Appendix D: Glossary
This glossary defines the key terms used throughout this book in the specific sense that the book uses them. Many of these terms have multiple definitions in common usage; the definitions here reflect the practitioner context of AI transformation in B2B SaaS.
AI council The governance body responsible for strategic oversight of an AI transformation program. Typically composed of the CEO, CTO, VP Product, COO, and Head of CS or equivalent. Distinct from the AI working group, which handles operational decisions. The AI council sets program direction, approves major investments, and reviews program health at quarterly intervals.
AI-augmented Describes an organization or process in which AI assists humans with tasks they were already doing, without redesigning the fundamental shape of the work. Contrast with AI-native. A CS rep who uses AI to draft responses faster is working in an AI-augmented process; a CS operation redesigned around AI triage with human escalation for complex cases is operating AI-natively.
AI enablement point A specific location in an existing workflow where an AI system produces an output that a human acts on. The work order classification assistant that suggests a category for an incoming work order before the dispatcher reviews it represents one AI enablement point. Distinct from an AI use case, which describes the problem being solved; the enablement point describes the specific human-AI interaction moment within the use case’s implementation.
AI function The organizational structure responsible for building, operating, improving, and governing AI systems. In a mature year-two program, the AI function has three components: AI systems engineering (building and operating), AI data and evaluation (the data flywheel and continuous improvement), and AI product management (portfolio prioritization and business outcome tracking).
AI-native Describes an organization or process in which AI constitutes new workflows that would not be practical without it, rather than accelerating existing workflows. Contrast with AI-augmented. An organization is AI-native in a specific process when removing the AI system would require the process to be fundamentally redesigned rather than just slowed down.
AI working group The operational governance body that meets weekly or biweekly to manage the day-to-day decisions of an AI transformation program. Composed of the technical lead, product manager, and representatives from the user teams affected by AI systems. Distinct from the AI council, which handles strategic oversight.
Calibration The process by which individual users develop accurate intuitions about when to follow an AI system’s suggestions and when to override them. Accurate calibration means the user follows the AI when it is right and overrides it when it is wrong. Miscalibrated trust means either over-trusting (following the AI even when it is wrong) or under-trusting (overriding the AI even when it is right).
Calibration protocol The structured process for helping the initial user cohort develop accurate calibration. Consists of three phases: structured exposure (guided first use under observation), calibration metrics review (reviewing the user’s override rate and its accuracy against outcome tracking), and independence (the user is operating without active coaching but with regular metric review). Described in detail in Chapter 11.
Chunking The process of dividing large documents into smaller segments for embedding and storage in the vector database. Chunking strategy affects retrieval quality: chunks that are too small lose context; chunks that are too large dilute the semantic signal. Optimal chunk size depends on the document structure and the query patterns of the use case.
Context window The maximum amount of text (measured in tokens) that a language model can process in a single API call, including both the prompt and the generated output. Context window size determines how much source material can be included in a single generation request, which affects retrieval strategy design.
Continuous evaluation discipline The organizational practice of systematically monitoring production AI systems, detecting quality drift, running improvement experiments, and deploying changes with confidence, as a continuous operational function rather than a one-time gate. Described as a key driver of compound competitive advantage in Chapter 14.
Data flywheel The virtuous cycle in which AI usage generates data that improves the AI systems, which generates more usage, which generates more data. The data flywheel has three required components: usage instrumentation (capturing AI interactions at sufficient granularity), structured feedback processing (converting interaction data into labeled examples and evaluation benchmarks), and improvement cycles calibrated to the data’s accumulation rate.
Embedding A vector representation of a piece of text that captures its semantic meaning in a way that allows semantic similarity to be computed mathematically. Embeddings are generated by embedding models and stored in vector databases to enable semantic search over document corpora.
Evaluation framework The set of benchmarks, metrics, and processes used to assess the quality of an AI system’s outputs. A well-designed evaluation framework defines what good looks like for the specific use case, provides reproducible quality measurements that can be tracked over time, and detects quality changes before they accumulate to a level that affects users.
Fallback behavior The behavior of an AI system when it has low confidence in its output or when a specific condition is not met. Designing fallback behavior is a critical product decision: the system should not surface uncertain outputs as if they were confident ones. Common fallback behaviors include showing a lower-confidence indicator, routing to human review, or declining to generate output.
Generation The process by which a language model produces text output given a prompt. In the context of retrieval-augmented generation (RAG), generation is the step that follows retrieval: the model receives the retrieved context and the user’s query and generates a response that synthesizes both.
Governance The structures, processes, and authority distributions that determine how an AI program makes decisions, resolves conflicts, and maintains accountability for outcomes. Described in terms of the AI council (strategic), the AI working group (operational), and the decision-making protocol that defines which decisions require which level of review.
Labeled examples Data records that pair an input with a human-verified correct output, used to evaluate AI system quality and, in some cases, to fine-tune models. In the context of the data flywheel, labeled examples are produced from production interaction data: an interaction in which a user edited an AI draft before using it can become a labeled example with the original draft as the AI output and the edited version as the human-verified correct output.
Loaded cost The total cost of an employee to the organization, including salary, benefits, employer taxes, and overhead. Typically 1.3 to 1.6 times base salary depending on location, benefits structure, and overhead allocation. Used in budget templates to represent the true cost of time allocation rather than salary alone.
Override rate The percentage of AI system suggestions that users override rather than accept. The primary metric for calibration accuracy: an override rate that is significantly higher than the system’s known error rate indicates under-trust; an override rate that is significantly lower than the system’s known error rate indicates over-trust. The target override rate is the system’s actual error rate, not zero.
Prompt engineering The practice of designing the instructions, context, and examples provided to a language model to produce outputs of the desired quality for a specific task. Prompt engineering is iterative and empirical: the right prompt for a specific task is discovered through testing against representative inputs, not derived from first principles.
Quality drift Degradation in an AI system’s output quality over time, caused by changes in the input distribution, changes in the underlying model (if using a hosted model that is updated by the provider), or deterioration in the retrieval corpus quality. Quality drift is typically gradual and is not visible without consistent evaluation against stable benchmarks.
Retrieval-augmented generation (RAG) An AI system architecture in which the language model’s generation is augmented by content retrieved from a document corpus. The retrieval step uses semantic search over the vector database to find the most relevant documents for the user’s query; the generation step uses those documents as context for producing the response. RAG is the primary architecture for knowledge-based AI use cases in B2B SaaS, as it allows the model to produce responses grounded in the organization’s specific data without fine-tuning.
Semantic search Search that retrieves documents based on meaning rather than keyword matching. Semantic search uses the cosine similarity between the query’s embedding and stored document embeddings to find conceptually related content even when the exact words do not match. Relevant for B2B SaaS use cases where users query knowledge bases using their own phrasing rather than the exact terminology of the underlying documents.
Sprint In the context of this book, the first 90-day implementation phase of an AI transformation program. Distinct from the Agile sprint concept: the AI sprint is a bounded program phase with a defined scope, timeline, and success criteria, not a two-week development cycle.
System prompt The instructions provided to a language model that define its role, behavior, and constraints for a specific use case. The system prompt is set by the developer and is not visible to the end user. It is distinct from the user prompt, which is generated from the user’s input at runtime.
Token The unit of text that language models process. A token is roughly three to four characters or about three-quarters of a word in English. Token count determines API costs (charged per token) and context window usage. Token cost estimation for a specific use case requires counting the average tokens in the system prompt, the retrieved context, the user query, and the generated response.
Token budget The planned allocation of API spending for a specific AI system or the full AI program, typically expressed as a monthly dollar amount. Token budget management involves monitoring actual spend against the plan, understanding the drivers of cost variation, and making architectural decisions (prompt length, context window usage, model tier) that keep costs within the plan.
Use case A specific business problem or workflow that an AI system is designed to address. In the prioritization framework of Chapter 3, a use case is evaluated against three criteria: business impact (how much the problem costs in its current state), technical feasibility (whether the organization has the data and technical capability to build an AI system that addresses it), and organizational readiness (whether the user cohort and supporting processes are ready to adopt the system).
Vector database A database optimized for storing and querying vector embeddings, enabling efficient semantic search over large corpora. In the AI stack described in this book, the vector database is the retrieval infrastructure layer: it stores the embedded representations of the organization’s knowledge base and responds to query embeddings with the most semantically similar documents.