Vocabulary: Governance, Security, and Azure

Responsible AI — the engineering practice of building AI that is safe, transparent, auditable, and accountable: advisory outputs, human review, measurement, and controls. Lives in: the duty statement and every state AI document. Say it: "I treat responsible AI as a design requirement, not a compliance afterthought — advisory output, human review, audit logging, measured accuracy."

Advisory vs. determinative — does the system inform a decision or make it? The load-bearing distinction for government AI; wording is part of the architecture. Lives in: responsible-AI design; now also in statute (§ 44259.8's "advisory analysis"). Say it: "'Appears to meet' is a different legal object than 'you are qualified' — we design the wording, not just the model."

Human-in-the-loop (HITL) — humans review, approve, or override AI outputs — especially low-confidence or high-stakes ones — with an escalation path that's cheap to use. Lives in: responsible-AI design; statutory language ("technology-supported analysis and human review"). Say it: "The AI diagnoses; humans act — and if overriding the AI is harder than accepting it, the loop is theater."

Prompt injection (direct and indirect) — untrusted content treated as instructions. Direct: the user smuggles instructions. Indirect: instructions hide inside content the system must read — an email, a web page, a document in the corpus. Lives in: AI security; the SQL injection of the LLM era. Say it: "Anything the model reads is an input channel — so we defend it like untrusted input: least privilege first, then delimiting, then output validation."

Audit logging — recording every AI invocation: who, when, what context went in, what came out, what action followed. Lives in: the control plane; your incident-response and public-records answer. Say it: "Every AI call is logged with its full context — we can reconstruct any answer the system ever gave."

Feature toggle (kill switch) — runtime control to disable an AI capability instantly, without a deployment. Lives in: the control plane. Say it: "When leadership asks 'can we turn it off,' the answer is a click — and when our pilot API key expired, the platform proved it by degrading gracefully."

Fail closed — when the AI is down, degraded, or uncertain, the system does the safe thing — route to a human, show the deterministic result — rather than guessing. Lives in: reliability design for AI features. Say it: "Uncertainty routes to people; the system never improvises under failure."

PII / data classification — personally identifiable information, and the tiering of data (public, internal, sensitive, federally protected like FERPA education records) that determines handling. Lives in: privacy architecture; drives model routing. Say it: "The data classification decides where the model runs — protected records argue for our government tenancy or local inference; public corpus content can use commodity services."

Data governance — policies for data quality, lineage, access, retention, and — for AI — what may enter prompts, indexes, and training sets. Lives in: the 20% duty ("data quality, model transparency, security, privacy, compliance"). Say it: "AI governance inherits data governance — the index is a data store and gets the same quality and access controls."

Model transparency — being able to explain what a model does, what it was built from, and why it produced a given output (citations for LLMs, feature importance for classical). Lives in: the duty statement, verbatim. Say it: "Transparency looks different per tier — citations for grounded LLM answers, feature importances for classifiers — but every output must be explainable to an auditor."

NIST AI Risk Management Framework — the U.S. reference vocabulary for AI governance: Govern (policies, accountability, inventory), Map (context and risks per use case), Measure (evaluate accuracy, bias, robustness), Manage (mitigate, monitor, respond). Lives in: governance frameworks; rhymes with California's guidance. Say it: "I'd frame it with the NIST AI RMF — map the use case and risks, measure before launch, manage with review and a kill switch, under our existing governance."

EO N-12-23 / CDT GenAI guidance — California's executive order and the Department of Technology guidance flowing from it: risk assessment keyed to public impact, procurement rules, AI inventories, pilots before production. Lives in: the state policy layer. Say it: "Any AI feature here starts from the state's GenAI guidance — documented use case, risk assessment, inventory entry, and an approved procurement path."

Azure OpenAI Service — OpenAI models hosted in Microsoft's cloud under enterprise agreements: your tenancy, contractual data protection, prompts not used for training. Lives in: the approved-cloud path for LLM work in a Microsoft shop. Say it: "Azure OpenAI gives us frontier models inside our own tenancy with data-protection terms — that's what makes cloud LLMs procurable here."

Azure AI services (Cognitive Services) — the managed pre-built AI APIs: vision, speech, translation, and Document Intelligence — OCR and structured extraction from forms and documents (the transcript-intake tool). Lives in: the duty statement's named services. Say it: "For transcript extraction I'd start with Document Intelligence — managed OCR with per-field confidence, already in our cloud."

Azure Machine Learning (ML studio) — the workbench for classical ML: train, register, version, deploy, and monitor models; includes the model registry and evaluation pipelines. Lives in: the duty statement's named services; the MLOps home for the classical tier. Say it: "LLM work runs through Azure OpenAI; classical models — like a processing-time predictor — live in Azure ML with its registry and monitoring."

Azure AI Search — the managed retrieval service: indexes documents, supports hybrid keyword-plus-vector search; the retrieval half of enterprise RAG on Azure. Lives in: the Azure RAG reference architecture. Say it: "Azure AI Search plus Azure OpenAI is the standard enterprise RAG pairing — retrieval as a governed, managed service."

Reference architecture — the blessed pattern for a class of problem; new designs follow it or explicitly justify divergence. Lives in: enterprise architecture; how you align with the Systems Architect. Say it: "I'd bring AI capabilities in through our reference architecture — same auth, logging, and deployment standards as everything else. No shadow AI stack."

ADR (architecture decision record) — a short written record of a significant decision: context, options, decision, consequences — making review cheap and history durable. Lives in: enterprise architecture practice. Say it: "For AI initiatives I'd bring the Systems Architect ADR-style proposals — a concrete artifact to review, not a hallway pitch."