AI Transformation Glossary

94 terms used across the AI Transformation Playbook — frameworks, acronyms, and concepts explained in plain English.

A

ADKAR

A change management framework with five stages: Awareness (why the change is happening), Desire (willingness to participate), Knowledge (how to change), Ability (putting knowledge into practice), and Reinforcement (sustaining the change). Used in the Change Journey Template to plan adoption activities.

See also: , Section 11

Adoption

People actually using AI tools as part of their daily work, as opposed to deployment (making tools technically available). The gap between deployment and adoption is where most AI programmes fail. Measured through usage data, behaviour change, and business outcomes.

See also: , Section 11

AI Agent

An autonomous AI system that uses tools, makes decisions, and operates in loops to complete multi-step tasks without continuous human direction. More capable than a simple prompt-response interaction but requires stronger governance. Represents the next evolution beyond copilots.

See also: , Section 05

AI Literacy

A basic understanding of what AI can and cannot do, how it works at a conceptual level, and what its limitations are. Required by the EU AI Act for organisations deploying AI. Distinct from technical AI skills — every employee needs literacy, not everyone needs to build models.

See also: Section 09

AI Readiness

An organisation's preparedness to adopt and benefit from AI, assessed across six dimensions: data foundations, digital maturity, people and skills, leadership and strategy, process and operations, and governance and ethics.

Algorithm Maturity Level

A five-stage model describing an organisation's analytical sophistication: Descriptive (what happened), Diagnostic (why it happened), Predictive (what will happen), Prescriptive (what should we do), and Autonomous (the system acts on its own). Most consumer businesses are between Descriptive and Diagnostic.

API (Application Programming Interface)

A method for accessing AI models programmatically, allowing them to be integrated into business systems and workflows. Using AI via API gives more control over data, costs, and customisation than using a web interface, but requires technical capability.

See also: Section 05

Assumption Map

A visual tool for plotting assumptions about an AI initiative by how important they are (if wrong, does it kill the idea?) and how much evidence supports them. Assumptions in the high-importance, low-evidence quadrant are the ones to test first.

See also: Section 07

Augmentation

Using AI to make a person better at their existing task, as opposed to automation which replaces the task entirely. The distinction matters for change management — augmentation changes how someone works, automation changes whether they do the work at all.

See also: , Section 11

Automation

Using AI to fully replace a human task. Distinct from augmentation, where AI assists but the human retains the task. Most early AI deployments in consumer businesses are augmentation, not automation.

See also: Section 11

Automation Bias

The tendency for humans to over-trust AI outputs without adequate scrutiny, especially when the AI is presented as authoritative. A key risk in human-in-the-loop designs — if the human always clicks 'approve', the oversight is meaningless.

See also: , Section 09
B

Baseline

The current-state measurement taken before AI is introduced, against which improvement is measured. Without a baseline, you cannot prove AI made things better. Common baselines: processing time, error rate, cost per transaction, customer satisfaction score.

Bias

Systematic unfairness in AI outputs that disadvantages particular groups. Can originate from training data, model design, or how outputs are used. Requires ongoing monitoring, not a one-time check — bias can emerge as data and context change.

See also: , Section 09

Business Case

The financial and strategic justification for investing in AI capability. Should include cost of inaction, total cost of ownership, confidence-banded benefits, and a clear narrative connecting AI investment to business strategy. Not just a spreadsheet — it's a story backed by evidence.

See also: , , Section 13

Business Model Canvas

A strategic tool for mapping how a business creates, delivers, and captures value. In the playbook context, used to identify where AI changes the business model itself (transformation) versus where it improves existing operations (optimisation).

See also: Section 02

Business Readiness

The state of preparedness across people, process, technology, and governance required before an AI capability can transition from experiment to production. Assessed via the Business Readiness Checklist (11g).

C

Centre of Excellence (CoE)

An organisational model where a small central team (typically 2–3 people) provides AI expertise, standards, and governance while business functions own their own AI-enabled outcomes. The recommended starting model for most consumer businesses.

Centaur Model

A human-AI collaboration pattern where there is a clear division of responsibilities — the human handles certain tasks and the AI handles others, with distinct handoff points. Named after the chess concept. Contrast with the cyborg model where roles are blended.

See also: , Section 04

Change Capacity

An organisation's ability to absorb concurrent changes without overwhelming people. AI programmes compete for change capacity with every other initiative. Overloading change capacity is one of the most common reasons AI adoption stalls.

See also: , Section 11

Change Impact Assessment

A structured analysis of how AI changes roles, processes, skills, and ways of working for affected teams. Maps current state to future state and identifies the transition requirements.

Change Management

The structured approach to transitioning people from current ways of working to new AI-enabled ways of working. Covers communication, training, support, and reinforcement. The most underestimated workstream in AI programmes — technology is the easy part.

See also: , Section 11

Chunking

Breaking documents into smaller segments (chunks) for storage in a vector database. Chunk size and overlap affect retrieval quality — too small loses context, too large dilutes relevance. A critical design decision in RAG systems.

See also: , , Section 10

Competitive AI Exposure Assessment

A structured analysis of how competitors and market entrants are using AI, and the resulting risk to your business if you don't act. Provides evidence for urgency (or confirms patience is warranted) rather than relying on FOMO.

See also: Section 02

Confidence Banding

Presenting business case benefits as conservative, expected, and optimistic scenarios with explicit assumptions for each, rather than a single point estimate. Builds board credibility by acknowledging uncertainty.

Context Canvas

A structured tool for mapping the external forces shaping the AI opportunity: competition, customer needs, technology trends, regulation, workforce trends, and economic environment.

Context Window

The amount of text an LLM can process in a single interaction, measured in tokens. Larger context windows allow more information to be included but increase cost and can reduce accuracy. A practical constraint when designing RAG and agent systems.

See also: , Section 10

Copilot

An AI assistant embedded within existing software (e.g., Microsoft 365 Copilot, GitHub Copilot). Augments the user's workflow without requiring them to switch tools. Lower adoption friction than standalone AI tools but limited to the host application's capabilities.

See also: , Section 05

Cost of Delay

The cumulative competitive, operational, and financial cost of not acting on AI over a defined period. Used in the Competitive AI Exposure Assessment to make the case for urgency — or to confirm that urgency is not warranted.

Cost of Inaction

The business impact of choosing not to invest in a specific AI capability. Included in business cases to provide balance — the alternative to investment is not zero cost, it is the cost of the status quo.

Cutover

The planned transition point where an AI capability moves from parallel running (old and new operating together) to full production. Requires defined rollback criteria and decision authority.

Cyborg Model

A human-AI collaboration pattern where the human and AI contribute to the same task in a blended way — the boundaries between human and AI work are fluid. Contrast with the centaur model where responsibilities are clearly divided.

See also: , Section 04
D

Data Classification

Categorising data by sensitivity level: public, internal, confidential, and restricted. Determines which data can be used with which AI tools — public cloud LLMs may be fine for internal data but not for restricted data containing PII.

See also: , Section 08

Data Governance

The framework of policies, roles, and processes that ensures data is managed consistently, securely, and in compliance with regulations. Covers access controls, quality standards, retention policies, and accountability. A prerequisite for trustworthy AI.

See also: , Section 08

Data Quality

The fitness of data for use in AI systems, assessed across accuracy, completeness, currency, consistency, and structure. Poor data quality is the most common blocker to AI adoption — garbage in, garbage out applies doubly to AI.

See also: , Section 03

Deployment

Making an AI tool technically available to users. Distinct from adoption — a tool can be deployed (installed, configured, accessible) without anyone actually using it. The gap between deployment and adoption is where change management matters.

See also: Section 11
E

Embeddings

Mathematical representations of text (or other data) as vectors of numbers, capturing semantic meaning. Documents with similar meaning have similar embeddings, enabling semantic search. The foundation technology behind RAG systems.

See also: , , Section 10

EU AI Act

European regulation that classifies AI systems by risk level (unacceptable, high, limited, minimal) and sets obligations proportionate to that risk. Applies to any organisation deploying AI that affects EU citizens, regardless of where the organisation is based. Phased implementation from 2024–2027.

See also: , Section 09

Experiment Canvas

A one-page framework for designing a structured AI experiment with a clear hypothesis, success criteria, timeline, and resource requirements. Ensures experiments produce evidence, not just activity.

Experiment Pipeline

The managed portfolio of AI experiments at various stages: proposed, active, completed, scaled, or stopped. Tracked via the Experiment Pipeline Tracker (07b) to prevent uncoordinated experimentation.

Explainability

The ability to explain how an AI system reached a particular output or decision in terms that the affected person can understand. Required by the EU AI Act for high-risk systems. Different from transparency — explainability is about the reasoning, transparency is about disclosure.

See also: , Section 09
F

Facilitation Guide

A structured guide for the facilitator running a playbook workshop. Includes session format, facilitation principles, recommended tool order, probing questions, and patterns to watch for. Every section with tools has a corresponding facilitation guide.

Fairness

Treating different groups equitably in AI outputs and decisions. Not the same as treating everyone identically — fairness may require different treatment to achieve equitable outcomes. Requires defining what fair means in your specific context.

See also: , Section 09

Fine-Tuning

Training an existing AI model on domain-specific data to improve its performance on particular tasks. More expensive and complex than prompt engineering or RAG, and usually unnecessary for most consumer business use cases. Consider it only after simpler approaches have been exhausted.

See also: , Section 05
G

Go/No-Go Gate

A formal decision point where the leadership team or steering group decides whether an AI capability proceeds to the next stage (experiment → pilot → production). Based on evidence from the Business Readiness Checklist.

See also: , Section 12

Golden Circle Canvas

A strategic framing tool based on WHY / HOW / WHAT that connects AI initiatives to organisational purpose. Ensures AI programmes can articulate why they matter beyond efficiency gains.

Governance Framework

The structure of roles, decision rights, escalation paths, and review cadences that ensures AI capabilities are deployed and operated responsibly. Proportionate to risk — not every AI use case needs the same level of governance.

Governance Tier

A risk classification for AI experiments and capabilities. Tier 1 (low risk): internal productivity tools needing light governance. Tier 2 (medium): customer-adjacent or data-sensitive uses. Tier 3 (high): customer-facing or consequential decision-making requiring full governance review.

See also: , Section 08
H

Hallucination

When an AI model generates plausible but factually incorrect output with full confidence. A fundamental characteristic of LLMs, not a bug to be fixed. Managed through human oversight, not through better prompting alone.

High-Risk AI

AI systems that affect people's fundamental rights or make consequential decisions (employment, credit, healthcare, education). Subject to the strictest requirements under the EU AI Act including conformity assessments and mandatory human oversight.

See also: , Section 09

Human-in-the-Loop (HITL)

A design pattern where AI generates output but a human reviews, approves, or modifies it before it reaches the end user or takes effect. The default pattern for most early AI deployments in consumer businesses.

See also: , Section 08

Hypercare

An intensive support period immediately after an AI capability goes live, with elevated monitoring, faster response times, and dedicated support resources. Typically 2–4 weeks depending on complexity.

I

Innovation Matrix

A visual tool for plotting AI initiatives by type of impact (efficiency, experience, capability, model) and scale of change required (optimise, extend, transform). Helps leadership see the portfolio balance and avoid clustering all bets in one quadrant.

K

Kill Criteria

Pre-defined conditions that trigger stopping an AI experiment before its planned end date. Agreed upfront so the decision to stop is based on evidence, not politics. Examples: data quality below threshold, cost exceeding budget by 50%, safety concern identified.

See also: Section 07

KPI (Key Performance Indicator)

A specific, measurable metric that indicates whether an AI capability is achieving its intended business outcome. Good AI KPIs measure business impact (cost saved, revenue generated, time freed), not just technical performance (model accuracy).

L

LLM (Large Language Model)

The category of AI that processes and generates text by predicting patterns in language. Includes models like GPT, Claude, and Gemini. The primary AI technology covered by this playbook because it offers the most accessible and immediate value for consumer businesses.

M

Maturity Stage

One of four stages in the AI Maturity Self-Assessment: Exploring (early awareness), Establishing (first experiments), Scaling (moving to production), and Optimising (continuous improvement of embedded AI capabilities).

Meaningful Human Oversight

Genuine human review of AI outputs where the reviewer has the knowledge, authority, and time to intervene — not rubber-stamping. Required by the EU AI Act for high-risk systems. Automation bias is the primary threat to meaningful oversight.

See also: , Section 09

Model Drift

Degradation of AI performance over time as the real world changes but the model's training data stays static. Requires ongoing monitoring and periodic retraining or re-evaluation. A reason why AI capabilities need operational ownership, not just project delivery.

N

No-Code / Low-Code Platform

Visual tools for building AI workflows without writing code (e.g., Make, Zapier, n8n). Lower the barrier to experimentation but can create governance challenges if workflows proliferate without oversight. Good for experiments, need careful scaling.

See also: Section 05
O

Operating Model

How the business actually operates: who does what, how decisions are made, what processes are followed, what technology supports them. Distinct from the business model (what the business does). AI changes the operating model even when the business model stays the same.

See also: Section 14

Optimisation vs Transformation

The strategic spectrum for AI adoption. Optimisation uses AI to do existing things faster, cheaper, or more consistently. Transformation uses AI to fundamentally change what the business does or how it operates. Most businesses start with optimisation and evolve.

Orchestration Framework

Software that manages the flow of data between AI models, vector databases, and business systems (e.g., LangChain, LlamaIndex). Handles the plumbing that connects RAG retrieval, prompt construction, and response generation.

See also: Section 10

Outcome Canvas

A structured template for documenting the results of an AI experiment: what was tested, what was learned, whether the hypothesis was validated, and the recommendation (scale, iterate, or stop). Ensures experiment outcomes are captured as organisational knowledge.

See also: Section 07
P

Parallel Running

Operating the old process and the new AI-enabled process simultaneously for a defined period to build confidence, identify issues, and validate output quality before full cutover.

See also: Section 11

Phase Gate

A structured decision point between phases of the AI journey (e.g., experiment to pilot, pilot to production). Requires evidence against defined criteria before proceeding.

See also: Section 12

PII (Personally Identifiable Information)

Any data that can identify an individual: names, email addresses, phone numbers, purchase history, location data. Sending PII to external AI services requires careful consideration of data processing agreements, GDPR compliance, and data residency.

See also: , Section 08

Pilot

A limited-scale deployment of an AI capability with real users and real data, but constrained to a specific team, location, or use case. The middle stage between experiment (proof of concept) and full production rollout.

See also: , Section 12

Pitch Canvas

A one-page narrative format for presenting AI business cases to board or executive audiences. Follows a storytelling structure (context → problem → solution → evidence → ask) rather than a data-heavy slide deck.

Prompt Engineering

The practice of crafting instructions and context to guide AI model behaviour and improve output quality. The simplest and cheapest way to improve AI performance — should always be tried before fine-tuning or custom model development.

See also: , Section 05
R

RAG (Retrieval-Augmented Generation)

A technique that connects an LLM to a business's proprietary data so the model can retrieve relevant information before generating a response. Makes AI answers grounded in your actual data rather than generic training data.

See also: , , Section 10

RACI

A responsibility assignment matrix: Responsible (does the work), Accountable (owns the outcome), Consulted (provides input), Informed (kept updated). Used in programme governance to clarify who does what. The most common mistake is having multiple people Accountable for the same thing.

Readiness Level

A visual maturity indicator showing where an AI capability sits on the path from experiment to full production. Assessed across technical, organisational, change, and commercial dimensions.

Reliability Categories

The three-category framework for AI capability: reliable (works consistently at volume), inconsistent (works sometimes, needs oversight), and impossible (AI cannot do this). The most important calibration tool for leadership teams.

Risk Classification

Categorising an AI capability by its risk profile across dimensions including data sensitivity, customer impact, regulatory exposure, and reputational risk. Determines the appropriate level of governance oversight.

See also: , Section 08

Risk-Reward Matrix

A visual tool for plotting AI opportunities by potential business value against implementation risk. Helps prioritise which opportunities to pursue first — high value, low risk initiatives are the obvious starting point.

Riskiest Assumption

The single assumption in an AI initiative that, if wrong, would most fundamentally undermine the case for proceeding. Identified and tested first to fail fast and cheaply rather than late and expensively.

ROI (Return on Investment)

The financial return generated by an AI capability relative to its cost. For AI, ROI calculations should include the full total cost of ownership, use confidence banding rather than point estimates, and account for the cost of inaction as the baseline.

See also: , Section 13

Rollback

Reverting to the previous process or system if an AI capability fails or produces unacceptable results after go-live. Requires pre-defined triggers, procedures, and decision authority.

See also: , Section 11

Runbook

Operational documentation describing how to run and maintain an AI capability in production: standard procedures, known issues and fixes, escalation paths, and monitoring thresholds. Essential for the handover from project team to operational team.

S

Scalability Assessment

A structured evaluation of whether an AI experiment is ready for production, covering four dimensions: technical scalability, organisational readiness, change impact, and commercial viability.

Sensitivity Analysis

Testing how business case outcomes change when key assumptions are varied. Reveals which assumptions the business case is most dependent on, and therefore which ones need the strongest evidence. Essential for credible AI investment proposals.

See also: , Section 13

Shadow AI

AI tools being used by employees without formal organisational approval or governance. Often a source of valuable use case evidence — the organisation's best AI ideas are frequently already being tested informally.

Stakeholder Mapping

Identifying and categorising the people who influence, are affected by, or need to support an AI initiative. Maps stakeholders by influence and impact to determine the right engagement approach for each group.

Steering Group

The governance body that oversees the AI programme, makes investment decisions at phase gates, resolves cross-functional conflicts, and ensures alignment with business strategy. Typically meets monthly.

SWOT Canvas

An AI-specific strengths, weaknesses, opportunities, and threats analysis. The intersections table (strength + opportunity, weakness + threat) is where strategic priorities emerge.

T

Target Operating Model (TOM)

The future-state design of how the business will operate with AI embedded. Covers value delivery, organisation and people, decision-making, core processes, technology, and data.

See also: Section 14

Task Decomposition

Breaking a role or process into its component tasks and assessing each for AI suitability. AI automates tasks, not jobs — this distinction is critical for both opportunity identification and workforce communication.

Tiered Governance

A proportionate approach to AI governance where the level of oversight matches the level of risk. Low-risk internal tools need less governance than customer-facing AI that makes consequential decisions.

See also: , Section 06

Token

The basic unit of text that an AI model processes — roughly three-quarters of a word in English. Tokens determine API costs (charged per token processed) and constrain context windows. Understanding tokens helps estimate costs and design efficient prompts.

See also: , Section 05

Total Cost of Ownership (TCO)

The full cost of an AI capability at production scale: licensing, compute, data preparation, integration, specialist hiring, change management, ongoing maintenance, and quality assurance. Experiment costs are misleading — TCO at scale is what matters.

Transparency

Disclosing to users, customers, and stakeholders when AI is being used and how it influences decisions or outputs. A legal requirement under the EU AI Act and a trust-building practice. Distinct from explainability, which is about the reasoning.

See also: , Section 09
V

Vector Database

A specialised database for storing and searching embeddings (e.g., Pinecone, Weaviate, Qdrant). The storage layer in RAG architectures that enables fast semantic search across large document collections.

See also: , , Section 10
W

Workflow Automation

Connecting multiple AI and non-AI systems to automate multi-step business processes. Tools like Make, Zapier, and n8n enable this without code. Distinct from single-task AI — workflow automation chains tasks together across systems.

See also: Section 05