Agent
An AI-enabled system designed to complete a defined job using instructions, knowledge, tools, and workflow logic. The word does not guarantee autonomy or reliability.
Plain-language AI glossary
Useful definitions.
No performance theater.
A practical vocabulary for teams discussing custom AI workflows, knowledge, actions, controls, testing, and measurement.
Grounding
Connecting an AI response to approved context or source material so the answer is based on relevant business information rather than general model memory alone.
Retrieval
Finding relevant source material at request time and supplying it to the model. Retrieval quality depends on content, structure, access, indexing, and the user question.
Guardrail
A rule, technical control, confirmation step, permission boundary, or review process intended to reduce unwanted behavior or impact.
Human handoff
A defined transfer from the AI system to a person, with enough context for the person to continue the work effectively.
Evaluation
A repeatable method for testing output or behavior against expected criteria, examples, policies, sources, or reviewer judgments.
Hallucination
A confident-sounding output that is unsupported, incorrect, or invented. Grounding and testing reduce risk but do not eliminate it.
Tool use
When an AI system calls an approved function or connected service to retrieve data or take an action. Access should be scoped to the workflow.
Human in the loop
A design in which a person reviews, approves, corrects, or takes over work at defined points instead of leaving the system fully unattended.
Pilot scorecard
A small set of agreed quality, operational, customer, and business measures used to decide whether a pilot should stop, change, or scale.