AI glossary
Essential vocabulary for the AI frontier. Precise language signals depth. Master these terms before any strategic interview or technical discovery.
Agile
An iterative approach based on the 2001 Manifesto. Values working software and customer collaboration over rigid plans.
Scrum
The primary Agile framework. Time-boxed Sprints with defined roles (PO, SM) and rituals (Daily, Review, Retro).
Sprint
A fixed-length iteration (typically 2 weeks) delivering a potentially shippable product increment.
Epic
A large business initiative spanning multiple sprints. Tied directly to a strategic objective.
User Story
Small unit of user value: "As a [persona], I want [action], so that [outcome]."
Acceptance Criteria
Testable conditions for story completion, often in Given/When/Then format.
Definition of Done
A team-wide checklist ensuring quality before work is moved to 'Complete'.
Velocity
The average story points a team completes per sprint. A forecasting tool, not a performance metric.
Backlog Refinement
Mid-sprint clarification and estimation session to ensure the backlog is 'sprint-ready'.
Story Points
Relative units of complexity and uncertainty, usually following the Fibonacci sequence.
Sprint Goal
A single objective for the sprint that gives the team focus and flexibility in how they achieve it.
Definition of Ready
A checklist that a story must meet before it can enter a sprint — clear AC, no open dependencies, estimable.
LLM (Large Language Model)
Models trained on massive datasets to predict and generate text. The core of Generative AI.
RAG (Retrieval-Augmented Generation)
Connecting LLMs to real-time facts, grounding responses in verified, retrieved data.
Prompt Engineering
Designing inputs to reliably produce desired outputs (Few-shot, CoT, System prompts).
Fine-tuning
Adapting foundation models for specific domains using labelled examples for target-style adherence.
Embeddings
Numerical vectors that capture semantic meaning. The silent backbone of search and RAG.
Hallucination
Confident fabrication of incorrect information. A fundamental risk requiring product mitigation.
Precision
Accuracy of positive predictions. High precision = low false positive rate (False Alarms).
Recall
Completeness of positive findings. High recall = low false negative rate (Misses).
F1 Score
The harmonic mean of Precision and Recall. The standard for balanced error evaluation.
Model Drift
Performance decay over time as real-world data patterns diverge from training data.
Agentic AI
Autonomous systems that reason and use tools to pursue multi-step goals without human direction.
MCP (Model Context Protocol)
An open standard for connecting models to tools via a standardised, composable interface.
Confidence Threshold
The probability floor for auto-action. Below this, logic routes to human review.
Vector Database
Infrastructure optimised for storing and searching high-dimensional embeddings (e.g. Pinecone).
RLHF
Reinforcement Learning from Human Feedback. Aligning models with human values via ranking.
Shadow Mode
Running models in production to log and validate results without showing them to users.
Shadow Deployment
Canary pattern for validating model shifts on 1–5% of traffic before full rollout.
Ground Truth
The verified correct answer used to evaluate model accuracy. A critical product artifact.
Data Drift
A shift in the statistical distribution of input data that degrades model performance over time.
Inference Latency
The time from input to model output. A non-functional requirement that must be specified before deployment.
Reasoning Model
A class of LLMs (o3, Claude with extended thinking) that generate internal reasoning chains before responding. Use for multi-step planning or complex evaluation at 5–15× the cost of a fast model — justify with cost-per-outcome, not cost-per-token.
Eval Suite
A structured set of test cases, golden-answer comparisons, and automated checks measuring LLM or agent quality across defined dimensions. The PM owns defining what it measures — not just accuracy, but the failure modes that block launch.
Model Governance
The policies, processes, and ownership structures controlling how AI models are selected, updated, evaluated, and retired in production. Governance gaps allow silent model degradation and unchecked cost escalation — the PM is accountable for both.
Prompt Registry
A version-controlled store of production prompts tracking changes, owners, test results, and deployment history. Without a registry, prompt edits are undocumented changes to a production system — apply the same review discipline as code.
Product Roadmap
A strategic communication tool (Now/Next/Later) showing approximately what and why — not a Gantt chart.
OKR
Objectives and Key Results. Separates inspirational goals from measurable outcomes.
RICE Score
Prioritisation: (Reach × Impact × Confidence) ÷ Effort. Relative ROI ranking tool.
Kano Model
Categorising features: Basic Needs, Performance Needs, and unexpected Delighters.
MoSCoW
Scoping framework: Must have, Should have, Could have, and Won't have (this cycle).
Discovery
The work of understanding problems and validating assumptions before building solutions.
North Star Metric
The single metric that best captures the core value delivered to users.
MVP (Minimum Viable Product)
The smallest version of a feature that generates maximum learning from real users.
JTBD (Jobs To Be Done)
A framework for understanding user motivation: what outcome are they hiring this product to achieve?
PRD (Product Requirements Document)
The PM's primary artifact — defines what to build and why, not how to build it.
Model Spec
AI-specific artifact co-authored with data science — defines model inputs, outputs, quality thresholds, and fallback behavior.
SOP (Standard Operating Procedure)
Step-by-step instructions for human reviewers in AI products — essential for the HITL loop to function consistently.
EU AI Act
The world's first comprehensive AI regulation (2024). Classifies AI by risk tiers from Unacceptable to Minimal.
High-Risk AI System
Strictly regulated AI used in critical areas like healthcare, recruitment, or credit scoring.
GDPR Implications
Data minimisation and the Right to Explanation for automated decisions about individuals.
Right to Explanation
The legal requirement to explain why an AI system made a specific automated decision about a person.
Model Card
A standardised artifact documenting model intent, performance, and known failure modes.
SR 11-7
US Fed guidance for model risk management, requiring validation and ongoing monitoring in financial services.
Privacy by Design
Architecting data protection into the system from Day 1, not as a compliance afterthought.
Algorithmic Accountability
The principle that organisations are liable for AI harm, bias, and errors — not just the model.
Bias Audit
Systematic evaluation across demographic groups to detect disparate impact or discrimination.
Guardrail Metrics
Threshold metrics that block release or trigger a pause regardless of primary metric performance.
MLOps
The practice of operationalising ML models — CI/CD for models, monitoring, retraining pipelines, and governance.
NIST AI RMF
The US National Institute of Standards and Technology AI Risk Management Framework — a voluntary governance standard.