Credit Union AI Glossary
Demystify AI technical, compliance, and governance definitions. Use credit union analogies to explain complex agentic concepts to board directors, managers, and staff alike.
Agentic AI
TechnicalAn AI architecture that uses large language models to reason, plan, select external tools, and execute multi-step processes autonomously rather than just responding to simple text queries.
Like hiring a virtual assistant who doesn't just read you the policy handbook but actually fills out the loan spreadsheet, double-checks the numbers, and files it for manager approval.
RAG (Retrieval-Augmented Generation)
TechnicalA method that searches internal documents (like policy manuals) to find factual reference texts, copying them into the AI prompt so it answers questions using real, up-to-date information rather than memory.
Like giving an employee an open-book exam. Instead of guessing from memory, they read the exact page in the loan guidelines before writing their answer.
PII (Personally Identifiable Information)
SecurityAny information that can be used to distinguish or trace an individual's identity, such as SSNs, names, account numbers, addresses, or biometric records.
The vault keys. Paste them in public AI systems and they are public. Always mask them first.
Human-in-the-Loop (HITL)
OversightAn operational design where an AI agent can execute low-risk tasks, but must pause and await explicit approval from a human employee before executing high-risk operations (e.g., approving a loan or wire transfer).
Like a new loan officer trainee. They compile the applicant file and run calculations, but the senior underwriter must physically sign the loan release.
System Prompt
OversightThe master instruction sheet given to an AI agent at initialization that defines its role, permissions, boundaries, and safety policies.
The official employee job description and code of conduct document given on day one.
Hallucination
TechnicalA phenomenon where an AI model generates responses that sound confident and grammatical, but are factually incorrect or unsupported by the source text.
Like a staff member who guesses an interest rate off the top of their head to sound helpful, instead of looking up the daily rate sheet.
Vector Database
TechnicalA database that stores text as mathematical coordinates (vectors), allowing AI models to find conceptually related ideas rapidly instead of matching exact keyword strings.
Like organizing the credit union filing cabinet by topics and concepts (e.g., placing all files about 'auto risk' together) rather than strictly alphabetical labels.
GLBA (Gramm-Leach-Bliley Act)
SecurityA federal law requiring financial institutions to explain their information-sharing practices and safeguard sensitive member data.
The privacy framework that makes sharing raw SSNs or tax files with third-party, non-compliant AI services illegal.
Model Drift
OversightThe decay of an AI model's accuracy over time due to changes in real-world data distributions or updates in systemic guidelines.
Like a lending policy getting outdated because economic conditions shift, making old underwriting scores unreliable under current market rates.
ReAct Framework (Reasoning + Action)
TechnicalAn agentic planning loop where the AI alternates between 'thinking' (reasoning about goals) and 'acting' (invoking database lookups or calculators) until a problem is solved.
Like a teller: 1. Think ('The member wants to withdraw $100'). 2. Act (Check account balance). 3. Observe ('Balance is $500'). 4. Think ('It is safe to withdraw'). 5. Act (Hand over money).