom systems in both the public and private sectors were designed in part to learn what others have already learned. I estimate the combined cost of waste and redundancy to be around 40% of the total budgets in EAI. While companies should apply AI to carve out unique products and services to maintain a competitive advantage, the majority of functions and behaviors in organizations are universal. The optimal path provides the benefits of universality with efficient, semiautonomous customization.
7) Top-tier talent is scarce
It’s well known that scarcity in top-tier AI science has resulted in a talent war. Compensation packages offered by a few companies for individual AI scientists can surpass the entire annual AI budgets for the average company, and it has generally proven to be a wise investment for a few companies with the largest AI programs.
However, less understood is that top-tier AI scientists are typically focused on a deep specialty, such as deep learning (DL) or genetic algorithms, not large-scale systems, particularly enterprise-wide. Google Research published a paper that captures part of the problem: ‘Everyone wants to do the model work, not the data work.’
A greater scarcity of talent exists at the confluence of organizational management, business strategy, operations, system architecture, software engineering, data physics and AI science. Similar to other types of systems, specific designs, tradecraft and trade secrets on how to optimize and secure those systems is what differentiates the field.
8) Not-invented-here syndrome
We have encountered the NIH problem in a surprising number of large organizations given the level of education on the topic, including the Department of Defense, banking, manufacturing, retail, and insurance. In one market leader, the large internal team pre-determined not to work with any external EAI vendor, regardless of any factor including cost, efficiency, ROI, or competitiveness.
EAI is a highly complex wheel that requires decades of learning to do well. Simplifying EAI to the point of natural integration with human work is quite challenging even for the most experienced. The size of budget and team are not nearly as important as the system design resulting from deep specialty knowledge and tradecraft. The most successful scenarios we’ve observed to date in EAI combine custom internal development in the organization’s specific areas of expertise built on top of external systems by EAI specialists.
9) Poorly designed data management systems pre-optimized for AI
If a company does not have an automated precision data management system (DMS) preoptimized for AI across the enterprise, and very few do, it’s physically impossible to provide an efficient and accurate EAI system. Only a handful of companies have developed a large-scale automated and semi-automated DMS, primarily for logistics, autonomous driving, automatic trading, etc.
The majority of investment to date in EAI has been in cloud infrastructure and consulting, which includes planning, manual data management, labeling, data cleansing, integration, research, and internal custom development. Due to maturation of AI systems and need for survival, Gartner expects hyperautomation to experience the greatest growth over the next few years, whereas IDC expects AI hardware to grow more rapidly.
10) Three little bear budgets
We often find ourselves comparing the Goldilocks fairy tale to AI budgets. Although a few companies waste billions of dollars on AI, running too hot and redundant, the super majority are underfunding AI systems, resulting in significantly increased existential risk. Only a fortunate few have found a Goldilocks budget that seems just right relative to their capabilities and needs. Size is much less important than design, though superstar firms invest the most and tend to realize the majority of benefits. The most common error we find is failure to invest in a refined DMS with the foundational properties necessary to establish and maintain a competitive digital enterprise.
EAI Management Principles
As organizations and policy makers began to issue guidance for AI, including for ethics and regulations, they typically discover they lack the physical ability to execute the enterprise-wide governance necessary. We therefore released our 15 EAI Management Principles earlier this year accompanied by a three-part video series hosted by our board member Dr. Robert Neilson who interviewed me on each of the principles, complete with rationale and implications for each. The 15 EAI Management Principles can help senior management determine whether they have the internal capabilities to compete in EAI, and serve as a guide as they embark on the disciplined journey of EAI.
Conclusion
Obstacles to adoption of competitive AI systems are primarily psychological and cultural, but they are no less damaging than structural and can be disastrous even for market leaders, so they must be overcome. Those who have just started to learn AI in the past two or three years may be too far behind on their own to catch up to competitors who mastered the same lessons much earlier.
CEOs and boards may want to review the history of generational change in technology, which historically favors incumbents who partner with emerging leaders that can provide a competitive advantage during the first decade of transformation, which often determines winners and losers.
Mark Montgomery is founder, CEO and Chairman of KYield, Inc., and inventor of the KOS (enterprise-wide AI OS) and Synthetic Genius Machine (SGM).
Visit KYield’s website.