Analytics Maturity Levels and “Enterprise Intelligence”

This blog revisits the familiar “analytics maturity model” from the perspective of my book, Enterprise Intelligence. This post was originally intended to be “Appendix B” of the book. But the book was already long and this appendix was too late to the party, so it was dropped to the cutting room floor.

Nonetheless, it’s a good way to frame the traditionally recited progression of analytics maturity levels in contrast to this AI-driven time.

To start, let’s review the traditional analytics maturity levels, along with an example:

1. Descriptive Analytics: Focuses on summarizing historical data.
   • What insights are helpful in addressing customer attrition? Bad customer support incidents, better deal elsewhere, customer lost interest?
2. Diagnostic Analytics: Analyzes data to find reasons behind occurrences.
   • How does what happened lead to customer attrition?
3. Predictive Analytics: Develop machine learning (ML) models used to predict (statistically) what will happen.
   • Knowing what we know about customer attrition, which customers are at risk for leaving?
4. Prescriptive Analytics: Recommends actions towards achieving desired outcomes.
   • What can we do to prevent an at-risk customer from leaving?

Assuming, you haven’t read my book, for the best results from this blog, please at least read the TL;DR of my book which will introduce the acronyms:

• DC – Data Catalog
• KG – Knowledge graph – ontologies about how things relate.
• ISG – Insight Space Graph
• TCW – Tuple Correlation Web
• EKG – Enterprise Knowledge Graph – KG + DC + ISG + TCW

Further, my architectural description provides a more technical take of the book than the “book cover description”.

Note: Enterprise Intelligence is available at Technics Publications and Amazon. If you purchase the book from the Technics Publications site, use the coupon code TP25 for a 25% discount off most items (as of July 10, 2024).

The concept of the analytics maturity levels is used to describe the stage of analytical capabilities within an organization. Each stage represents a more advanced level of data analysis, building upon the previous one to provide deeper insights and more actionable outcomes. Assessing the level of analytics maturity shows us where we add further analytics capabilities.

An understanding of these traditionally recognized analytics maturity levels can help frame how the primary structure of my book, the Enterprise Knowledge Graph (EKG), function and contribute to enterprise intelligence. A crucial point is that the levels are not superseded by higher levels. Each level has its place in the analytics processes. This is analogous to how algebra remains equally relevant even at geometry and calculus.

I do think the term, “maturity”, can be a little misleading. For example, if our enterprise’s analytics maturity level is found to be at descriptive, that doesn’t mean diagnoses, predictions, and prescriptions don’t happen there. The analytics process of even a mango cart proprietor in reality covers all four of those levels. The question is, at what point do analytics databases, software, and automation come into play? Each level is composed of human and machine intelligence. Hopefully, the bigger the role of machine intelligence at various levels, the “smarter” the analysis and decisions will be.

Descriptive Analytics is roughly what we have traditionally thought of as Business Intelligence (BI). That is, we’ve successfully developed some set of BI data sources such as a data warehouse, data marts, and OLAP cubes. A human analyst opens up Tableau/PowerBI/Excel, connects to a BI data source, and slices and dices away, viewing data in a variety of visualizations. The visualizations reveal insights such as trends and inflection points (line graphs), and distributions (bar and pie charts, scatter plots).

The higher levels generally involve a growing proportion of machine learning (ML) and the accompanying data science processes and data scientists. But the abrupt and profound emergence of readily usable and highly-performant (relative to the smartest of human experts across all fields) large language models (LLM) into the mainstream require revisiting the analytics maturity levels.

Diagnostic (2) and predictive analytics (3) is dominated by the ML we’ve grown to love over the past decade or so. For the diagnostic level, ML is often applied as associations, regressions, and decision trees which reveal nuggets of information hidden in the data. For predictive analytics (which is the application of what we’ve learned in diagnostic), there are recommendation engines, fraud detection, even medical image analysis. ML also “wades” in the shallow-end of the Prescriptive Analytics (4) pool.

With the impact of LLMs over the past year and half since ChatGPT rudely burst into the daily lives of the general public, there is greater feasibility for reaching the deeper end of the Prescriptive Analytics pool. Here is a short outline of how LLMs affect the Analytics Maturity Levels:

1. Descriptive Analytics: This is what we traditionally think of as BI, enhanced by LLMs for automated insights.
2. Diagnostic Analytics: Uses ML for pattern recognition, enhanced by LLMs for unstructured data analysis.
3. Predictive Analytics: Relies heavily on advanced ML, with LLMs providing nuanced predictions.
4. Prescriptive Analytics: Uses optimization and advanced ML, with LLMs aiding in the crafting of robust recommendations.

Following are brief discussions of each level. My descriptions probably will disagree with what others have grown to know or how others see LLMs altering these maturity levels. But that’s to be expected during the early phases of paradigm shifts. My opinion is that:

• LLMs narrow the gap between machine and human communication, therefore, the maturity levels aren’t as tech-centric.
• LLMs drastically expand intellectual capacity by taking on more mundane, lower-skilled intellectual tasks, which is the bulk of analysis.
• The intellectual expansion enables us to follow models of problem solving exemplified by Sherlock Holmes, the TV character, House, and that resourceful guy, MacGyver.

However, before diving into the first level, descriptive analytics, I need to prepend two items at the top: The problem we’re trying to solve and the empirical data. Starting our story at descriptive analytics might be like watching Rocky II without having seen Rocky I (or at least the highlights of Rocky I).

Problem We’re Solving

Analytics doesn’t make sense without a problem to solve. A problem can be clear (specific), such as investigating how someone died, or it can be open (vague), such as how to improve profits. Of course, that’s obvious, but a discussion of levels of analytics requires us to get into the spirit being presented with problems and relieving ourselves of the problems.

Empirical Data

To analyze, of course, we begin with data. As “duh” as that is, I wish to provide the reminder to build context around the levels.

For a detective, these are clues left at the crime scene. For a doctor, these are symptoms presented verbally from the patient and the doctor’s observation of the patient.

But there are three levels of data for analysis, which is very well stratified using Databricks’ “Medallion Architecture”:

1. Bronze Layer: Raw Data – raw, unprocessed data ingested from various sources.
2. Silver Layer: Cleansed and Enriched Data. Silver generally equates to a Data Warehouse, but the DW could also be considered gold, especially if implemented on a powerful platform such as Snowflake.
3. Gold Layer: Integrated (across many domains) and aggregated (optimized for performance) business-level data. OLAP cubes of the SQL Server Analysis Services (SSAS) category are gold. Other examples might be a data mart derived from a DW and the “business vault” of a data vault.

“ETL/ELT” is the process of upgrading data from bronze to gold. The silver and gold data are generally the input into the Descriptive Analytics level. My feeling is that pre-LLM, this empirical data would have been lumped into descriptive analytics – ETL and the BI analysis using visualization tools.

Descriptive Analytics (Higher-Order Data)

Descriptive analytics is usually said to be about “what happened”. But I like to look at descriptive analytics as a continuation of data gathering that began in empirical data. That is, create new data (higher order data) from the empirical data. This high order data provides more context and insights.

Traditional BI

• Purpose: Repackage historic data into many different angles for viewing.
• Tools and Techniques: Dashboards, reports, data visualization, and basic statistical analysis.
• Data Sources: Data warehouses, data lakes, and structured databases.
• Example: Monthly sales reports, customer segmentation based on past behavior.

Maturity Level

• Focus: Generate higher order data from existing data.
• Machine Learning Integration: Minimal ML; mostly relies on associative reasoning and manual data exploration.
• LLMs Impact: Enhanced capabilities in natural language processing (NLP) prioritize exploration based on the context of the problem and available data.

In the BI context, data (usually gold, but sometimes silver or even bronze) is accessed through the likes of Tableau and PowerBI. The slice and dice activity we do in Tableau and PowerBI is the process of descriptive analytics. We view slices and dices of data through visualizations that reveal insights to us. However, all three medallion tiers of data can serve as input into the description analytics level.

As mentioned, the descriptive level is usually where BI is thought to end. The silver and gold tier data are the input, and the output are insights gleaned through those visualizations. Normally, those insights exist only in the heads of analysts viewing the visualizations, slicing and dicing away in Tableau/PowerBI. The analyst notices the insights and might communicate it to other people through email, conversations, PowerPoint slides, or public dashboards. A very small minority of what are considered important insights might be transcribed to some sort of report.

The EKG automatically captures those insights into two structures:

• ISG: The ISG is descriptive as it represents a memory of what BI data would manifest in visualizations. It captures salient points from BI analysts’ visualizations, reflecting historical data and insights.

• TCW: The TCW is diagnostic as it represents a network of correlations and relationships between event tuples within the enterprise data. It captures the intricate web of how different data points interact and influence each other, enabling a deeper understanding of the underlying patterns and causes behind observed outcomes. The TCW leverages both historical data and advanced analytics to map out these connections, providing insights into why certain events occur and how different factors are interrelated.

Descriptive analytics are, for the most part, just more data points. For example, the history of PSA levels is empirical data, but noticing a continual trend upwards is a data point derived from the time series of empirical PSA readings.

Higher-order data is more valuable because it is structured, aggregated (compacted into a succinct form), and enriched, making it easier to analyze and understand. For instance, a raw dataset of sales transactions can be transformed into a sales report that highlights trends, patterns, and key performance indicators. This higher-order data is crucial for making informed decisions and setting the stage for more advanced levels of analytics.

By starting with a clear problem and empirical data, and then applying descriptive analytics to create higher-order data, organizations can build a solid foundation for deeper analysis and insights.

Diagnostic Analytics (The Model of Why it Happened)

Diagnostic analytics dives into the data to discover the reasoning and mechanics of the problem. It’s this point where a detective develops and tests her hypothesis after gathering clues at the crime scene and interviewing persons of interest. Or where the doctor has listened to the patient (hint, hint), examined the patient, received all test results, run it through her ten years of medical training, and is prepared to offer a diagnosis.

The process of the diagnostic level is to compose hypotheses build on the input of the descriptive data until it becomes a theory (an artifact of deductive reasoning), which is the output of the diagnostic level.

Purpose

• Purpose: Analyzes data to understand why certain events occurred.
• Tools and Techniques: Root cause analysis, drill-down, data mining, and anomaly detection.
• Data Sources: Enriched and cleansed data from data warehouses and data lakes.
• Example: Analyzing why sales dropped in a specific region by examining related factors.

Maturity Level

• Focus: Generating hypotheses from data (abductive reasoning) – abductive diagnostics.
• Machine Learning Integration: Moderate ML; includes clustering, classification, and regression techniques to identify patterns and correlations.
• LLMs Impact: Ability to analyze unstructured data (e.g., customer reviews, social media posts) to provide deeper insights into reasons behind trends and anomalies.

Twenty years ago, this was almost the exclusive realm of human intelligence. It was and is still mostly that quanty person with not much else but the keyboard, monitor, and MBA. Sherlock Holmes and House (the TV show) are the memes of the epitome of detective work and medicine, respectively.

Over the past two decades, ML plays an increasing role at this level of analytics. ML models automatically sort out hundreds of features of millions of “cases” (ex. visits to the doctor, crime instances) automatically discovering correlations and other kinds of relationships between the features. From these pieces, a chain or web of cause and effect is composed as a hypothesis.

It’s important to note that ML models don’t just automate analysis (relieving humans of drudgery), but it expands the analytical capability pie. ML models can do the majority of tasks much faster and deeper, leaving more time for humans to attend to the tasks requiring the advanced intellectual and creative capability that’s still mostly in our realm.

LLMs such as ChatGPT having recently gone viral into the mainstream greatly expands the scope of our diagnostic capabilities. LLMs are very useful in taking in a collection of relationships and applying its “knowledge” towards coalescing hypotheses.

As for the EKG, TCW is diagnostic as it represents event tuples and their correlations and probabilities, enabling analysis of the relationships and patterns that provide pointers as to why certain outcomes happened. By capturing these correlations, the TCW aids in understanding the causes behind observed data trends.

As a business example, we might discover there is a big base of 49ers fans in Boise due to the massive influx of people from the San Francisco Bay Area. Most still love the Bay Area (thus still love the 49ers) but have been pushed out by high prices. Further, they still might even work in the Bay Area thanks to remote work technology and it’s less than a two-hour flight away.

Although the diagnosis level is early in the levels as traditionally laid out (2nd of 4) and it’s sometimes not even included, this is the meatiest level as I present it in this AI world. The output of the diagnostic level isn’t just the diagnosis. It’s the web of logical rules and/or step by step chain of facts that was learned to get to the diagnosis.

What we learn about the “why” in this diagnosis level should be the basis for the following levels (Predictive, Prescriptive).

Predictive Analytics (Apply the Knowledge)

Predictive analytics composes statistical models from historic data to predict future outcomes. Data is fed through ML algorithms which outputs prediction models-chains or webs of statistically correlated/measured relationships. In the analytics world I grew up in, I’d say this has been the primary target of ML and data science.

Purpose

• Purpose: Uses historical data and statistical models to forecast future outcomes.
• Tools and Techniques: Time series analysis, forecasting models, and advanced ML algorithms (e.g., neural networks, decision trees).
• Data Sources: Historical and current data from various sources, often in near real-time.
• Example: Predicting next quarter’s sales based on historical data and market trends.

Maturity Level

• Focus: Applying diagnostic insights to forecast future events (inductive and deductive reasoning).
• Machine Learning Integration: Advanced ML; includes predictive modeling, supervised learning, and ensemble methods to improve accuracy.
• LLMs Impact: LLMs enhance predictive capabilities by interpreting complex datasets and generating nuanced predictions based on diverse data sources.

These predictive models could be thought of as the application of what we’ve learned at the diagnostic level. Once we’ve found a chain/web of relationships that correctly results in some value, we can apply that to many other cases, which an expectation of being correct some percentage of the time.

Continuing with the diagnosis of the growing base of 49ers fans in Boise, we might predict rising sales in Boise for NFL teams at cities sharing salient characteristics such as Los Angeles, Seattle, Denver, and Phoenix (all very crowded and within a short flight from Boise). Perhaps the same applies for other sports teams of other leagues (MLB, NBA, NHL) in those cities.

Prescriptive Analytics (Wisdom)

Prescriptions are a set of steps to address a specific problem. These steps are derived from experience or data and tested/proven for “real-world” accuracy. Prescriptive analytics goes a step further from predicting a condition to suggesting actions that can influence desired outcomes. It involves optimizing decision-making processes and recommending specific courses of action.

Purpose

• Purpose: Recommends actions to achieve desired outcomes.
• Tools and Techniques: Optimization algorithms, simulation models, decision analysis, and advanced ML.
• Data Sources: Outputs from predictive models and enriched datasets.
• Example: Suggesting optimal inventory levels to minimize costs and meet demand forecasts.

Maturity Level

• Focus: Providing actionable recommendations based on predictive insights (inductive reasoning) towards a defined problem.
• Machine Learning Integration: Sophisticated ML; uses reinforcement learning, optimization techniques, and scenario analysis.
• LLMs Impact: LLMs assist in generating prescriptive insights by integrating vast amounts of data and providing context-aware recommendations, facilitating decision-making processes.

When AI (LLM, multi-modal recognition, RAG) is integrated with the EKG, prescriptive analytics can be achieved. By leveraging insights from the KG, ISG, and TCW, AI can recommend actions to optimize business processes and achieve strategic objectives.

Proscriptive Analytics (“Life Advice”)

Wait … what?! Is “Proscriptive Analytics” a thing? I thought it was. I recall someone introducing me to the concept about a decade ago. My impression is that proscriptive is like the Judaism’s “Ten Commandments” (“thou shalt not”) and the things “my mother always said”. In other words, don’t do these things and you’ll be good and golden.

But the term rarely comes up in the course of my BI practice since. When it does come up, it’s usually more advanced that prescriptive analytics.

Researching (Googling and reading a few articles) the term today, there really isn’t much more than what I learned all those years ago about “proscriptive analytics”. From the few sources I’ve read, it seems like people usually refer to predictive analytics, but occasionally someone chooses to use the word “proscriptive” in an almost interchangeable way.

Purpose

• Purpose: Advises on actions to avoid or prevent undesirable outcomes.
• Tools and Techniques: Risk management frameworks, rules-based systems, and advanced ML for pattern recognition.
• Data Sources: Insights from diagnostic and predictive analytics, enriched with real-time data feeds.
• Example: Advising against specific marketing strategies that have historically led to negative ROI.

Maturity Level in this Time of AI

• Focus: Creating systems of advice based on inductive and deductive reasoning to avoid risks and negative outcomes.
• Machine Learning Integration: Integrates the outputs of ML models to define constraints and avoidance strategies.
• LLMs Impact: LLMs enhance the ability to generate comprehensive advice and identify potential pitfalls by analyzing patterns across extensive datasets.

Let’s start with definitions. From dictionary.com:

• Prescriptive – Saying exactly what must happen, especially by giving an instruction.
• Proscriptive – involving, imposing, or defining limits and prohibitions.

We won’t get any help thinking about the prefixes, pre and pro. Both pretty much suggest something before. It might help to think of the pro in proscriptive as “prohibit” (don’t do that).

We could use as examples:

• Prescriptive (do this): Lose weight to control of even possibly reverse a diabetic condition.
• Proscriptive (don’t do this): Don’t become obese in order to avoid a diabetic condition.

Think about preventive care. The idea is to take steps to avoid bad conditions before they happen, thus requiring very expensive treatments. For example, don’t smoke to avoid lung cancer – that’s a “don’t”. But we could also say keep slim to avoid diabetes – that’s a “do”.

The difference between prescriptive and proscriptive is confusing if we think about prescriptive as “do this” and proscriptive as “don’t do this”. We can always rephrase a “do” to a “don’t” and vice-versa. So is there a need to differentiate prescriptive and proscriptive?

Is “Eat more fruits and vegetables for better health” prescriptive or proscriptive? It’s telling us to do something. We could rephrase to “Eat more fruits and vegetables to avoid malnutrition”. Or better yet, we can make it proscriptive like this: “Don’t not eat more fruits and vegetables for better health”.

Honestly, at the time I write this, the notion of prescriptive versus proscriptive as “to recommend” versus “to forbid”, respectively, doesn’t seem helpful in a practical sense.

However, I do think it’s helpful to think of the difference as how to fix a clear and present problem (prescriptive) versus how to avoid a problem (proscriptive) before it happens. For example, when facing a diagnosis of lung cancer, the prescription involves a series of procedures and lifestyle changes. To avoid lung cancer, the proscriptive recommendation is to stop smoking (which goes for those diagnosed with lung cancer as well).

Another example is calling an ambulance when I’m hit by a bus (prescriptive) versus looking both ways before crossing the street (proscriptive). Or the latter rephrased in a more properly proscriptive way: Don’t cross the street before you’re sure it’s clear.

In the context of AI/BI systems, a more useful way might to be think of prescriptive versus proscriptive in the performance management terms, tactical versus strategic. For example, if we have slumping sales at a store probably due to a bad location, we can deploy a tactical marketing campaign. We can deploy a strategy of fantastic customer service to ensure a high level of repeat customers.

A big problem with such blanket proscriptive advice is that there are countless events we might want to avoid. Eventually we collect so many of them that they start to compete or contradict each other. Somehow we must engineer a system that balances all the competing goals-and engineering is tricky.

Systems Thinking

The notion of collections of proscriptive advice and avoiding confusion reminds me of Scott Adams’ (among others) concept of building systems instead of focusing on goals. If we build the right system, we inherently avoid many pitfalls and negative outcomes along the way.

Scott Adams emphasizes that systems are superior to goals because:

1. Sustainability: Systems provide a structured approach to consistently achieve small, incremental improvements, which are more sustainable over time.
2. Flexibility: Systems can adapt to changing circumstances, whereas goals are fixed and can become irrelevant as conditions change.
3. Ongoing Improvement: Systems focus on continuous improvement and learning, rather than achieving a static goal and then stopping.
4. Resilience: Systems build resilience by preparing for and mitigating risks, just as proscriptive analytics aims to avoid negative outcomes by advising against certain actions.

By incorporating “proscriptive” (the way I’m defining it here) analytics into our decision-making framework, we are essentially building a robust system that not only guides us towards positive outcomes but also helps us avoid multiple potential pitfalls. Just as Scott Adams suggests focusing on systems to ensure long-term success, proscriptive analytics helps create a structured approach to consistently avoid undesirable results, leading to more resilient and effective strategies.

In other words, a system of “proscriptive” advice.

Proscriptive analytics aligns with the very wise philosophy of building systems over setting rigid goals. It provides a proactive approach to data-driven decision-making, ensuring that we not only aim for positive outcomes but also systematically avoid negative ones. This dual focus can lead to a more balanced and resilient strategy in both personal and business contexts.

Lastly, I discuss Strategy Maps, in the book and in my blog, Revisiting Strategy Maps. A strategy map is a kind of knowledge graph encoding theoretical cause and effect. It’s a model of how events in the enterprise affect each other.

Mapping Analytics to Reasoning Types

Another recent blog related to Enterprise Intelligence, Exploring the Higher Levels of Bloom’s Taxonomy, is a 1st cousin to this blog. Within the context of the application of AI as mainstream, it dives deep into the meaning of inductive, deductive, and abductive reasoning.

Spoiler alert: Abductive reasoning is the least known term of the three and is the sort of reasoning that has been very difficult for non-AI analytical systems. It’s primarily the realm of human reasoning. It’s what engineers, scientists, detectives, and doctors do when faced with novel problems. As “abductive” is an obscure term, so is “proscriptive”.

Following are summaries of how the analytics maturity levels map to the types of reasoning.

Descriptive Analytics

• Purpose: Mostly data gathering and data preparation.
• Reasoning Type: Inductive.
• Example: Identifying blood spatter as a potential clue for a crime scene based on prior knowledge.

Diagnostic Analytics

• Purpose: Generating hypotheses out of data to determine why something happened.
• Reasoning Type: Abductive.
• Example: Hypothesizing that a drop in sales is due to a failed marketing campaign based on observed data.

We mostly think of doctors as “diagnosing”. But most such diagnosticians are highly trained pattern recognizers. In other words, their intense medical training enables them to incredible combinations of symptoms to a medical condition. However, some doctors, such as TV’s House, go beyond this sort of trained pattern recognition to developing a hypothesis for a novel set of symptoms.

Predictive Analytics

• Purpose: Applying diagnostic insights to forecast future outcomes. Using historical data, identified patterns, and rules to predict something.
• Reasoning Types: Inductive and Deductive.
• Example: Using historical sales data and identified patterns (ML models) to predict the effect of an ad campaign for a particular cohort of customers.

Prescriptive Analytics

• Purpose: Recommending actions to achieve desired outcomes. When we have a diagnosis, we need to address mitigation or elimination of the symptoms/problems.
• Reasoning Type: Inductive and deductive reasoning delivered from “conventional” ML models.
• Example: Suggesting a marketing strategy to increase future sales based on predictive insights.

Proscriptive Analytics

• Purpose: Providing systems of advice, often about what not to do, to avoid undesirable outcomes.
• Reasoning Types: Inductive and Deductive reasoning hashed out by LLM prompt-engineered with a strategy map of relationships and goals.
• Example: Advising against launching a similar marketing campaign that previously failed

Proscriptive analytics is a system of inductive and deductive reasoning that leads to comprehensive advice on avoiding undesirable outcomes. For example, how can I have my cake and eat it too, while keeping my BMI and glucose levels in check, and benefiting from a Keto diet?

Healthcare’s Triple Aim (A Proscriptive System)

Proscriptive analytics involves creating a robust system of advice and constraints to prevent negative results, simultaneously addressing multiple dimensions. This approach emphasizes a proactive, system-based strategy over focusing on individual goals. The Triple Aim framework in healthcare exemplifies this methodology.

Triple Aim Framework

The purpose of the Triple Aim framework, developed by the Institute for Healthcare Improvement (IHI), is to optimize health system performance by focusing on three dimensions:

1. Improving Individual Health Outcomes: Enhancing the patient experience of care, including quality and satisfaction.
2. Improving Population Health: Addressing the broader determinants of health to improve the health of entire populations.
3. Reducing Per Capita Costs: Lowering the cost of healthcare without compromising quality.

Proscriptive Analytics in Triple Aim

The Triple Aim framework leverages proscriptive analytics to provide structured and sustainable solutions by creating a system of validated guidance and rules. Here’s how it works across the three dimensions:

1. Improving Individual Health Outcomes:
   • Rule: Avoid unnecessary medical procedures to prevent patient harm and reduce costs.
   • Proscriptive Advice: Implement evidence-based clinical guidelines to ensure patients receive appropriate care. For instance, using decision support systems to recommend against prescribing antibiotics for viral infections.
   • Outcome: Better patient outcomes through the avoidance of unnecessary interventions, improving overall care quality.
2. Improving Population Health:
   • Rule: Prevent the spread of infectious diseases to improve public health.
   • Proscriptive Advice: Initiate public health campaigns promoting hygiene practices and healthy lifestyles. Utilize predictive modeling to identify high-risk populations and intervene early.
   • Outcome: Healthier populations with reduced incidence of preventable diseases, leading to better overall community health.
3. Reducing Per Capita Costs:
   • Rule: Avoid overuse of expensive diagnostic tests to control healthcare costs.
   • Proscriptive Advice: Establish protocols for cost-effective diagnostic testing and encourage the use of primary care services for initial assessments. Implement value-based care models that reward healthcare providers for cost-efficient practices.
   • Outcome: Lower healthcare costs per capita by minimizing unnecessary spending while maintaining high-quality care.

System-Based Strategy

Proscriptive analytics within the Triple Aim framework illustrates how creating a system of advice and constraints can proactively address multiple objectives simultaneously. This system-based strategy is inherently proactive, focusing on preventing negative outcomes by leveraging historical data, validated models, and best practices. By doing so, it ensures that healthcare providers not only strive to achieve specific goals but also maintain a sustainable approach that balances quality, cost, and population health.

For example, consider a hospital implementing the Triple Aim framework with proscriptive analytics:

• Clinical Decision Support Systems (CDSS): These systems use historical patient data and clinical guidelines to advise physicians on the most appropriate treatments, avoiding unnecessary procedures and improving patient outcomes.
• Population Health Management: Using data analytics to identify at-risk populations and proactively engage them with preventive care programs, such as chronic disease management and vaccination drives, to improve overall health.
• Cost Management Protocols: Implementing protocols to reduce the overuse of expensive diagnostic tests and treatments by promoting evidence-based practices, thereby reducing per capita healthcare costs. While maintaining the quality of individual outcomes.

The Triple Aim framework serves as an example of proscriptive analytics in action, demonstrating how a well-structured system of advice and constraints can prevent negative outcomes across various dimensions. By focusing on a system-based strategy, proscriptive analytics ensures sustainable, high-quality, and cost-effective healthcare delivery.

Conclusion

The proliferation of AI into the mainstream of enterprise analytics facilities the mainstreaming of proscriptive analytics and abductive reasoning. These two terms have been almost exclusively the domain of humans. However, AI is beginning to forge its way into the domain of human intelligence. AI is far from a Nicola Tesla, Holmes, or “House”, but it is on a path in that direction. When (or if) it makes it all the way down the path is debatable.

Regardless, the LLMs of today are a completely different kind of intelligence from those of even the smartest of us. Unlike Tesla and Holmes, it has a much wider range of knowledge locked and loaded.

The EKG components described in Enterprise Intelligence aligns with various types of analytics:

• Descriptive: ISG, KG – Capture what analysts would look for in a visualization.
• Diagnostic: TCW, relationships between things.
• Predictive: Machine Learning and Data Science – Find patterns in data for the purpose of taking really data-driven, educated guesses.
• Prescriptive and Proscriptive: AI recommendations accentuated with widely-scoped BI insights and correlations stored in the EKG.

What I’ve written here is just my attempt to map what’s new in AI to what we currently understand. Attempting such mapping can hold back progress. But as long as it’s not pounding a square peg into a round hole, it’s a good intellectual exercise during the early stages of a paradigm shift.