AI is software that infers how to produce predictions, content, recommendations or decisions, and it runs on a stack of models, data, tools and vendors. Governance starts by knowing what counts as AI, how it is built and run, who owns which layer, and whether the organization is ready to govern it.
An AI system is software that works out for itself how to turn inputs into outputs, instead of only following rules a person wrote line by line. Think of a spam filter that learned from millions of emails versus one that blocks a fixed list of words: the first infers, the second just executes. That ability to infer, with some level of autonomy, is what puts a system inside AI governance scope.
diagram
The OECD and EU AI Act definition in one line: objectives and input go in, the system infers, and its outputs influence an environment. Inference is what separates AI from rule-only software.
Key points
The Organisation for Economic Co-operation and Development (OECD) defines an AI system as a machine-based system that, for explicit or implicit objectives, infers from its input how to generate outputs.
Those outputs are predictions, content, recommendations or decisions that can influence physical or virtual environments; the OECD added content in November 2023 to cover generative AI.
The EU AI Act, Regulation (EU) 2024/1689, Article 3(1), mirrors the OECD wording and adds varying levels of autonomy and possible adaptiveness after deployment.
EU AI Act Recital 12 makes the capability to infer the key characteristic and excludes systems based only on rules defined solely by people to execute operations automatically.
Inference techniques include machine learning that learns from data and logic- and knowledge-based approaches that reason from encoded knowledge (EU AI Act Recital 12).
The NIST AI Risk Management Framework (AI RMF) adapts its AI system definition from the 2019 OECD Recommendation and ISO/IEC 22989:2022, the AI concepts and terminology standard.
The Adopt, Defend, Govern (ADG) framework covers every AI system in this sense, from classical machine learning to diffusion, multi-modal and composite systems; a model is only one part of a system.
Exam tip
When a question asks what makes something AI rather than ordinary software, pick inference, meaning deriving outputs or models from data, not automation alone. A rules-only system written entirely by people is the classic distractor.
Machine learning is how most AI gets its skill: instead of being told the rules, it studies examples. Predictive AI sorts or scores things, such as flagging a risky loan, while generative AI creates new text, images, audio or code. Most generative AI sits on a foundation model, one huge general-purpose model that many organizations adapt, like a single engine fitted into many different cars.
diagram
Predictive AI learns from labeled data, unlabeled data or rewards to classify, score and forecast. Generative AI usually starts from a self-supervised foundation model that teams adapt through fine-tuning, RAG or agents.
Key points
NIST AI 100-2 E2025 splits AI systems into predictive AI (PredAI) and generative AI (GenAI), and notes that most industrial AI applications are still predictive.
Supervised learning trains on labeled data; classification predicts discrete classes and regression predicts continuous values.
Unsupervised learning uses unlabeled data, semi-supervised learning mixes a few labels with many unlabeled samples, and reinforcement learning has an agent learn a policy that maximizes reward.
Foundation models are pre-trained on large-scale data with self-supervised learning, then aligned with supervised fine-tuning (SFT) or reinforcement learning from human feedback (RLHF).
Teams adapt a foundation model by fine-tuning it or wrapping it in a retrieval-augmented generation (RAG) application or an agent, so one model's weaknesses can reach many applications.
The EU AI Act defines a general-purpose AI (GPAI) model as one with significant generality that competently performs a wide range of distinct tasks and can be integrated into many downstream systems.
The Adopt, Defend, Govern (ADG) framework defines an AI agent as an AI system that can autonomously plan, run multi-step tasks, invoke tools and take actions for a user or organization.
Exam tip
Match the learning type to the scenario: labeled examples point to supervised learning, a reward signal to reinforcement learning. For governance, a flaw in a shared foundation model is a supply chain risk for every application built on it (see 3.2).
The same AI technique can be harmless in one job and dangerous in another, so governance judges the use, not the algorithm. A model that ranks movie suggestions and one that ranks job applicants may be built almost the same way, yet the second can shape someone's livelihood. That is why regulators and frameworks sort AI by what it is used for and how much it acts on its own.
Predictive AI
Generative and agentic AI
Typical use cases
Scoring and ranking
Credit scoring, hiring, fraud flags
Forecasting demand
Drafting and summarizing
Coding and search assistants
Agents acting through tools
Governance watch-outs
Annex III areas such as employment and credit
Unequal outcomes for groups
Inventing facts or policies
Choosing HITL, HOTL or HOOTL
The technique matters less than the use. Predictive scoring of people lands in EU AI Act Annex III areas, while generative and agentic uses raise questions of invented answers and how much the system acts on its own.
Key points
EU AI Act Annex III high-risk areas include biometrics, critical infrastructure, education and vocational training, employment and workers' management, and access to essential private and public services.
The remaining Annex III areas are law enforcement, migration, asylum and border control management, and administration of justice and democratic processes.
Under the Digital Omnibus on AI, Regulation (EU) 2026/1744, Annex III high-risk obligations now apply from 2 December 2027 instead of 2 August 2026.
Some uses are banned outright by EU AI Act Article 5, such as social scoring and manipulative techniques that cause significant harm; these prohibitions have applied since 2 February 2025.
The Adopt, Defend, Govern (ADG) framework sets three autonomy tiers: human in the loop (HITL), human on the loop (HOTL) and human out of the loop (HOOTL).
ADG examples: an email-drafting copilot is HITL, a trading agent acting within position limits is HOTL, and an overnight data-pipeline agent reporting in the morning is HOOTL.
Use drives harm: a copilot inventing a refund policy is an operational harm, while a loan agent denying one group credit more often is a societal harm in the ADG harm taxonomy.
Exam tip
When two answer options describe the same algorithm, choose the one that ties risk to the use case and the people affected. Governance classifies the deployment context and the level of autonomy, not the model type.
Building AI is less like shipping a finished product and more like tending a crop: you plan, gather data, grow a model, check it, put it to work, and keep watching because conditions change. Models can drift as the world moves away from the data they learned from, so operations only end when the system is retired. Every stage needs testing and evidence, not just a final check before launch.
diagram
The lifecycle as NIST and the OECD describe it, with retirement added by the OECD. The loop back shows why it is iterative: drift or a new need sends the system back to planning.
Key points
The NIST AI Risk Management Framework (AI RMF) lifecycle stages are Plan and Design, Collect and Process Data, Build and Use Model, Verify and Validate, Deploy and Use, and Operate and Monitor.
NIST adds a Use or Impacted by dimension for people and planet, and says risk management should ideally start in Plan and Design and continue throughout the lifecycle.
The Organisation for Economic Co-operation and Development (OECD) lifecycle adds retire or decommission and treats phases as iterative, not strictly sequential.
Test, evaluation, verification and validation (TEVV) runs through every stage, from audit and impact assessment at design to integration and compliance testing at deployment.
AI systems may need more frequent corrective maintenance than ordinary software because of data, model or concept drift (NIST AI RMF Appendix B).
ISO/IEC 5338:2023 defines AI system life cycle processes, building on ISO/IEC/IEEE 15288 and 12207 with AI-specific processes from ISO/IEC 22989 and ISO/IEC 23053.
Adopt, Defend, Govern (ADG) Minimum Control 5 (MC-5) puts changes to prompts, tools, models and retrieval sources through governed change control with approval records.
Exam tip
If an option treats testing as a single pre-launch gate, it is the distractor; NIST places TEVV across the whole lifecycle. Drift is the reason operations and monitoring never stop while the system is live.
Pin the validated model and data, then sign the release · bash
# Record hashes of the exact model and dataset that passed validation
sha256sum models/contoso-credit-v3.onnx data/train-2026-10.parquet > release-v3.sha256
# Re-check before deployment; any changed file fails the gate
sha256sum -c release-v3.sha256
# Signed tag ties the approved version to an accountable approver
git tag -s model-v3.0.0 -m "Approved for release by the AI Governance Council"
git tag -v model-v3.0.0
An AI product is rarely one model. It is a stack: the model, the data it retrieves, the tools and connectors it can call, the guardrails around it, and the logs that record what happened. Some layers you build, some you rent from a model provider, and some arrive inside a vendor's software, like a building where you own some floors, lease others, and share the lobby.
Foundation model API
SaaS or embedded AI
Organization owns
Prompt, Context, Tools
Orchestration, Identity
Safety Layer, Telemetry
Context, Identity
Telemetry
Governance policy
Vendor owns
Model training and alignment
Infrastructure and API availability
Model, Orchestration, Tools
Runtime stack, Learning Loop
Who owns which ADG governance surface when AI is consumed as an API versus embedded in a vendor product. A homegrown stack owns all nine, and shared items such as model governance or the safety layer sit between the two parties.
Key points
The Adopt, Defend, Govern (ADG) technology layer covers a model gateway, retrieval and memory, a tool registry, guardrails and harm detection, telemetry and replay, training and evaluation pipelines, and input checks.
The Model Context Protocol (MCP) is an open-source standard for connecting AI applications to external data sources, tools and workflows; ADG governs MCP servers as data access paths.
ADG names nine governance surfaces: Model, Prompt, Context, Tools, Orchestration, Identity, Safety Layer, Telemetry and Learning Loop.
With a homegrown, self-hosted stack the organization owns all nine surfaces; with a foundation model application programming interface (API) it owns seven and shares model governance and the learning loop.
With software as a service (SaaS) or embedded AI the organization keeps Context, Identity, Telemetry and its governance policy, while the vendor owns the model, orchestration, tools, runtime and learning loop.
Ownership shifts toward the vendor but accountability for use never fully transfers; ADG Minimum Control 12 (MC-12) requires documented responsibility boundaries and vendor due diligence (see 3.2).
In generative AI, data collection, model training, deployment and application development are often done by different organizations, which widens the AI supply chain (NIST AI 100-2 E2025).
Exam tip
In shared-responsibility questions, a vendor can take over operating a layer, but the deploying organization stays accountable for how the AI is used. Pick the answer that keeps consumer-side telemetry and governance policy in house.
Before an organization scales AI, it needs to know what AI it already has, who owns each system, and whether its people understand the tools they use. Readiness is like checking the foundations before adding floors: an inventory, named owners, trained staff, and a body that can say yes or no. Without that base, every later control has nothing to attach to.
diagram
The ADG four-phase roadmap moves from inventory and ownership to continuous assurance in about 24 months. Foundation delivers MC-1 to MC-3, and Control Deployment rolls out MC-4 to MC-9 plus MC-11 and MC-12.
Key points
EU AI Act Article 4 on AI literacy has applied since 2 February 2025 to every provider and deployer of AI systems, not only those using high-risk AI.
The Digital Omnibus on AI, Regulation (EU) 2026/1744, in force since 27 July 2026, rewrote Article 4: organizations must take measures to support AI literacy, with no specific level guaranteed.
Deployers of high-risk AI systems must still make sure staff who use and oversee those systems are sufficiently trained, under Article 26 of the EU AI Act.
NIST AI Risk Management Framework (AI RMF) GOVERN 1.6 calls for an AI system inventory, GOVERN 2.1 for documented roles and responsibilities, and GOVERN 2.2 for AI risk training.
NIST AI RMF MAP 1.3 documents the organization's mission and goals for AI, and MAP 1.5 determines and documents organizational risk tolerances.
The Adopt, Defend, Govern (ADG) Foundation phase, months 0 to 3, publishes the AI inventory, names owners, classifies criticality and charters the AI Governance Council, covering Minimum Controls MC-1 to MC-3.
The ADG roadmap then runs Control Deployment (months 3 to 9), Agentic Readiness (9 to 15) and Maturity (15 to 24), reaching continuous assurance in about two years.
Exam tip
The first readiness step is almost always the AI inventory with a named owner per system, because you cannot classify, test or govern what you have not found. UNESCO's Readiness Assessment Methodology assesses countries, not companies, so it is a distractor in enterprise questions.
AI inventory record (ADG MC-1 and MC-2 fields) · json
{
"systemId": "AI-0042",
"name": "Contoso claims triage assistant",
"accountableOwner": "claims-ops-lead@example.com",
"deploymentClass": "Foundation model API",
"riskClassification": "High",
"autonomyTier": "HOTL",
"intendedPurpose": "Rank incoming claims for human review",
"lastGovernanceReview": "2026-10-15"
}