AI: The Fifth Generation of Business Intelligence. Is it Worth it for CIOs?

AI (Artificial Intelligence) is a part of our daily work landscape—embedded within our daily work apps, at our desks via communication tools, and woven into our business processes. It has been glorified in marketing and popular media as a nearly an immortal entity, with vast memory, perpetually optimizing returns against the organic, approaching the singularity, ready to take your job (and it is... sorta). However, AI can fundamentally be described as something that "acts intelligently." And at a price tag of $15 - $150 USD per seat for most businesses, is adding a layer of AI to your business worth it? Before we can answer that question, let's take a look at what AI is, how it's packaged as services, and what impacts you can expect from installing it.

The Evolution of Optimization in Business

Businesses have always sought to optimize operations, from the dawn of the assembly line and replaceable parts to the invention of vacuum tubes for large-scale computation, to the development of smaller transistors, integrated circuits, and faster computation methods.

From a programming perspective, we have moved from the difficult-to-write and understand 1st-Generation Languages (Binary Machine Code) through generations of programming languages, including assembly language, then compiled 3GL (third-generation languages), and then 4GL which sits at a higher level of abstraction. Now, we are at the 5th Generation Language (Natural Language) or what could be perceived as Artificial Intelligence. However, Natural language voice recognition technology, as an I/O (Input / Output) device, and its entire functional stack, while advanced, can sometimes fall short in appearing as truly intelligent due to various limitations.

Within this spectrum, there is an increasing focus on “what” needs to be done versus “how” it is done, and on “how people think” versus “how the machine implements it.” The higher the abstraction away from logic gates and transistors, the easier it is for humans to understand.

What is AI?

A more comprehensive definition of AI involves the simulation of human intelligence by hardware machines such as CPUs, GPUs, NPUs, and TPUs. These hardware components, powered by electrical pressure (voltage), traditionally process bit-stored software through primitive logic gates. These combined and sequential gates represent 'god-level' cleverly designed mathematical functions, enabling the processing of simple instructions, classical statistics, and more contemporary approaches like neural networks.

A clearer distinction can be made between classical linear logical programs and the latest NN (Neural Network) models. Classical programs consist of hand-written subroutines and algorithms that follow explicit rules and logic, which CPUs excel at performing computations on. In contrast, modern NN models, which are also initially hand-written and assembled through hand-written code, are "pre-trained" to extract features from data. This creates a model of those features through a learning process akin to creating a generalized keyhole capable of recognizing the topology of a key.

Similar to a wooden model representing a molecule or a mathematical model representing a phenomenon, the model produced from training data represents the features in a body of observations. Although these models are initially designed by humans, NNs learn to recognize patterns through waves of functions over data points to make predictions based on the data they process. Modern hardware, such as NPUs and TPUs, is now optimized for NN-type computation, further enhancing the performance.

Deployment and Effectiveness of AI

In various computational approaches—whether classical logical programs, statistical methods, or advanced deployed models—poor design or irrelevant deployment can result in a work environment that lacks intelligent behavior. Regardless of the method, whether classical logic, simple automated workflows, or contemporary AI, the programs simulate functions and can appear intelligent if their responsiveness is effective.

At the end of the day, it's all about how well you deploy it. Simply "slapping it on" for the sake of using "Artificial Intelligence" and expecting it to do the work for you is still unadvisable. Proper deployment requires careful planning, understanding the specific needs of your business, and ensuring that the AI solution is tailored to meet those needs effectively.

Across the spectrum of methods and results, classical approaches typically produce discrete outputs with clear boundaries. In contrast, modern AI systems, such as neural networks, usually generate probabilistic outputs with confidence levels, providing scores or probabilities rather than definitive answers. This flexibility allows contemporary AI to manage more complex and nuanced tasks, mimicking a broader range of human-like intelligence. Examples include high-dimensional pattern recognition of phenomenological or biological entities or states, reflecting the AI's ability to handle intricate and varied scenarios.

Key Questions for AI Deployment in Business

Q1: When it comes to procuring AI for your business, where do you deploy it?

Q2: Which models are valuable for your business, and which are more of a headache so far?

In most cases, your business will have purchased off-the-shelf ERP (Enterprise Resource Planning) or CRM (Customer Relationship Management) systems that come with new AI-powered features and/or optional upgrade subscriptions for $15 - $150 per seat. For the rare few who are adventurous and looking to boost company value by deploying their own services, you have the advantage of choosing from a wide variety of newer methods for processing data to elevate business intelligence and performance. If you want absolute accuracy on repetitive tasks, then hard-coded automation is the way to go. If you want visibility into hard-to-see patterns in the large sea of data, then employ a model.

The Reality of "Powered by AI"

Before we go further, let’s examine the common product packaging of “Powered by AI.” While this phrase sounds impressive, in most cases, the underlying technology is basic generalized "If This, Then That" type automation, coupled with feature-rich Business Intelligence dashboards and recommended report cards you can add. Often, these solutions may recommend interesting yet irrelevant insights, and in some cases can be a hinderance to rarity and adaptation, as the rules may abhor the novel.

In the landscape of general purpose CRM, ERP SaaS systems, there are only a few classical analysis tools available, which is understandable, as the depths of EDA (Exploratory Data Analysis), statistics, ML and beyond are partitioned off into siloed services. Additionally data preparation and feature engineering remain crucial parts of the journey toward enabling contemporary machine learning, which is still out of reach for many mid-sized companies. Simply buying “AI” does not guarantee the delivery of the perceived promise it expresses.

In general, models that give you between 85% - 95% accurate results can be very assistive in a work environment but may require validation. However, of course, performance levels this low would be unacceptable in a low-tolerance, high-variability environment such as autonomous driving in a dense city, so we're not at point where AI services will have high perception equal to a human. That being said, humans are good at breaching frontiers, and biological consciousness is perhaps the last great final domain, where we will finally get to meet AGI (Artificial General Intelligence). **And if we do arrive at that destination, I'll print this article out and eat it on a zoom call with AGI (because it would have read this article, and demanded my compliance).

Future Hardware

Traditionally, models are processed on CPUs with support from GPUs, however the new NPUs and TPUs are specifically designed for the matrix multiplication needed for neural networks, and are ready to be deployed at the heart of networks on servers and on the edge embedded in devices.

With snapdragon NPUs in embedded devices like your phone, tablet, tv, video game console, laptop, kitchen microwave – developers and bring 5th Generation Natural Language closer to its you. And with nVidia NPUs in server farms, processing of large volumes of data can happen more efficiently than on conventional CPUs and GPUs.

Final Thoughts

At the end of the day, AI is best used to: streamline a functioning process, discover an optimal solution through parameter tuning, or enable the detection of a relevant observation in an overwhelming sea of high-dimensional data that exceeds organic perception. However, "AI" (meaning: "trained models") can have unforeseeable outcomes (usually driven by unbalanced data) resulting in bias and unfair decision-making, that humans may follow.

If the data is unbalanced, the model will provide recommendations sorted by its own bias, which company staff may follow as policy, slowly tilting the business's direction. In essence, this results in AI running the business. An AI’s unforeseen results lead to multiple hidden business decisions along the way, causing a systemic negative trend. So until AGI arrives, keep excellent talent in control of your daily operations with assistive AI tool to quicken delivery of work units.

However, that said any business logic, policy, recommendation engine, cultural allowance by leadership and management OR AI model will drive the business's way forward. So in short, AI will not save your business by default. Good hiring practices, retention of top talent, efficient and active innovation and optimization, and considerate application of AI will give you an edge and help you streamline your business.