Enterprise AI Glossary: 200+ Terms Defined

Enterprise context: Never use accuracy as the sole metric for credit default, fraud detection, medical diagnosis, or any domain with significant class imbalance. Use precision, recall, AUC-ROC, or F1 instead.

Enterprise context: Required under ECOA (credit), FCRA, and equivalent regulations in other jurisdictions. AI-generated adverse action notices must use model-specific reason codes, not generic statements. SHAP values are commonly used to generate these codes.

Enterprise context: Agentic AI introduces qualitatively different governance requirements compared to single-turn LLM applications — specifically around tool authorization (what the agent can do), human-in-the-loop design (when the agent must seek approval), and audit logging (traceable records of all agent actions).

Enterprise context: The EU AI Act applies to any organization deploying AI that affects EU residents, regardless of where the organization is headquartered. High-risk system operators must maintain technical documentation, conduct conformity assessments, register systems in the EU database, and establish post-market monitoring.

Enterprise context: The most common AI CoE failure mode is becoming an "ivory tower" — building capabilities without business unit engagement. The most effective CoEs prioritize time-to-value for business units over internal platform perfection.

Enterprise context: AI governance is distinct from data governance and IT governance, though it intersects with both. The most common gap is a governance framework that exists as a written policy but is not operationalized into the model development and deployment workflow.

Enterprise context: 73% of AI implementation failures trace to readiness gaps identified in retrospect but not assessed upfront. A structured readiness assessment before program launch reduces average implementation delay from 8.4 months to under 3 months.

Enterprise context: Risk classification determines the level of model documentation, testing, validation, and governance oversight required. Systems classified as high-risk under internal frameworks should align with EU AI Act high-risk classification criteria to avoid regulatory surprises.

Enterprise context: Anomaly detection models are highly sensitive to the definition of "normal" at training time. Models trained on pre-COVID transaction patterns, for example, may generate excessive false positives during unusual economic periods. Regular recalibration of baseline distributions is essential.

Enterprise context: API-based access to LLMs (OpenAI, Anthropic, Google) introduces data privacy, latency, cost, and vendor dependency considerations. Most enterprise use cases require data processing agreements confirming that API inputs are not used for model training.

Enterprise context: AUC is threshold-independent, making it the preferred metric for comparing models when the optimal operating threshold depends on business context (e.g., the relative cost of false positives vs. false negatives in fraud detection or credit underwriting).

Enterprise context: AutoML is effective for well-defined tabular prediction problems with clean data. It is less appropriate for complex feature engineering requirements, specialized model architectures, or use cases requiring deep customization. AutoML output must still pass model risk governance review.

Enterprise context: Establishing a baseline is one of the most commonly skipped steps in enterprise AI development. Without it, teams cannot determine whether their model is actually adding value or simply codifying existing rules in a more opaque form.

Enterprise context: Most enterprise AI use cases are batch, not real-time — demand forecasting, churn prediction, credit scoring in underwriting queues, and document classification are common examples. Real-time inference adds significant infrastructure cost and complexity that batch does not require.

Enterprise context: Model bias is both a regulatory risk (disparate impact requirements in credit and employment) and a business risk (customer churn, reputational damage, regulatory action). Bias assessment must be conducted before production deployment and monitored continuously.

Enterprise context: In enterprise RAG deployments, BM25 combined with dense vector search in a hybrid retrieval architecture handles 20-30% of queries more accurately than dense-only approaches. Critical for regulatory citation, product code, and proper noun queries.

Enterprise context: The most common reason AI business cases fail CFO review is an incomplete cost model — typically missing data infrastructure, integration, governance, change management, and ongoing monitoring costs. See our AI ROI Guide for a complete TCO framework.

Enterprise context: Required by SR 11-7 for material model updates in financial services. Champion/challenger testing reveals production performance differences that validation testing on historical data cannot — particularly for distribution shift and edge case behavior.

Enterprise context: Hierarchical chunking with structure-aware document parsing is recommended for contracts, financial reports, and regulatory documents. Naive fixed-size chunking is inappropriate for these document types, as it destroys clause context and cross-reference relationships.

Enterprise context: A larger context window does not mean performance is equal across the full window. Most LLMs exhibit degraded accuracy for information located in the middle of very long contexts ("lost in the middle" phenomenon). Retrieval-augmented approaches remain valuable even with large context windows for precision and cost reasons.

Enterprise context: Organizations running more than five models in production require automated CI/CD-ML pipelines to manage the complexity of simultaneous retraining, validation, and deployment cycles. Manual processes do not scale beyond this threshold.

Enterprise context: Data drift typically precedes accuracy degradation by 4-8 weeks. Monitoring input feature distributions and alerting on PSI above 0.2 provides an early warning that enables proactive retraining rather than reactive incident response after business metrics have already been affected.

Enterprise context: A data lake alone does not make an organization AI-ready. The most common data lake failure for AI use cases is absent data quality standards — the lake contains the data, but it cannot be relied upon without significant cleaning and validation before model training.

Enterprise context: Disparate impact in AI models used for credit underwriting, employment screening, insurance pricing, or housing decisions is subject to legal challenge under ECOA, Title VII, Fair Housing Act, and equivalent regulations. Regular disparate impact analysis is required, not optional.

Enterprise context: The quality of your embedding model has a larger impact on RAG system accuracy than most teams expect. General-purpose embeddings underperform domain-specific embeddings for specialized vocabulary (legal, medical, financial). Domain-fine-tuned embeddings show 12-18% recall improvement in these domains.

Enterprise context: SHAP (SHapley Additive exPlanations) is the most widely used local explainability method in enterprise contexts because it provides consistent feature attribution values that satisfy adverse action notice requirements and audit documentation needs.

Enterprise context: Experiment tracking is not optional for regulated industries. SR 11-7 requires that model development decisions be documented and reproducible. Untracked experiments mean that a model's development cannot be recreated for model risk review or regulatory examination.

Enterprise context: Faithfulness is the primary hallucination metric for enterprise RAG systems. It should be measured both on a labeled evaluation set during development and continuously on production samples. A faithfulness drop of more than 0.05 on a 7-day rolling average should trigger investigation.

Enterprise context: Feature stores add significant infrastructure cost and operational complexity. They are justified when: multiple teams are building models on similar data, real-time feature serving is required for latency-sensitive use cases, or the organization has more than 8-10 models in production consuming similar feature sets.

Enterprise context: Fine-tuning is appropriate when RAG cannot achieve the required quality (typically when the task requires internalized domain knowledge rather than retrieved facts), when inference cost at scale makes RAG economically unviable, or when the organization cannot send data to external APIs for privacy reasons. Fine-tuning requires high-quality domain training data, which is often the binding constraint.

Enterprise context: The risk profile of hallucinations varies significantly by use case. A hallucination in an internal knowledge base assistant is an inconvenience; a hallucination in a clinical decision support tool or a legal document review system is a significant patient safety or professional liability risk. Use case risk classification should drive hallucination mitigation investment.

Enterprise context: The EU AI Act requires human oversight for high-risk AI systems. The practical design question is not whether to include humans but when and how — specifically, what the uncertainty threshold is for automatic escalation, what information humans need to review AI recommendations effectively, and what the workflow is for disagreement between human and AI judgments.

Enterprise context: LLM selection for enterprise deployment should be based on task-specific performance benchmarks conducted on your actual data — not on public leaderboard scores, which measure general capabilities and may not predict performance on your specific domain vocabulary and task requirements.

Enterprise context: SR 11-7 applies to all "models" as broadly defined — including ML and AI systems used in credit decisions, risk measurement, capital planning, and stress testing. Financial services firms deploying AI must build model development processes that satisfy SR 11-7 validation requirements from the start, not as a retrofit.

Enterprise context: Multi-agent systems multiply the governance complexity of single-agent systems. Each agent-to-agent interaction is a potential failure point, and the combined system can exhibit emergent behaviors not present in individual agents. Comprehensive audit logging of inter-agent communications is essential for production deployments.

Enterprise context: Organizations with five or more models in production require a formal MLOps platform. Teams without MLOps infrastructure spend an estimated 40-60% of ML engineer time on manual operational tasks that should be automated, significantly reducing capacity for model development.

Enterprise context: Prompt engineering should be version-controlled and treated as software — not ad-hoc text. Prompt regression testing (validating that prompt changes do not degrade performance on a standard evaluation set) is necessary for production LLM applications.

Enterprise context: Prompt injection is the most significant security concern for enterprise LLM deployments, especially agentic systems with tool access. Mitigation requires input sanitization, output filtering, privileged instruction separation, and limiting the scope of tools available to AI agents.

Enterprise context: RAG is the default architecture for most enterprise GenAI use cases involving proprietary documents — contract review, policy Q&A, knowledge base assistants, regulatory guidance lookup. Naive single-stage RAG implementation consistently underperforms due to poor chunking and lack of retrieval optimization.

Enterprise context: Most large organizations have published responsible AI principles. The gap is in operationalization — translating principles into specific controls, checklists, review gates, and monitoring requirements that are embedded into standard development and deployment workflows.

Enterprise context: Shadow mode deployment should run for a minimum of two weeks before production cutover to catch input distribution differences, integration edge cases, and performance anomalies that only appear on real data. Teams that skip shadow mode under schedule pressure consistently encounter production surprises in week one of go-live.

Enterprise context: SHAP is the industry standard for adverse action reason codes in credit models, satisfying ECOA requirements for specific, accurate reason statements. SHAP is also required documentation under many model risk management frameworks for the validation of high-impact models.

Enterprise context: SR 11-7 compliance is not optional for US bank holding companies and national banks. Financial services AI programs must design model development processes (training documentation, experiment tracking, validation evidence) to satisfy SR 11-7 requirements from inception, not as a retrofit after development is complete.

Enterprise context: Target leakage is the most commonly missed data quality issue in enterprise AI implementations. It is found in 44% of programs we audit. The impact is severe: validation metrics overstate production performance by 20-40%, leading to model risk approval on inflated results. A structured leakage audit before training prevents this entirely.

Enterprise context: Vector database selection should be based on corpus size, deployment model (managed cloud vs. on-premises), filtering requirements, and performance benchmarks on your specific data. Recall quality at production scale differs significantly across vendors — our RAG Architecture Guide contains benchmarks at 10M, 100M, and 1B vector scales.

Enterprise context: Zero-trust principles are particularly important for agentic AI systems that have access to tools (email, databases, file systems, APIs). Without zero-trust controls, a compromised or misbehaving agent can cause outsized harm before detection.