Glossary
Glossary
Plain-language definitions for operators building practical AI systems. No hype — terms you will meet in Discovery and in production.
Agent observability
Agent observability means you can see traces of what an agent did: prompts or tool calls, data scopes, latency, cost, and outcomes. Without it, incidents become blame games.
Agentic AI
Agentic AI means systems that pursue goals with tools and multi-step plans — not single-shot answers. It only pays off when scope, approvals, and evaluation are designed in from the start.
AI agent
An AI agent is software with a goal, tools, and limits — not a chatbot that only chats. It can triage, draft, or update systems under rules you set, with humans approving consequential steps.
AI Operating Layer
An AI Operating Layer is the production stack that runs practical agents on your data — open models by default, frontier when justified, MCP tools, and human approval on consequential actions — as one governed system.
Article 50 transparency
Article 50 of the EU AI Act covers transparency obligations for certain AI uses — such as making people aware they interact with AI when it matters. For many customer-facing agent uses, these duties apply from 2 August 2026.
Audit trail
An audit trail is a durable record of who or what accessed which resource, for what purpose, when, and with what approval. For AI agents it must include tool use and model routes, not only human logins.
Context window
The context window is how much text a model can consider at once (prompt plus history). Bigger is not always better: stuffing irrelevant context raises cost and can hurt accuracy.
Data lakehouse
A data lakehouse combines warehouse-style governance with lake-style flexible storage — often with bronze/silver/gold layers, catalog, and fine-grained access. For AI, it is where trusted tables and documents live so agents do not scrape unmanaged shares.
Data residency
Data residency is the requirement that certain data stay in a region or environment you control. For AI, it shapes which model routes are allowed: open models in your region versus vendor clouds.
Embedding
An embedding is a numeric representation of content that places similar meaning nearby in vector space. Embeddings power semantic search and RAG.
EU AI Act
The EU AI Act is the European Union's risk-based law for AI systems. Obligations differ by use and risk class.
Evaluation (evals)
Evals are structured tests of agent or model behaviour against examples and criteria — accuracy, tone, cost, refusal, tool use. Without evals, "it worked in the demo" is the only evidence.
FinOps
FinOps is the practice of making cloud and AI spend visible, attributable, and decided — not just paid. For AI, it means knowing cost per workflow, per model route, and per environment, then changing routing or scope when the numbers do not match value.
Frontier model
A frontier model is a leading closed or high-capability model typically offered as a vendor API. It often costs more per call than capable open-weight routes.
Guardrails
Guardrails are technical and process limits that keep agents inside allowed behaviour: tool allowlists, data scopes, output filters, rate limits, and escalation paths. They complement human approval — they do not replace clear ownership of who may change the rules.
Hallucination
A hallucination is a fluent model output that is false or unsupported. In operations, the risk is not poetry — it is inventing a booking, a policy, or a customer fact.
Human-in-the-loop
Human-in-the-loop means a person must approve or correct certain steps before they take effect — especially outbound messages, payments, deletions, or policy exceptions. It is not a failure of automation; it is how you keep agents safe where mistakes are costly.
Inference
Inference is the act of running a model to produce an output for a request. For operators, inference is the recurring cost and latency of automation — not the one-time build.
LLM
A large language model (LLM) predicts text tokens from context. Operators use LLMs inside workflows — triage, drafts, extraction — not as the whole product.
MCP
MCP (Model Context Protocol) is a standard way for agents to talk to tools and data sources through consistent interfaces. For operators, it reduces one-off glue code and makes least-privilege tool access easier to reason about — when servers are scoped and logged like any other integration.
Model routing
Model routing sends each workflow or step to the model that fits — open for volume, frontier when justified, region-constrained routes when data requires it. A gateway sits between agents and providers so you can swap models, track spend, and keep policy in one place.
Open-weights model
An open-weights model publishes its parameters so you can run it yourself or via a host you choose. For operators, the point is control: region, cost, and the option to change hosts without rewriting every workflow.
PBAC
PBAC (purpose-based access control) grants access for a stated purpose and time, rather than standing privileges that linger. For agents, it means a triage bot may read inbox fields for classification but not export the whole CRM.
Prompt caching
Prompt caching reuses stable prompt prefixes (policies, schemas, long instructions) so repeated calls pay less and run faster. For operators with high-volume similar requests, cache hit rate often moves the bill more than switching models.
RAG
RAG (retrieval-augmented generation) fetches relevant documents or records before the model answers, so responses ground in your content rather than memory alone. It needs clean retrieval, permissions, and evaluation — not only a vector database.
Token
A token is a chunk of text the model reads or writes — roughly a word piece. Pricing and context limits are often expressed in tokens.
Vector database
A vector database stores embeddings so you can find similar text, images, or records by meaning. It is infrastructure for search and RAG, not a substitute for access control or retention policy.
Yerbabuena Digital
Ready to plant something good together?
Tell us what season you are in. We will walk the field with you — honestly — and suggest the smallest next step that can take root.