Reference · AI Operations
50+ terms. Defined precisely. Mapped to the architecture.
Authored by InteractiveAIReviewed by domain experts running Production AI
53 terms
4 categories
AI Operations is the discipline of authoring, governing, and improving AI in production. This document is its working dictionary. Fifty-plus entries. Four navigation categories. Every term defined with the precision required for procurement diligence, executive briefing, and engineering specification. Imprecise vocabulary is the failure mode turning a Production AI programme into a pile of pilots. The Glossary fixes the vocabulary at its definitional root.
The Glossary is organised into four navigation categories: Category Foundations, Architecture, Operating Model, and Vocabulary. Each entry carries a hero sentence, a 100 to 150 word body, two to three cross-references to sibling entries, and one "Read further" link to the sister asset that carries the deeper argument. Readers arriving from search land on the entry they queried. Readers arriving from a sister asset land on the term that asset linked. Both paths converge on the same definitions.
A generic AI glossary lists generic AI terms. A framework reference defines framework primitives. This Glossary defines the operating model and the architecture together, with the domain expert as the protagonist who authors the AI. Authoring without governance is not Production AI. Engineering without authorship is not AI Operations. The discipline is defined by the union, not the sum.
Authored by the InteractiveAI team. Reviewed by domain experts running Production AI inside iGaming, Financial Services, and Travel and Hospitality operators. Every definition is written to hold in 2028. No dated metrics. No current-vendor references. No reference to the calendar year of publication.
Category 1
This category names the discipline. Before architecture, before operating model, before supporting vocabulary, the reader needs to know which category they are in. The eight entries below plant AI Operations as a discipline parallel to DevOps and MLOps, with its own primitives, lifecycle, and accountable role. Every other category in this Glossary depends on the definitions in this one.
AI Operations is the discipline of authoring, governing, and improving AI in production, parallel to DevOps and MLOps, with its own primitives, operating model, and accountable role inside the enterprise.
AI Operations is a discipline, with primitives, an operating model, and a lifecycle, not a product category and not a buzzword.
The AI Operations Lead is the role inside the enterprise accountable for the policy library, the evaluation suite, and the autonomous improvement loop that governs Production AI.
MLOps governs the lifecycle of models. AI Operations governs the lifecycle of authored AI behaviour. The two disciplines are complementary, not adversarial, and sit at different layers of the stack.
DevOps governs the deployment of code. AI Operations governs the authored decisioning that runs on top of deployed code, sitting one layer above DevOps in the production stack.
Production AI is the state in which an AI system runs against real users, with traceability, deterministic enforcement of critical policies, and structured evaluation at scale, governed under the AI Operations discipline.
A Compounding Asset is the authored policy library when the five architectural requirements hold and the autonomous improvement loop runs with human governance gates, accumulating value with every release.
Authored Orchestration is the method by which domain experts compose policies, routines, and the rest of the Eight Components into production-grade AI behaviour, distinct from prompt engineering and workflow automation.
Category 2
The architectural layer of AI Operations. Interactive Agents are InteractiveAI's proprietary class within AI Operations, defined by the Eight Components (Prompt, Policy, Routine, Glossary, Macro, Variable, Tool, Knowledge Base) and five architectural requirements (per-decision traceability, deterministic enforcement, EvalOps, autonomous improvement loop, sovereignty by construction). This category plants each primitive at its definitional root. Each entry is precise: a CTO or CAIO running procurement diligence can cite it without ambiguity. Sovereignty by Construction is treated as a first-class architectural requirement, not a residency footnote. The Eight Components are canonical as a group: each Component entry opens with the binding phrase "Component N of the Eight Components of an Interactive Agent."
Interactive Agent is InteractiveAI's proprietary class of AI within AI Operations, defined by Eight Components and five architectural requirements that together produce traceable, governable, improvable decisioning in production.
Dynamic Context Engineering is the per-decision assembly of the right context for the right decision at the right moment, executed on every turn an Interactive Agent makes.
Universal Router is the model-orchestration layer of the InteractiveAI Platform that selects the right model per decision against operator-authored Priorities including cost, latency, capability, and jurisdiction.
A Priority is the authored ordering principle the agent applies when Policies conflict on a decision or when the Universal Router selects between candidate models. It is a routing and conflict-resolution control, not one of the authored components.
Component 1 of the Eight Components of an Interactive Agent. A Prompt is the authored linguistic surface where authored components render into model input or user-visible output. Treating Prompts as components removes prompt engineering as a separate skill from AI Operations.
Component 2 of the Eight Components of an Interactive Agent. A Policy is an authored rule with a criticality tier that determines deterministic or probabilistic enforcement at runtime, expressed in natural language and operated against by the agent on every decision.
Component 3 of the Eight Components of an Interactive Agent. A Routine is an authored sequence that traverses adaptively at runtime, supporting backtracking and linking, governed by the Policies in scope.
Not to be confused with this document. Component 4 of the Eight Components of an Interactive Agent. The Glossary component is the authored vocabulary the agent uses to interpret terms inside its domain.
Component 5 of the Eight Components of an Interactive Agent. A Macro is a reusable block of authored message text, written once and interpolated wherever the agent needs consistent wording, governed centrally and owned by the domain expert who authored it.
Component 6 of the Eight Components of an Interactive Agent. A Variable is authored, not inferred. Named state the agent reads and writes during a session, exposing values to Policies, Routines, Prompts, and Tools.
Component 7 of the Eight Components of an Interactive Agent. A Tool is the authored interface to an external system, classified by consequentiality and governed by authored Tool Gating conditions.
Component 8 of the Eight Components of an Interactive Agent. The Knowledge Base is the authored retrieval surface that grounds the agent's answers in the operator's documents, optional by construction and governed for residency, freshness, and index scope.
Architectural requirement 1 of an Interactive Agent. Per-Decision Traceability means every decision the agent makes is inspectable down to the Policy, the Routine step, and the context that produced it.
Architectural requirement 2 of an Interactive Agent. Deterministic Enforcement means critical policies execute deterministically against decisions, not probabilistically through model adherence.
EvalOps is architectural requirement 3 of an Interactive Agent: the operating practice of structured evaluation at scale, run against authored Policy criticality and instrumented through LLM-as-a-Judge, evaluation suites, and per-turn evaluation.
Architectural requirement 4 of an Interactive Agent. The Autonomous Improvement Loop is the cycle by which an agent improves through authored refinement, gated by human governance.
Architectural requirement 5 of an Interactive Agent. Sovereignty by Construction means sovereignty is enforced through the architecture of the InteractiveAI Platform, not declared through residency policy alone.
The Policy Library is the operator's authored estate of Prompts, Policies, Routines, Glossary entries, Macros, Variables, Tools, and Knowledge Base, governed centrally and owned distributedly across the domain experts who authored them.
Category 3
The operating model is the how-it-runs layer of the discipline. Architecture defines what is built. The operating model defines who authors, who governs, who improves, and how the lifecycle moves through the organisation. These entries define the human and procedural structure that makes the architecture executable inside an enterprise. The domain expert is the protagonist. Co-Pilot is the analytic engine that surfaces patterns from production traces and drafts refinement proposals for the protagonist to review.
Domain Expert Authoring is the operating principle that the humans who understand the domain compose the AI's behaviour through the Eight Components, rather than translating their knowledge through engineers or prompt specialists.
Centralised Governance is the operating principle that the policy library is governed as one estate, under one set of standards, with one accountable function, regardless of how many domain experts author into it.
Distributed Ownership is the operating principle that the domain experts who author the policies own them across their lifecycle, while the central function governs the standard under which all authors operate.
The Human Governance Gate is the named checkpoint at which every refinement proposed by the autonomous improvement loop is reviewed by an accountable human before it merges into the live policy library.
Co-Pilot is the analytic engine on the InteractiveAI Platform that surfaces patterns from production traces and drafts refinement proposals. It performs the analytical work that previously required an engineering sprint: trace forensics, pattern extraction, refinement scoping, draft authoring. The domain expert remains the protagonist.
The Build-Connect-Deploy-Govern-Improve Lifecycle is the five-stage standard reference cycle of an Interactive Agent under AI Operations, run continuously rather than once per release.
Policy Criticality is the three-tier framework (Low, Medium, High) that determines the enforcement mode of an authored policy. High criticality requires deterministic enforcement. Low criticality permits probabilistic adherence.
Routine Adaptive Traversal is the property of an authored routine that allows the agent to adapt its path mid-execution based on context, rather than following a fixed sequence or a rule-based branch.
Backtracking is the authored capability of an Interactive Agent to revisit a prior routine step when newly available context changes the right answer, governed by the policies in scope at the step the agent returns to.
The Annotation Workflow is the structured process by which domain experts label production traces against authored evaluation criteria, generating the labelled corpus that feeds the autonomous improvement loop.
Agent Memory is the persistent state surface an Interactive Agent reads and writes across sessions, governed by authored policies and distinct from session-scoped variables that reset at each interaction boundary.
A Refinement Proposal is the artefact Co-Pilot drafts and a domain expert reviews, the unit of the autonomous improvement loop, gated by the Human Governance Gate before any merge into the live policy library.
Category 4
The supporting vocabulary of AI Operations. The Architecture and Operating Model categories rely on these terms, and procurement diligence and engineering teams use them when evaluating Production AI. Each entry is defined tightly and cross-linked upward to the primary entry it supports. These are reference terms, ordered alphabetically inside the category. The vocabulary is consistent with the canon: every term here resolves to its load-bearing parent in Architecture or Operating Model. Treat this category as the procurement-glossary surface, the place a CTO's team checks when an RFP response uses the term.
Adjacent Visibility is the authored capability that controls which context outside the current Routine step is visible to an Interactive Agent's decision.
A Consequential Tool is a Tool whose invocation produces an external effect, including payments, transactions, notifications, and writes to systems of record.
A Continuous Policy is a Policy defined by cadence: it evaluates on every turn of an Interactive Agent's session, not only at Routine entry or at named checkpoints.
An Evaluation Suite is the structured set of evaluations run against an Interactive Agent under EvalOps, bundled as a versioned artefact governed inside the policy library.
LLM-as-a-Judge inside AI Operations is the EvalOps instrument that scores Production Traces against authored criteria owned by the AI Operations Lead.
Per-Turn Evaluation is evaluation that runs at the granularity of a single conversational turn, not at session end, providing the unit-level grain of EvalOps.
A Persistent Policy is a Policy defined by scope: it applies across all Routines for an Interactive Agent, agent-wide rather than scoped to any single Routine or Routine step.
A Production Trace is the live production record of an Interactive Agent's decision-making, distinguished from development traces and sandbox traces, and the substrate of audit response and improvement-loop input.
A Routine-Scoped Policy is a Policy that applies only within a specified Routine, the counterpart to a Persistent Policy that applies agent-wide.
Routine Linking is the authored connection between two Routines that allows an Interactive Agent to traverse from one Routine into another without losing trace continuity or context.
A Session is the bounded interaction unit between an Interactive Agent and a counterparty, distinct from a single turn and distinct from the Trace that records the session.
Tool Gating is the set of authored conditions under which a Tool may or may not be invoked by an Interactive Agent inside a Routine.
Top-K Policy Injection is the mechanism by which the most relevant Policies for the decision at hand are surfaced into the decision context assembled by Dynamic Context Engineering.
A Trace is the full record of an Interactive Agent's decision-making for a Session, the substrate that per-decision traceability, EvalOps, and the autonomous improvement loop operate against.
Vector Retrieval inside AI Operations is one input source to Dynamic Context Engineering, not a standalone capability. It retrieves over embedded representations under authored governance for residency, freshness, and index scope.
Closing
The Glossary is the working vocabulary of AI Operations. Use it. Cite it. Ship against it. The terms that govern Production AI are the terms in this document.
The Glossary is the working dictionary of AI Operations. Ship against it, link to it, and bring the precision into your next procurement conversation.