Beyond the Hype: Delivering AI Today

“When can we use this?” is the question echoing around companies' executive suites worldwide. Almost every hyped technology has CEOs asking their CIO and CTO that question, and the latest technology to garner this attention is ChatGPT. Often the motivation for the question is the desire to create the appearance of innovation or the desire to be first.

Software licenses for data lakes were bought at the cost of millions on the “build it, and they will come” basis. The same is true of blockchain-based technology solutions. In many ways, both have under-delivered on their hype.

The interest in ChatGPT is valid. The potential is exciting for science, exploration, or just for fun. The ability of generative AI to create new content such as text, dialog, art, ideas, or music is the realization of the once mythical sci-fi notion of a computer simulating a human that we’ve seen in movies and books since the 1950s.

ChatGPT is based upon the ability to exercise massive learned models of knowledge (Large Language Models or LLM), such as the language of a business or industry domain like financial services, healthcare, government, or software engineering.

So everyone would agree that the potential is vast, and the opportunity is unlimited. So back to our favorite CEO question, “how soon?” We know that companies are already racing into this space. Firms like Microsoft, Google, and Meta have already made multi-billion-dollar bets to even further monetize their search and advertising businesses, in which each of us is their product. They will also use these capabilities to create stickiness for their existing software and cloud businesses.

However, there’s a current downside to ChatGPT. Simply put, it is unpredictable and inaccurate in a way that a consumer or enterprise may not realize or understand. The news is full of examples of ChatGPT dialogs that are racist, sexist, homophobic, or just simply incorrect. Imagine a significant brand deploying a virtual agent that learns to insult its customer or provide wildly incorrect information, albeit information that, at face value, looks authoritative or correct. Imagine a source code generated by ChatGPT that is not functionally correct and bills a customer incorrectly as the outcome of a learned behavior. I don’t know many brands that could risk or tolerate those outcomes. It is going to take real work and time to overcome these issues. For the over-excited CEOs, there are unlikely to be quick wins or short-term innovation marketing advantages they seek. Rather, there will be headaches and reputational-risk.

For those responding to that question from a CEO, I’d like to suggest that there is a re-framing question they can use. “Let’s park ChatGPT for now as a research activity. It isn’t game-day-ready.?However, how about we use enterprise-ready AI to make better decisions for our business? That is something we can do today!”

“Let’s actually use AI” seems like a strange answer until you consider how many firms truly use AI at scale today. The CEO will say, “Don’t we already do that?” For clarity, almost every large firm is exploring AI. However, the two important words in the sentence were “at scale.” Yes, every major firm has groups of data scientists or enthusiasts. Yet when I survey CDOs and Lines of Business technology executives across the financial services industry, few firms have deployed more than a handful of Machine Learning models into production usage. With such limited production usage, the radical potential is still unrealized.

Why the limited success of AI so far??The answer is simple. Delivering AI at scale has inherent challenges that need to be overcome. Fortunately, progress and acceleration are possible with the right technology and skills.

Before considering the “how?”, let’s evaluate the “why?” Financial services businesses make complex and wide-ranging decisions every second of the day. They do so through a combination of rules and statistical models. Of course, by leveraging human decision-making, people look at each situation, recommend an outcome, or make a specific binary (yes/no) or quantifiable decision.

Rather than focus on the millions of decisions being made, let's explore whether there are decision-making patterns. Patterns help generalize and operationalize the application of process and technology capabilities.

• Knowing a customer. Understanding and anticipating their context and needs.
• Choosing the best way to engage or create an experience
• Identifying an opportunity for the firm and its customer
• Understanding what happened before now or what might happen next.
• Protecting from risk
• Detecting anomalies
• Suggesting a next step
• Approving, declining, or referring an action
• Conducting a conversation
• Answering a question or offering guidance
• Fulfilling a transaction
• Optimizing a task, situation, or outcome to create simplicity, streamline, de-risk, or reduce expense.

The list above is simply a first framing and is not intended to be definitive. Even within this list, there is natural and intentional overlap. Reducing the list to its purest form might result in an abstract list that renders the business context and content invisible. For example, protecting cyber-security risks goes hand-in-hand with detecting situational anomalies within events emitted from infrastructure, middleware, and application platforms.

However, the value of creating abstract/higher level patterns is that we generalize the technical approach and limit unnecessary variation and complexity. Many scenarios in the list above can be solved with a single technological and operational procedure of defined steps.

1. Identify a specific decision opportunity. An example could be identifying potential customers to take advantage of a new product capability. Two goals could be to drive customer loyalty and profitability.
2. Identify the data required to support decision-making. The data could exist in many internal or external systems for a business. It becomes essential to know where it is stored, how it is accessed, and what the data fields mean. This is often a manual process in organizations that still need a governance catalog.
3. Connecting and virtualizing access to a data source to inspect, catalog, and consume data.
4. In the case of some vendors, they are mandating the moving of data from a source system into a central repository.
5. Applying controls to data, such as access and privacy guards, ensures that PII is not visible and cannot be re-distributed.
6. Combining multiple data sources, often with “table joins,” to create an aggregated view
7. Cleansing data for format, anomaly, or quality. This is where data scientists spend the most time. It seems wasteful that the most valuable resources focus on basic tasks. Data scientists typically use standard tools and programming languages such as Jupyter notebooks and python. From jupyter.org, “The Jupyter Notebook is the original web application for creating and sharing computational documents. It offers a simple, streamlined, document-centric experience.”
8. Using data sources to execute a model training process. Training a model uses an algorithm to understand the importance or significance of combination characteristics within a data set.
9. Assessing the efficacy and quality of a trained model.
10. Using an analytic capability to understand the potential bias of a trained mode. Bias could be a characteristic of the mindset of a data scientist, for example, the choice not to include specific fields of data. Also, it could be a characteristic of the training data itself. For example, a training data set might consist of information only about men rather than include information about women. In either case, the bias is likely unintentional but also usually undetected. A well-known example occurred with a prominent bank deploying a credit card approval model that was not trained on female data. As a result, women were granted lower credit limits than equivalently qualified men.
11. Deploying a model for testing and production and publishing a callable API
12. Calling a model from an application function via an API call
13. Explaining model decisions.?Suppose a model is used for a regulated function, such as a credit card application or a money laundering decision. In that case, a regulator may need to understand the specific characteristics of data that caused an outcome from the AI model. A regulator will use this to govern the efficacy and quality of decision-making and ensure that firms act socially responsibly.
14. Governing, versioning, and iterating a model with new insights to create a feedback loop that ensures an AI model adapts to further information over time—in the same way, ensuring that a model does not drift from acceptable quality.

While it seems like an arduous number of steps, knowing the repeatable process is a great start. Naturally, a defined process is also a strong candidate for tooling and automation.

IBM pioneered AI capabilities, with teams researching the domain since its inception in the 1950s. AI capabilities are built directly into all our technology platforms to deliver business or operational value today. We don’t just talk about the potential of AI… we make it tangible and actionable for our clients today.

The IBM Data Fabric approach combines data virtualization across almost any source in any cloud or on-prem location, discovery, a self-service catalog, in-place consumption without movement, data aggregation, privacy, and security controls, data science tools for preparation and training, a model training, testing, and evaluation pipeline, model deployment, and model governance. Unsurprisingly, these capabilities closely map to the steps above.

Working in partnership IBM Client Engineering, the IBM Financial Services Industry team can help your business bring these ideas to life and deliver real value with AI in weeks, not months. Challenge us. ?When can we get started? To close with the words I used to open this note, “When can we use this?”

Please do share your comments and feedback. I'm here to learn and help.

Steve Canepa Thad Vorozilchak John Emmert Jennifer McComas Stephen R. Tim Paydos Jeff Schlageter