Abstention
A controlled system response that declines to answer or act when evidence, authorization, confidence, or scope is insufficient.
Acceptance criteria
Predefined quality, value, risk, security, cost, and operating thresholds used to decide whether a pilot or release passes.
Agentic system
A system that uses models and software tools to plan or execute multiple steps toward a goal. Intelligence724 treats tool authority and side effects as the critical control boundary.
AI system register
An inventory of models, applications, data, tools, owners, uses, risks, evidence, changes, incidents, and lifecycle status.
Architecture decision record
A durable record of the decision, context, options, evidence, tradeoffs, owner, conditions, and review triggers for a system architecture choice.
Bounded agency
Machine action limited by explicit process scope, identity-aware permissions, tools, budgets, approval gates, monitoring, and rollback.
Build, buy, configure, defer
Four sourcing paths that allocate product, integration, operating, control, and lifecycle responsibility differently.
Client-owned evaluation set
Representative test cases and expected outcomes controlled by the client so providers can be compared consistently and changes can be regression-tested.
Control plane
The policies, identity, permissions, budgets, approvals, monitoring, logging, and change mechanisms that constrain system behavior.
Cost per successful outcome
Allocated total cost divided by outcomes meeting the business acceptance definition—not merely cost per token, request, or seat.
Critical error
An error capable of causing defined material legal, financial, safety, privacy, customer, or operational harm.
DMAIC
Define, Measure, Analyze, Improve, and Control—the process-improvement cycle underlying the Intelligence724 method.
Evidence confidence
A separate rating of how much of a score or decision is supported by verified client facts rather than assumptions or unresolved unknowns.
Evidence ledger
A structured record separating verified facts, consultant inferences, unresolved unknowns, and client-retained decisions.
Ground truth
The authoritative answer, outcome, or expert adjudication procedure used to evaluate a system.
Human in the loop
A workflow that pauses before a material action until a qualified person approves the exact action with enough context and authority.
Human on the loop
Low-risk, observable, reversible operation that a qualified person monitors and can stop in practice.
Knowledge assurance
Governed source stewardship, identity-aware retrieval, evidence-based answers, evaluation, correction, and production operation—not simply a chatbot.
Least intelligence necessary
The rule that the simplest intervention meeting the business and control requirement should be preferred.
Model change risk
The risk that provider, version, policy, behavior, price, or availability changes invalidate prior evaluation and operating assumptions.
Process baseline
A reproducible measure of the current workflow before intervention, including volume, time, quality, error, rework, cost, experience, and control effort.
Prompt injection
Untrusted content attempting to alter model behavior or induce unauthorized disclosure or tool action. It must be treated as an authorization and architecture problem, not only a wording problem.
RAG
Retrieval-augmented generation: a system that retrieves external evidence for a model. Retrieval alone does not guarantee authority, authorization, correctness, freshness, or citation quality.
Residual risk
Risk remaining after controls, with a named owner authorized to accept, remediate, transfer, avoid, or stop the use.
Rollback
The tested ability to return to a known-safe system, process, data state, model version, or manual workflow.
Scale gate
A decision point requiring verified value, quality, operating readiness, control evidence, ownership, and acceptable economics before expansion.
Source authority
A business-governed classification that determines whether a source may establish fact or policy within a defined scope.
System-level evaluation
Testing the complete workflow—people, data, retrieval, prompts, tools, interfaces, infrastructure, permissions, and controls—rather than only a base model.
Vendor neutrality
The ability to compare suppliers, custom development, open components, conventional automation, and no purchase under disclosed commercial incentives.
