Expert System
Also known as: rule-based AI, rule engine, BRMS, knowledge-based system
An expert system is an AI program that encodes the knowledge and decision-making logic of a human expert in a specific domain as a set of 'if-then' rules, and applies those rules to make decisions, diagnoses, or recommendations.
Unlike modern machine learning models that learn patterns from data, expert systems are explicitly programmed with rules written by domain experts. They were the dominant form of AI in commercial and industrial settings throughout the 1980s and 1990s and remain in use today in domains where explicit, auditable rules are required.
Expert System, explained at three levels
At home
An expert system works like a digital flowchart based on expert knowledge. A medical expert system might say: 'IF the patient has a fever AND a rash AND traveled to a tropical country in the last two weeks THEN suggest testing for dengue fever.' A financial expert system might say: 'IF the applicant's credit score is below 600 AND they have had a missed payment in the last 12 months THEN decline the loan.' Each rule was written by a human expert, not learned by the AI from data.
At work
Expert systems are used when decisions need to be explainable, auditable, and consistent. A loan decision made by an expert system can be explained to the customer rule by rule. A medical expert system's recommendation can be traced back to the specific clinical guidelines that triggered it. This contrasts with machine learning models, where the path from input to output is often opaque. Regulated industries often prefer expert systems or hybrid approaches for decisions that require regulatory explanation.
At home
An expert system has three main components: a knowledge base (the rules, encoded as IF-THEN statements), a working memory (the current facts about the situation being assessed), and an inference engine (the reasoning mechanism that applies rules to facts to reach conclusions). The inference engine fires rules forward (working from facts to conclusions) or backward (starting from a possible conclusion and checking whether the facts support it). Modern Business Rule Management Systems (BRMS) like Drools, IBM ODM, or FICO Blaze Advisor implement this architecture in commercial software.
At work
Expert systems suffer from brittleness when they encounter situations outside their rule set. A medical expert system trained on adult symptom patterns may produce incorrect recommendations for paediatric patients if no rules for age-specific differences were coded. Maintaining expert systems requires ongoing rule updates as domain knowledge evolves -- changes to clinical guidelines, tax regulations, or technical standards must be manually translated into rule updates by knowledge engineers. This maintenance burden was a key factor in the decline of large-scale expert systems in the 1990s.
At home
Hybrid AI architectures combine expert systems (rules) with machine learning (statistical pattern recognition) to get the benefits of both: explainability and consistency from rules, adaptability and pattern-finding from ML. In credit scoring, a rules engine might enforce hard regulatory limits (no lending above certain debt-to-income ratios, mandated by law) while an ML model predicts default probability within the rules envelope. In insurance underwriting, rules enforce regulatory constraints while ML models price the policy.
At work
From a practitioner standpoint, the key design question when choosing between an expert system and a machine learning model is whether the decision logic can be fully articulated as rules by domain experts. If domain experts can describe every relevant scenario and the appropriate response, a rules-based approach may be more appropriate -- it is auditable, explainable, and does not require training data. If the patterns in data are too complex or numerous to enumerate as rules, an ML model is more appropriate. Many production systems use both: ML for prediction, rules for constraint enforcement.
Real-world examples
At home
Thomas used the UK government's online benefits calculator, which is built on an expert system: the calculator applies HMRC rules (IF income is above X AND household size is Y AND rent is Z THEN Universal Credit entitlement is W) to his answers to produce an entitlement estimate. Every output can be traced back to the specific rules that generated it.
At work
The underwriting team at Meridian Insurance maintained a BRMS (Business Rule Management System) for home insurance pricing. Business analysts updated pricing rules using a rule editor (no code changes required), allowing the business to respond to actuarial model updates without waiting for a software release cycle. The rules engine processed 2,000 quotes per minute with full audit trails for every pricing decision.
How it compares
| Feature | Expert System | Machine Learning Model | Hybrid AI |
|---|---|---|---|
| Knowledge source | Explicitly coded rules from human experts | Patterns learned statistically from data | Rules for constraints + ML for predictions |
| Explainability | Fully auditable -- every decision traceable to rules | Often opaque (black box), especially deep learning | Rules layer is auditable; ML layer may require explainability tools |
| Adaptability | Requires manual rule updates as domain evolves | Can be retrained on new data | Rules updated manually; ML layer retrained on data |
| Best for | Regulated decisions, compliance, consistent enforcement | Pattern recognition in complex, high-volume data | Combining regulatory constraints with statistical prediction |
Expert systems are best when decisions must be fully explainable and rules can be enumerated. ML models handle complex pattern recognition but lack native explainability. Hybrid architectures use both to balance regulatory compliance with adaptability.
For analysts
Expert systems are directly relevant to BAs who work on business rules and decision automation. When documenting requirements for a decision-automation system, distinguish between rule-based requirements (fixed business logic that must always apply, regardless of data patterns) and prediction requirements (outputs that require learning from data). For rule-based requirements, BAs should document the decision table or decision tree structure, the data inputs required, the output classifications, and the exception handling process. Useful format: decision tables mapping input conditions to output actions, which can be directly implemented in a BRMS.
History & origin
Expert systems emerged from the AI research labs of the 1960s and 1970s. MYCIN (Stanford, 1972) was an early expert system for diagnosing bacterial infections and recommending antibiotics; DENDRAL (Stanford, 1965) analysed chemical compounds. Commercial expert systems proliferated in the 1980s, driven by optimism about AI -- companies including DEC, Ford, and the major banks deployed them for configuration, diagnosis, and lending decisions. The 'AI winter' of the late 1980s and 1990s followed when the limitations of rule-based systems (brittleness, high maintenance cost, inability to handle uncertainty well) became apparent. Modern rule engines and business rule management systems (BRMS) are descendants of expert systems.