Short definitions of the terms this course uses. The lesson in brackets is where each is explained.
| Term | Meaning |
|---|---|
| A2A | Agent2Agent protocol: how one agent discovers and delegates to another across a boundary. (Workflows and Agents) |
| Abstention | The system declining to answer and handing off, rather than guessing. (Requirements and Success Metrics) |
| ADR | Architecture decision record: one page with the context, the decision, the options rejected and the consequences. (The Solution Design Document) |
| Adoption | The share of intended users who actually use the system, measured after the novelty period. (PoC to Production) |
| Agent | A model in a loop that chooses its own next action and tools until a task is done. (Workflows and Agents) |
| Agent registry | The catalogue of every agent with its owner, purpose, tools and risk tier. (Agents at Enterprise Scale) |
| Agent washing | Relabelling chatbots or scripts as agents. (Workflows and Agents) |
| AI gateway | The single internal entry point to models: keys, limits, routing, fallback, cost tags. (Choosing a Model) |
| AI literacy | The EU AI Act duty to ensure staff using AI understand it sufficiently. (Governance and Regulation) |
| Autonomy dial | Five levels from assisting a person to acting across systems unsupervised. (Workflows and Agents) |
| Baseline | How the task performs today, in real units, before any change. (Discovery) |
| Batch processing | Submitting requests for results within hours at about half price. (Cost, TCO and ROI) |
| C4 model | Diagrams at four zoom levels: context, containers, components, code. (The Solution Design Document) |
| Canary | Sending a small share of traffic to a new version and comparing. (PoC to Production) |
| Centre of excellence | The central team that runs the platform, sets standards and reviews high-risk systems. (The AI Platform and Operating Model) |
| Chargeback | Billing each team’s budget for its AI use. (The AI Platform and Operating Model) |
| Chunk | A retrievable passage of a document. (Retrieval and Knowledge) |
| Context engineering | Deciding what goes into the context window on each call. (How AI Works, for Architects) |
| Context window | Everything the model can see on one call; a fixed budget in tokens. (How AI Works, for Architects) |
| Cutoff | The date after which a model has no training data. (How AI Works, for Architects) |
| Data residency | Where data is stored and processed. (Security, Privacy and Residency) |
| Data sovereignty | Whose law can compel access to data. (Security, Privacy and Residency) |
| Degraded mode | What the system does when no model route is available. (Capacity, Quotas and Resilience) |
| Deployer | Under the EU AI Act, an organisation using an AI system in its own operations. (Governance and Regulation) |
| Deterministic control | A control enforced by code or configuration that a model cannot be talked past. (Security, Privacy and Residency) |
| Discovery | Finding the real problem, its numbers, data and people before proposing anything. (Discovery) |
| Distillation | Fine-tuning a small model to do a narrow task as well as a large one. (The Pattern Ladder) |
| Embedding | A vector representing the meaning of text, used for search. (How AI Works, for Architects) |
| Evaluation (eval) | A set of cases with known good outcomes plus a way to score against them. (Evaluation as the Contract) |
| Exit plan | One page on how you would leave a strategic dependency, written at adoption. (Portability and the Model Lifecycle) |
| Fallback | A second model or route, already evaluated, used when the first fails. (Capacity, Quotas and Resilience) |
| Fine-tuning | Continuing a model’s training on your examples to change how it behaves. (The Pattern Ladder) |
| Foundation model | A large general-purpose model that applications are built on. (How AI Works, for Architects) |
| Gate | A measurable condition for moving from one delivery stage to the next. (PoC to Production) |
| Groundedness | Whether every statement in an answer is supported by a cited source. (Evaluation as the Contract) |
| Guardrail | A check on model input or output, usually itself a model or classifier. (Security, Privacy and Residency) |
| Hallucination | A fluent, confident statement that is not true. (How AI Works, for Architects) |
| Harness | The code around a model in an agent: tools, policy, budgets, state. (Workflows and Agents) |
| High-risk system | An EU AI Act class with extensive obligations, defined by use. (Governance and Regulation) |
| Hub and spoke | Central platform and standards, with builders in business teams. (The AI Platform and Operating Model) |
| Hybrid search | Keyword and vector search combined. (Retrieval and Knowledge) |
| Inference | Using a trained model to produce output. (How AI Works, for Architects) |
| Inventory | The register of every AI system with owner, purpose, data and risk class. (Governance and Regulation) |
| Landing zone | The pre-approved cloud accounts, networks and policies teams must deploy into. (The Cloud and Software Ground Floor) |
| LLM | Large language model. (How AI Works, for Architects) |
| MCP | Model Context Protocol: the standard way to connect agents to tools and data. (Workflows and Agents) |
| Model as judge | Using a strong model with a rubric to grade outputs. (Evaluation as the Contract) |
| Model retirement | The date a provider stops serving a model version. (Portability and the Model Lifecycle) |
| Non-goal | Something the system will deliberately not do. (Requirements and Success Metrics) |
| On-demand | Paying per token on shared capacity. (The Cloud AI Platforms) |
| Open-weight model | A model whose weights can be downloaded and run yourself. (Choosing a Model) |
| Operating model | Who runs the platform, who builds, who sets rules, who pays. (The AI Platform and Operating Model) |
| Paved road | A shared route so easy that teams choose it, with controls built in. (The AI Platform and Operating Model) |
| Payback | Months until cumulative net benefit equals the one-off cost. (Cost, TCO and ROI) |
| Pilot | A fixed-length trial with real users and data, run as an experiment. (PoC to Production) |
| PoC | Proof of concept: evidence that something can work once. (PoC to Production) |
| Prompt caching | Paying a fraction of the input price to reread an unchanged prompt prefix. (Cost, TCO and ROI) |
| Prompt injection | Text in the model’s input acting as an instruction. (Security, Privacy and Residency) |
| Provider | Under the EU AI Act, whoever develops an AI system and places it on the market. (Governance and Regulation) |
| Provisioned throughput | Reserved model capacity for a fixed fee. (Capacity, Quotas and Resilience) |
| Quota | The requests and tokens per minute a provider allows an account. (Capacity, Quotas and Resilience) |
| RAG | Retrieval-augmented generation: fetch relevant facts, then answer from them. (Retrieval and Knowledge) |
| Realisation factor | The share of saved time that becomes countable value. (Cost, TCO and ROI) |
| Recall | Whether the passage containing the answer was among those retrieved. (Retrieval and Knowledge) |
| Reranking | A second model reordering search results so the best few are kept. (Retrieval and Knowledge) |
| Risk tier | An internal class that sets how much review and control a system needs. (Governance and Regulation) |
| ROI | Return on investment: benefit minus cost, divided by cost. (Cost, TCO and ROI) |
| Routing | Sending each request to the cheapest model that can handle it. (Choosing a Model) |
| Shadow AI | Unapproved AI tools used on company data. (The AI Platform and Operating Model) |
| Shadow mode | Running a new version on real traffic without showing its output. (PoC to Production) |
| Shared responsibility | The split of duties between model maker, host and application builder. (Well-Architected AI) |
| Showback | Showing each team its AI spend without charging it. (The AI Platform and Operating Model) |
| Slice | A tagged subset of an evaluation set, gated separately. (Evaluation as the Contract) |
| Structured output | A provider feature that forces replies to match a schema. (The Pattern Ladder) |
| TCO | Total cost of ownership over a period, including people. (Cost, TCO and ROI) |
| Token | The unit a model reads and writes; about four characters of English. (How AI Works, for Architects) |
| Tool | A function a model may ask your code to run. (Workflows and Agents) |
| Unit cost | What one task costs to run. (Cost, TCO and ROI) |
| Well-architected review | A structured examination of a design through fixed pillars. (Well-Architected AI) |
| Workflow | Several model calls arranged in an order fixed by your code. (The Pattern Ladder) |