Revolutionizing Credit Surveillance (Part II): Does Generative AI Live up to the Hype? Data as the Power behind AI Initiatives

Where are we in the AI hype cycle?

After several years of buzz around generative artificial intelligence (GenAI), S&P Global Market Intelligence has noticed its clients are taking a candid look at their digitisation initiatives in an effort to determine if indeed they are delivering value (see figure 1).

Figure 1: Expectations vs. Reality in the Generative AI Hype Cycle

Re-prioritizing AI and digitalization projects can encompass the following factors:

• Evaluating Costs vs. Benefits: Assess the value of AI and digitalization initiatives against their delivery costs.
• Re-assessing Criticality and Value: Analyze the competitive advantages of new solutions compared to existing ones, ranking their importance to achieve strategic initiatives, and determining their scalability and future-proof capabilities.
• Quantifying Implementation Accuracy and Impact of Risks: Investigate the accuracy of automated solutions, identify any critical information that may be overlooked, and evaluate potential cyber and data security risks associated with these applications.

Suitable, fit-for-purpose data as the power behind the AI engine

Through our interactions with clients and the development of our own AI capabilities, we have recognized that data is a pivotal factor in the success of AI applications. Even though comprehensive data testing typically occurs post-proof-of-concept (POC) review, where the potential value of AI initiatives have been communicated to stakeholders, the significance of suitable, fit-for-purpose data cannot be overstated. Such data, as illustrated in Figure 2, can integrate seamlessly into internal workflows, provide substantial value at reduced costs, and facilitate the visualization of information for business units. These attributes help the proposed digitization or AI initiatives to realize its intended benefits. Nevertheless, it is common practice for data sourcing and verification to take place only after the business case has been established at the POC stage and stakeholder approval has been obtained to move projects forward.

Figure 2: Desirable data attributes for AI applications

AI Applications Illustrated: NLP-Driven Financial Thresholds and Automated Credit Memo Processing

To help address a potential mismatch between expectations from POCs and the reality of project deliverables, S&P Global Market Intelligence has been investing to providing these data attributes upfront. These desirable data attributes benefit our internal AI-powered applications as well, including S&P Global Market Intelligence’s Natural Language Processing (NLP)-powered Financial Thresholds on RatingsDirect (Figure 3), which is one indicator that may be helpful to assess how far an entity is from a potential credit rating change; and the beta version of our credit research chatbot on RatingsDirect ^® on CIQ Pro.

Figure 3: NLP powered Financial Thresholds by S&P Global Market Intelligence

This data also can be fed into clients’ own internal AI or automation solutions. For instance, credit memos can be automated for large-scale generation, a common use case in commercial lending, capital markets, and wealth management. Below are illustrations of live data accessible through Xpressfeed™, XpressAPI, and XpressCloud delivery platforms.

Figures 4 and 5 illustrate descriptive information and peer comparisons derived by S&P Global Market Intelligence from S&P Global Ratings’ research text, which can be automatically integrated into credit memos and linked to other datasets by leveraging the cross-reference capabilities of Business Entity Cross References (BECRS). In the future, there is potential for credit memos to be populated based on responses to customer-specific chatbot prompts via an API, which are designed and validated by the customer to ensure consistent and reliable responses tailored to each use case, context, user, and product type.

Finally, data coverage is an important consideration in selecting a data solution, as it alleviates the need to manage multiple data pipelines and harmonize data definitions across providers.  For entities that are not rated by credit ratings agencies, the RiskGauge (RG) pre-calculated Credit Scores and Bond Implied Scores (BIS) from Credit Analytics leverages market- and fundamentals-driven analytical models that deliver 880,000+ lower-case credit scores for public and private entities that aims to broadly align with ratings from S&P Global Ratings.  To supplement our extensive coverage, today, we allow clients to load and score the financials of their counterparties which they may acquire directly.  We are looking to automate these processes in the future to seamlessly expand the universe of credit scores available to that client

Figure 4: Visualization of an entity’s descriptive information, financial statements and transactions history from S&P Global’s data delivery solutions 

Figure 5: Visualization of blurbs created from RatingsXpress® Research Sections and Credit analyst adjusted financials for peer comparisons from S&P Global’s data delivery solutions

A data-oriented take on whether Gen-AI lives up to the hype

Ultimately, Generative AI relies heavily on data to drive outputs through learning models. While traditional data processes prioritize the timeliness and geographical coverage of data as key competitive advantages — attributes that are still relevant for AI applications—the same Gen AI model can yield markedly different results based on factors such as machine readability, consistency of presentation, cross-referencing and indexing, as well as the inherent noise within the data etc.  Furthermore, biases present in the data, subsamples of the data, or similar content across multiple sources may cause AI/Gen-AI to develop systematic biases in their outputs at run-time.  Ultimately, models require data that is used for learning and at run-time to be fit-for-purpose – suitable data can help these Generative AI projects live up to expectations, hyped up or otherwise.