Glossary
Short, stable, citeable definitions for AI waste, ROI measurement, and the Harmondale method.
Last updated: 25 June 2026
Artificial intelligence is the field of building machine-based systems that infer from inputs to produce predictions, content, recommendations, or decisions. In business, the useful question is not whether software is called AI, but what objective it serves and what output it changes.
Read definitionAn AI system is a machine-based system that uses inputs to infer how to generate outputs such as predictions, content, recommendations, or decisions. It may operate with different levels of autonomy and may adapt after deployment, which makes ownership and monitoring important.
Read definitionMachine learning is a branch of AI where a system improves or configures behavior from data rather than from only hand-written rules. The model learns statistical patterns during training, then applies those patterns to new inputs during inference.
Read definitionDeep learning is machine learning based on neural networks with many layers. These networks learn representations from data, which is why they are useful for language, images, speech, and recommendation tasks, but also harder to explain and govern than simpler models.
Read definitionAn AI model is the trained mathematical component that maps inputs to outputs. It is not the whole product: the complete system also includes prompts, retrieval, user permissions, logging, human review, integrations, policies, and the workflow where the model is used.
Read definitionA foundation model is a large model trained on broad data so it can be adapted to many downstream tasks. It can support chat, classification, extraction, coding, image generation, or agents, but still needs context, controls, evaluation, and cost discipline.
Read definitionA large language model, or LLM, is a model trained on large collections of text and related data to predict and generate language. It can draft, summarize, classify, translate, or reason over context, but it does not automatically know what is true for your company.
Read definitionGenerative AI is AI that produces new content such as text, images, code, audio, video, or structured drafts from a prompt or other input. Its value depends on whether generated output reduces work, improves quality, or creates new risk and rework.
Read definitionA transformer is a neural-network architecture that uses attention mechanisms to relate parts of an input sequence to each other. Transformers made modern LLMs and many multimodal models practical because they handle long context, parallel training, and flexible sequence tasks well.
Read definitionA token is a unit of text or data that a language model processes, often a word part, word, punctuation mark, or symbol. Tokens determine context limits, pricing, latency, and how much information can be read or generated in one request.
Read definitionA context window is the amount of information a model can consider in a single request, usually measured in tokens. It includes the system instructions, user prompt, retrieved documents, conversation history, and any output the model must generate.
Read definitionA prompt is the instruction, question, data, or context given to an AI model to guide its output. A useful prompt defines the task, constraints, audience, source material, and success criteria instead of only asking for a generic answer.
Read definitionPrompt engineering is the practice of designing, testing, and maintaining instructions that make AI outputs more reliable for a specific task. In mature workflows it includes examples, constraints, source boundaries, evaluation criteria, and version control, not only clever wording.
Read definitionAn embedding is a numeric representation of text, image, audio, or other data that places similar items near each other in a vector space. Embeddings help AI systems search by meaning, cluster documents, deduplicate content, and retrieve context for generation.
Read definitionA vector database stores embeddings and retrieves the nearest matches to a query vector. It is often used for semantic search and retrieval-augmented generation, where the model needs relevant company documents before answering.
Read definitionRetrieval-augmented generation, or RAG, is a pattern where a system retrieves relevant source material before asking a model to generate an answer. The model is still generative, but the answer is grounded in selected documents, records, or knowledge-base passages.
Read definitionFine-tuning is additional training that adapts an existing model to a narrower task, style, domain, or output format. It changes model behavior more permanently than a prompt, but it is not a substitute for clean data, retrieval, evaluation, or governance.
Read definitionInference is the moment an AI model applies what it has learned to new input and produces an output. In production, inference is where latency, cost, safety checks, retrieval, logging, and user experience come together.
Read definitionTraining data is the data used to fit the learnable parameters of an AI system. It shapes what patterns the model can learn, which means gaps, bias, duplication, outdated examples, or unclear rights can become model behavior.
Read definitionValidation data is data used to evaluate a trained model during development and tune non-learned choices such as thresholds, prompts, hyperparameters, or stopping decisions. It helps detect overfitting before the final independent test.
Read definitionTesting data is data used for an independent evaluation of an AI system before launch or after major change. It should represent real conditions without being used to train or tune the model, otherwise the evaluation becomes inflated.
Read definitionAn AI hallucination is an output that sounds plausible but is false, unsupported, or disconnected from the provided sources. Hallucination is not only a model flaw; it can come from weak prompts, missing context, poor retrieval, or no human review.
Read definitionAn AI agent is a system that uses a model to pursue a goal through steps, tool calls, memory, or actions in other software. The risky part is not the word agent, but what permissions, data, approvals, and rollback paths it receives.
Read definitionAn AI workflow is a repeatable sequence where AI changes how work is produced, checked, approved, or delivered. It includes the task, input data, model or tool, human review, handoff, KPI, and decision rule for keeping or stopping it.
Read definitionModel evaluation is the process of measuring whether an AI model or system performs well enough for a specific use. It can test accuracy, reliability, bias, robustness, latency, cost, safety, and business impact before and after deployment.
Read definitionAI ROI is the measurable return produced by a specific AI use case after subtracting the full cost of tools, human review, integration, training, governance, and risk. It is not usage volume or self-reported time saved; it is net business value that survives operational measurement.
Read definitionAn AI ROI audit maps every meaningful AI use case, subscription, workflow, owner, cost, risk, and expected result, then compares the promise with evidence. The audit separates useful deployment from theater, identifies waste, and gives leadership a ranked action plan for stop, consolidate, govern, or reinvest decisions.
Read definitionAI waste is the gap between what an organization spends or risks on AI and the value it can prove. It includes duplicate tools, unused seats, pilots that never ship, unsafe shadow use, outputs that need rework, and workflows where adoption rises while business results stay flat.
Read definitionThe AI Waste Index is Harmondale's diagnostic score for estimating how much AI value is leaking before deeper analysis. It weighs spend dispersion, adoption without value, data or control risk, and role drift, then turns the result into a priority lever and a 30/60/90 recovery path.
Read definitionThe Four Leaks of AI ROI is Harmondale's framework for the four places AI value usually escapes: overspend and dispersion, theater adoption, leaks and risk, and role drift and demotivation. It keeps discussions concrete by naming the operational failure mode before recommending another tool.
Read definitionThe AI ROI Recovery Method is Harmondale's five-step process for turning scattered AI activity into measurable value: audit the real AI footprint, map value, rationalize waste, redeploy useful workflows, then govern and measure over time. The method starts with evidence, not another model choice.
Read definitionShadow AI is the use of AI tools, prompts, extensions, agents, or automations without official approval, inventory, security review, or ownership. It often begins as helpful experimentation, then becomes a hidden risk because company data, vendor dependency, cost, and output quality sit outside governance.
Read definitionTheater adoption is AI usage that creates the appearance of progress without a measurable business result. Teams prompt, generate, summarize, or demo more work, but the workflow cost, quality, delay, revenue, or risk does not improve. Adoption becomes a performance, not evidence of return.
Read definitionAI tool rationalization is the disciplined review of AI subscriptions, licenses, copilots, agents, and vendor dependencies to decide what should stay, merge, stop, or be renegotiated. The goal is not fewer tools by default, but a stack where each tool has value, ownership, and control.
Read definitionAI governance is the operating system that defines which AI uses are allowed, who owns them, what data they can touch, how outputs are reviewed, how costs are monitored, and when a use case must be changed or stopped. Good governance enables investment instead of slowing it.
Read definitionThe AI productivity plateau is the point where more AI usage stops creating additional measured value. People generate more drafts, summaries, and automations, but bottlenecks, quality checks, exceptions, meetings, or unclear ownership absorb the gains. The organization sees motion, while throughput and profitability stay flat.
Read definitionUnused AI licenses are paid seats for copilots, assistants, or AI platforms that employees rarely open or cannot connect to useful work. They are easy to miss because the invoice renews quietly, the unused seat looks small, and no one compares paid access with real workflow adoption.
Read definition