Fuzzy logic and Expert systems in the modern AI context

AI conversations often revolve around generative models but not all forms of intelligence need to be creative to be effective. Techniques like fuzzy logic and expert systems excel at capturing nuance, interpreting ambiguity and navigating the grey areas upon which most of the real-world decisions depend on. They provide structure without being rigid and handle complexity in a way that feels closer to human reasoning. By looking beyond the generative wave, we can uncover a broader toolkit for building AI solutions that are dependable, transparent and thoughtfully engineered.

What is fuzzy logic?

Fuzzy logic is a mathematical theory that allows systems to handle real life ambiguity, much like humans do. Unlike the traditional set Theory where statements are either true or false, fuzzy logic works with degrees of truth and its values fall anywhere between 0 and 1. In other words, in classical set theory, an element either belongs to a set or it doesn’t: there’s no middle ground.

For instance, if you want to group all the “tall” people in a set, traditionally you choose a threshold (let’s say 180cm) and all those taller than it are in the set, all the others aren’t. Fuzzy logic instead allows for the classification of a person as “almost” tall, introducing the concept of membership function, which determines to what degrees an element belongs to a fuzzy set. Membership functions are usually chosen by an expert in the domain being modelled with fuzzy logic, but for simplicity we can use a function as below.

As shown, all those shorter than 180cm, still belong to the Tall fuzzy set to a degree that is determined by the ramp on the lefthand side of the diagram.

Although a membership function is a simple concept, it extends the capabilities of logic processing. This allows us to mathematically treat uncertainty and vague definitions. A computer becomes capable of “understanding” what almost tall, or not very tall mean. Of course the example is quite trivial, but we can model membership functions for any set that comes to mind. For instance we can model the risk of a project, the credit score of a borrower, the suspicion level of a transaction in a fraud detection system and so on..

Expert systems

When dealing with complex business decisions, most often information is ambiguous or uncertain and that’s where fuzzy logic comes into play. Unlike rigid yes/no rules, fuzzy logic allows systems to deal with degrees of truth, capturing nuances such as customer being moderately engaged, a risk being somewhat high etc. Expert systems can manipulate information using fuzzy variables, which are variables which accept fuzzy sets as values rather than numbers.

For example an human expert could model the Customer Engagement variable using three fuzzy sets (not active, moderately active and very active) measuring the number of interactions with the customers in a given period of time.

This allows the system to process information like a “Customer is somewhere in between very active and moderately active”. In fact, using the diagram above, a customer who interacts 18 times a year with the company is equally moderately and very active. A customer which interacts 23 times a year with the company is not very different from someone who does it 24 times, and their belonging to the very active set will be close to 1.

When fuzzy logic is combined with an expert system, the result is a fuzzy expert system. Like any other expert system It uses rules provided by human specialists and an inference engine that applies those rules and calculates the appropriate output using operators such as and and or, which are often respectively the Min and Max of the degrees of membership, but there are many more operators one can choose. In fact, there are many known functions called T-norms and S-norms that satisfy certain defined mathematical properties a. These can be selected to define the “and” and “or” operators in a way that meets specific requirements of smoothness and strictness, balances input relevance or favours certain inputs over others.

The rules are usually provided in the form of IF THEN rules, and are something like :

IF the customer engagement is moderate and the purchase history is strong then the priority assigned to the customer is medium.

IF the customer engagement is not active and the purchase history is weak then the priority assigned to the customer is low.

In the following trivial example we use Min as the AND operator as it’s easily visualisable and we provide an input of Customer Engagement as 11 and Purchase history as 10.

Once the input is provided, each member function named in the rule is “activated” and through the logic operators and, or, not an output is inferred. Then all the rules output are combined together and a single output value is calculated through a process named defuzzification. In the example above, the priority is calculated as about 4.7 out of ten, which belongs to a high degree to the set of medium priority.

Here we have used a Mamdani inference system because it’s intuitive and easy to visualise. Exploring other forms of inference, like the Takagi-Sugeno which is a common valid alternative goes beyond the scope of this article.

Naturally, this example is quite trivial, and real-life applications of expert systems can be far more complex. In the medical field, expert systems can assist doctors in diagnosing diseases and suggesting treatments by analyzing patient symptoms and medical history. In engineering and manufacturing, they help detect faults, optimize processes, and maintain quality control. Financial institutions use expert systems for credit approval and risk assessment. Even in transportation, expert systems power driverless trains, monitoring tracks, controlling speed, and responding to obstacles to ensure safe and efficient operations. These examples demonstrate how expert systems can handle complex decision-making tasks that traditionally required human expertise.

Where expert systems lie in the AI landscape

Expert systems shine in environments where decisions must be consistent, explainable and defensible. Unlike machine learning and the more modern generative AI, an expert system doesn’t infer patterns from past data or produce responses based on probabilities: it relies on rules defined by specialists and so every conclusion is traceable to a transparent rationale. This is particularly important in industries governed by strict regulations such as banking, health care, insurance, energy etc. For instance, when a bank algorithmically declines a loan application, an expert system can pinpoint the exact lending criteria that weren’t met; when a hospital triage system assigns a priority to a patient it can show the clinical reasoning that led to that decision. In any case any result is explainable and defensible.

Everyday business reality is rarely black and white. Risk, urgency, customer value, product demand, health indicators range more often in the gray area where traditional boolean decision systems may struggle as everything needs to be forced to a yes or no. By introducing the ability to work with degrees, expert systems remove the need for absolutes and allow machines to work with human-like ways of processing information. They work on a knowledge base that is explicitly encoded by human specialists in the form of rules. They don’t discover new patterns on their own and never deviate from the given logic. In a simple way of looking at it, we could say that expert systems take the opposite approach to data processing than machine learning.

Generative AI, instead, represents a third category, going beyond prediction to create new content based on probabilistic language modelling. Instead of deciding whether a transaction is risky, it can draft a report explaining risk criteria; instead of choosing an underwriting decision, it can propose underwriting rules drawn from historical notes, guidelines, and policy documents. GenAI can summarize huge volumes of unstructured information much faster than humans and often more consistently. However, its creative flexibility comes with unpredictability; it may produce incorrect or incomplete logic unless guided by constraints. This makes it a valuable and powerful tool but not a final decision-maker in regulated environments.

Taken together, the differences create a complementary ecosystem rather than competition. Expert systems deliver airtight reasoning, machine learning captures complex statistical behavior, and generative AI accelerates knowledge discovery.

A good example of synergy between these technologies are the ANFIS: adaptive neuro-fuzzy inference systems. ANFIS is a hybrid intelligent system that combines the pattern-recognition capabilities of neural networks with the reasoning and uncertainty-handling strengths of fuzzy logic. The neural network component enables the system to learn adaptively from input-output data, potentially adjusting the system parameters (such as the membership functions, the weights etc.) while the fuzzy inference component provides the human-like reasoning capability. A key advantage of ANFIS is that its rule set can adapt to changing conditions or data, making it highly effective in dynamic or uncertain environments.

As we look ahead, there is no question that GenAI holds immense promise, and is already starting to deliver on them. Its ability to synthetise information, deal with unstructured data, automate content creation and accelerating decision making is transforming industries at a remarkable speed.

But GenAI it’s not a universal remedy. Certain applications demand predictability, structure and reliability, areas where other solutions like Expert Systems continue to excel. It’s easy to fall into the “when you have a hammer, everything looks like a nail” trap, but let’s not forget modern computing offers us a huge variety of tools each designed to solve different kinds of problems and each delivering their best when used together in the right combination and context: the real opportunities lie in understanding when to use which tool and how to combine them to build solutions that are genuinely aligned with the problem at hand.