Automation
Foundations
Definition: Rule-based execution of tasks without human intervention.
Why it matters: Many “AI wins” are actually automation wins; confusing the two increases cost and risk.
Use when: Deciding if a workflow needs intelligence or just consistency.
Generative AI
Foundations
Definition: Models that create new content (text, images, audio, code).
Why it matters: Multiplies ideation and first-draft speed.
Use when: You need options/variations to edit rather than start from scratch.
Large Language Model (LLM)
Foundations
Definition: A model trained on vast text to predict the next tokens; powers chat/assistants.
Why it matters: Core engine behind drafting, summarizing, and Q&A.
Use when: You need language tasks done quickly (briefs, emails, reports).
Inference
Foundations
Definition: Running a trained model to generate predictions/outputs.
Why it matters: Where latency and per-output cost are felt.
Use when: Estimating throughput and SLAs for workflows.
Deterministic vs Probabilistic Systems
Foundations
Definition: Deterministic systems produce the same output given the same input; probabilistic systems may vary.
Why it matters: Sets expectations for consistency and review needs.
Use when: Deciding where AI is appropriate vs where strict repeatability is required.
Context Window
Foundations
Definition: The maximum text the model can consider at once.
Why it matters: Too little context → omissions; too much → cost.
Use when: Deciding how much source material to include.
Token / Tokenization
Foundations
Definition: Sub-word units models read and bill on.
Why it matters: Drives cost/latency and limits prompt size.
Use when: Estimating run cost and fitting long inputs.
Transformer
Foundations
Definition: The neural architecture most modern LLMs use.
Why it matters: Explains why attention to context matters (and costs scale).
Use when: Framing limits and behavior of LLMs to non-technical peers.
Prompt
Prompting
Definition: The instructions and context you give the model.
Why it matters: Clear prompts produce better, more consistent results.
Use when: You specify role, task, context, constraints, and format.
Prompt Framework (RTF/TAG/RISE)
Prompting
Definition: Simple structures for writing prompts (e.g., Role-Task-Format).
Why it matters: Reduces trial-and-error; improves reproducibility.
Use when: You want teammates to get similar results for the same task.
System Prompt
Prompting
Definition: A persistent instruction that sets the model’s role/behavior.
Why it matters: Standardizes outputs across people and time.
Use when: Building repeatable assistants or SOP-driven workflows.
Temperature
Prompting
Definition: A setting that controls randomness (lower = more deterministic).
Why it matters: Balances creativity vs. consistency.
Use when: Lower for compliance/SOPs; higher for brainstorming.
Chain-of-Thought (CoT)
Prompting
Definition: Prompting the model to reason step-by-step.
Why it matters: Improves problem decomposition and transparency.
Use when: You want structured reasoning or math-like steps.
Hallucination
Reliability & Risk
Definition: Plausible-sounding but false output.
Why it matters: Brand, SEO, and trust risk.
Use when: Designing checks: grounding, citations, and human review.
Grounding / RAG (Retrieval-Augmented Generation)
Reliability & Risk
Definition: Supplying trusted documents/data to the model at answer time.
Why it matters: Reduces hallucination; enables citations.
Use when: Outputs contain facts, numbers, or client specifics.
Guardrails
Reliability & Risk
Definition: Automated checks/policies that block or flag risky outputs.
Why it matters: Enforces standards without manual policing.
Use when: Validating brand/SEO rules or blocking restricted terms.
Human-in-the-Loop (HITL)
Reliability & Risk
Definition: A design where humans review, approve, or intervene in AI outputs.
Why it matters: Maintains accountability while enabling speed and scale.
Use when: Outputs affect clients, brand, compliance, or decisions.
Alignment
Reliability & Risk
Definition: Steering model behavior toward human values and policies.
Why it matters: Prevents unsafe or off-brand outputs.
Use when: Setting constraints, safety filters, and tone rules.
Explainability (XAI)
Reliability & Risk
Definition: Methods to understand why a model produced an output.
Why it matters: Aids trust, debugging, and client communications.
Use when: Reviewing higher-impact decisions or sensitive content.
Embedding
Knowledge & Retrieval
Definition: Numeric representation of text used to measure similarity.
Why it matters: Powers search and retrieval over our own content.
Use when: Building knowledge search or RAG.
Vector Database / Index
Knowledge & Retrieval
Definition: Storage optimized for embeddings and similarity search.
Why it matters: Makes retrieval fast and relevant.
Use when: You want repeatable grounding over many documents.
Workflow Decomposition
Workflow & Experimentation
Definition: Breaking work into discrete steps to identify automation or AI assist points.
Why it matters: Makes AI opportunities visible and scoping realistic.
Use when: Mapping a workflow to find assistable steps.
Impact–Feasibility–Risk (IFR) Lens
Workflow & Experimentation
Definition: A framework for evaluating opportunities across value, practicality, and downside.
Why it matters: Prevents novelty-chasing; supports responsible prioritization.
Use when: Comparing pilot ideas and picking “smallest safe tests.”
AI Pilot / Experiment
Workflow & Experimentation
Definition: A time-boxed, scoped test designed to learn—not to scale.
Why it matters: Prevents premature rollout and reputational risk.
Use when: Testing assumptions before committing resources.
Agent / Agentic System
Architecture
Definition: An agent plans/acts in steps; an agentic system orchestrates multiple tools/agents toward a goal.
Why it matters: Agentic systems have more power, but also greater complexity and risk.
Use when: Start with simple agents; justify multi-agent with a clear need.
Fine-tuning
Architecture
Definition: Training a base model further on task-specific examples.
Why it matters: Boosts consistency for narrow, high-volume work.
Use when: You have many labeled examples of the same task.