Case-Based Reasoning Systems: How They Work in Practice

Case-based reasoning (CBR) is a problem-solving paradigm in which an automated system retrieves historical cases, adapts their solutions to fit a new problem, and stores the outcome to build a growing knowledge base. This page describes how CBR systems are structured, how they operate across a retrieval-adaptation-revision cycle, where they are deployed in enterprise and public-sector technology contexts, and where their reliability boundaries constrain appropriate use. The subject is relevant to AI architects, procurement officers, compliance analysts, and researchers mapping the types of reasoning systems used in production environments.

Definition and scope

Case-based reasoning is a form of analogical reasoning in which a system solves new problems by drawing on a library of previously solved problems rather than by applying explicit programmed rules. The paradigm was formally characterized by Roger Schank and his colleagues at Yale University in the 1980s, and the canonical four-phase model — Retrieve, Reuse, Revise, Retain — was articulated by Aamodt and Plaza in their 1994 paper "Case-Based Reasoning: Foundational Issues, Methodological Variations, and System Approaches," published in AI Communications (IOS Press).

The scope of CBR spans systems that vary substantially in case representation format, similarity metric design, and adaptation mechanism. A case is a structured data record encoding a problem description, the solution applied, and (when available) the outcome. The case base may contain thousands to millions of records; IBM's early diagnostic CBR tools for hardware support operated case libraries exceeding 50,000 entries.

CBR differs from rule-based reasoning systems, which encode knowledge as explicit IF-THEN conditions. CBR does not require engineers to articulate rules in advance — an advantage in domains where expert knowledge is tacit, inconsistent, or too voluminous to express procedurally. The tradeoff is that CBR performance degrades when the case base is sparse, poorly indexed, or unrepresentative of the problem distribution the system encounters in deployment. For a broader structural map of where CBR fits, see the Reasoning Systems Defined reference on this site.

How it works

The operational core of a CBR system follows a four-phase cycle, each phase presenting distinct engineering challenges:

1. Retrieve — Given a new problem, the system queries the case base using a similarity metric. Common metrics include weighted Euclidean distance for numeric features and edit distance or TF-IDF cosine similarity for text features. The output is a ranked set of candidate cases. Retrieval accuracy depends heavily on feature selection and index structure; poorly weighted features cause semantically irrelevant cases to rank above relevant ones.

2. Reuse — The retrieved case's solution is adapted to fit the new problem. Adaptation may be null (the solution is applied unchanged), transformational (parameters are adjusted algorithmically), or generative (a new solution is synthesized guided by the retrieved case as a template). Null adaptation is the most common in deployed commercial systems because generative adaptation requires domain-specific repair knowledge that is expensive to encode.

3. Revise — The proposed solution is evaluated — either by a human expert, by outcome feedback, or by a secondary reasoning module. If the solution fails or is modified, the corrected version is recorded. This phase is where human-in-the-loop oversight is most commonly embedded, particularly in regulated domains such as healthcare and financial services, where explainability in reasoning systems is a compliance requirement.

4. Retain — The resolved case (problem + final solution + outcome) is stored in the case base. Retention policies govern what is stored: indiscriminate retention degrades retrieval performance by increasing noise and redundancy. Effective CBR systems apply case maintenance strategies to consolidate, prune, or weight cases over time.

This cycle is described in detail in the DARPA-funded Lockheed CBR research of the early 1990s, and the four-stage model remains the reference architecture cited by the American Association for Artificial Intelligence (AAAI) in technical literature on case-based approaches.

Common scenarios

CBR systems are deployed across a concentrated set of domains where historical precedent is dense and well-documented:

Legal and regulatory analysis — Legal research platforms use CBR to surface precedent cases matching the factual profile of a new matter. This mirrors the doctrine of stare decisis, making CBR a structurally natural fit. Systems in this domain must handle jurisdictional filtering and temporal precedent weighting. See Reasoning Systems: Legal and Compliance for sector-specific deployment patterns.

Medical diagnosis and clinical decision support — CBR systems in clinical settings retrieve prior patient cases with similar symptom profiles, laboratory values, and demographic features. The Veterans Health Administration has documented CBR-adjacent approaches in clinical decision support within its VistA electronic health record infrastructure. Reasoning Systems in Healthcare Applications covers the regulatory framing under FDA guidance on clinical decision support software.

Help desk and technical support — Enterprise IT service management platforms have used CBR since the mid-1990s to match incoming incident tickets to resolved cases. The approach reduces mean time to resolution by surfacing previously documented fixes without requiring a technician to search unstructured knowledge bases.

Financial services credit and fraud analysis — CBR supports anomaly detection by comparing transaction profiles to known fraud cases. The Federal Trade Commission's guidance on algorithmic decision-making in credit contexts, including adverse action notice requirements under the Equal Credit Opportunity Act (15 U.S.C. § 1691 et seq.), applies where CBR outputs influence credit decisions. See Reasoning Systems in Financial Services for compliance implications.

Supply chain exception handling — Logistics operators use CBR to recommend responses to disruption events — port closures, carrier failures, customs delays — by retrieving analogous disruption cases and their resolution paths. Coverage of supply chain deployment is at Reasoning Systems: Supply Chain.

Decision boundaries

CBR is not universally applicable. Four structural conditions define its appropriate operating range:

Condition 1: Case base adequacy. CBR requires a case library that is dense enough to cover the problem space. In domains where fewer than a threshold number of historical cases exist — or where cases cluster in narrow regions of the feature space — retrieval will consistently surface low-similarity matches, producing unreliable solution proposals.

Condition 2: Feature encodability. Problems must be representable as structured feature vectors or retrievable text. Problems involving continuous sensor streams, unstructured visual data, or highly contextual narrative require preprocessing pipelines or hybrid architectures before CBR retrieval is viable. Hybrid reasoning systems combine CBR with statistical or neural components to handle these inputs.

Condition 3: Adaptation tractability. When the gap between retrieved cases and new problems requires complex, multi-step adaptation, CBR systems without deep domain adaptation knowledge produce solutions that fail revision. In these scenarios, rule-based reasoning systems or model-based approaches may be more reliable.

Condition 4: Regulatory transparency requirements. CBR produces inherently interpretable justifications — "this decision is recommended because Case #4471 presented the same 7 features and was resolved by action X" — making it preferable to opaque statistical models in contexts governed by explainability mandates. The National Institute of Standards and Technology (NIST) AI Risk Management Framework (NIST AI RMF 1.0) identifies explainability as a core trustworthy AI characteristic, and CBR's case-trace audit trail directly satisfies that requirement in a way that probabilistic reasoning systems often cannot.

CBR also underperforms relative to machine learning models when problem volume is high and solution patterns are smooth and continuous rather than episodic. For a structured comparison, see Reasoning Systems vs. Machine Learning.

Procurement and deployment decisions for CBR systems should reference the broader Reasoning Systems Authority index for navigating the full landscape of available reasoning architectures and qualified vendors.

References

• Aamodt, A. & Plaza, E. (1994). "Case-Based Reasoning: Foundational Issues, Methodological Variations, and System Approaches." AI Communications, IOS Press.
• NIST AI Risk Management Framework (AI RMF 1.0)
• NIST Special Publication on AI Standards and Trustworthiness — NIST AI 100-1
• Federal Trade Commission — Reports on Algorithmic and AI Decision-Making
• Equal Credit Opportunity Act, 15 U.S.C. § 1691 et seq. — via Cornell Legal Information Institute
• AAAI — Association for the Advancement of Artificial Intelligence, Technical Publications
• Veterans Health Administration — VistA Electronic Health Record Program Documentation