---
title: "Expert System"
url: "https://mahouse.world/glossary/expert-system"
type: "glossary term"
topic: "AI concepts"
author: "Ihor Vlasov"
reviewed: "2026-09-20"
published: "2026-08-10"
---

# Expert System

An expert system is a program that encodes a human expert's knowledge as explicit rules — thousands of if-then statements over a knowledge base, run by an inference engine — to reach conclusions in a narrow domain. The dominant commercial AI of the 1980s, it fell from favour when the rules proved brittle and costly to maintain, and it lives on as the rules engines that run transaction monitoring, underwriting and compliance at every fintech.

**Also known as:** Rule-based system, Knowledge-based system, Expert systems (1980s AI)

## Key facts

- **Used at:** historically medicine (MYCIN), chemistry, configuration (DEC's XCON), finance; today, under the names rules engine, decision engine, policy engine — AML monitoring rules, credit policy rules, fraud rules, KYC decisioning, and the guardrails around machine-learning models
- **Binding:** no standard; but regulators like rules, because a rule is explainable and auditable by construction, and a lender's credit policy is usually rules the model must respect
- **Typical structure:** a knowledge base of rules ("if utilisation above 80% and two missed payments then decline"), an inference engine that applies them (forward or backward chaining), an explanation facility (the rules that fired), and an editor for the experts who maintain them
- **Typical failure:** the knowledge-acquisition bottleneck — extracting rules from experts is slow and the rules interact in ways nobody foresaw; brittleness — a case the rules did not anticipate produces nonsense; maintenance — thousands of rules nobody dares change; these ended the first AI boom in the late 1980s ("AI winter")
- **Parties:** the domain experts whose knowledge is encoded, the engineers who maintain the rules, the users of the decisions, and the regulators who can read them

## How an expert system works in practice

A fintech's AML transaction-monitoring system is an expert system by another name: three hundred rules written by compliance experts — thresholds by customer segment, velocity patterns, jurisdiction lists, structuring typologies — applied to every transaction by an engine that raises an alert when a rule fires and records which one. An analyst reads the alert, sees the rule, and investigates; a regulator reads the rule set and asks why each threshold is where it is. The system is explainable by design, auditable line by line, and brittle in exactly the way the 1980s systems were: a laundering pattern the rules do not describe passes; a rule tuned for one segment fires on another; and the rule set grows as every incident adds one.

The modern architecture pairs the rules with a model. The rules encode policy — what the business will and will not do, what the regulator requires, what must be explainable — and the model finds the patterns the rules cannot: an anomaly score that ranks alerts, a fraud model that catches what no rule describes. The division is deliberate: rules where the decision must be explained and controlled, models where the pattern is too complex to write down, and rules again as guardrails around the model's outputs. A credit decision at most lenders is a policy rule set (the hard eligibility criteria) wrapped around a scoring model (the risk estimate), which is the expert system and its successor sharing the job.

For a buyer of a fintech, the rules engine is where the company's judgement is written down — and the place to look for rules nobody has reviewed since they were added after an incident in 2019.

## Example

A payments company's fraud system began as 40 rules written by its first risk analyst. Six years later it has 1,100, added one incident at a time; nobody can say which are still needed; the false-positive rate is 94%; and a new fraud pattern took three weeks to write a rule for while it ran. The rebuild: a machine-learning model scores every transaction; 120 rules remain, expressing hard policy (blocked countries, velocity limits, sanctioned merchants, regulatory requirements) and explainability requirements; the rest are retired after a back-test shows the model catches what they caught. False positives fall to 70%, new patterns are caught by the model before anyone writes a rule, and the 120 rules are reviewed quarterly with an owner each. The regulator's inspector, reading the new rule set, asks fewer questions than about the old one.

## Common mistakes

- Rules added after every incident and never retired.
- A model deployed with no policy rules around it, so it decides what the business never authorised.
- Rules that encode one expert's 2019 opinion as permanent policy.
- Thresholds set by round numbers and defended to a regulator as expertise.
- Buyers reading a fintech's "AI-driven" risk system and finding 1,100 if-statements.

## Related terms

- [Symbolic AI](https://mahouse.world/glossary/symbolic-ai.md)
- [Machine Learning](https://mahouse.world/glossary/machine-learning.md)
- [Transaction Monitoring](https://mahouse.world/glossary/transaction-monitoring.md)
- [Explainable AI](https://mahouse.world/glossary/explainable-ai.md)
- [Risk-Based Monitoring](https://mahouse.world/glossary/risk-based-monitoring.md)
- [Artificial Intelligence](https://mahouse.world/glossary/artificial-intelligence.md)

— Ihor Vlasov, M&A FinTech Expert. Last reviewed 2026-09-20. https://mahouse.world/authors/ihor-vlasov