# AI Lexicon Red Finch Advisory. Version 1.0, current as of 2026-10-21. 183 terms. https://redfinchadvisory.com/lexicon ## What is AI ### AI A machine-based system that makes predictions, recommendations, decisions or new content toward goals that people set. The NIST CSRC Glossary says: "A machine-based system that can, for a given set of human-defined objectives, make predictions, recommendations, or decisions influencing real or virtual environments." Also called: Artificial intelligence Source: https://redfinchadvisory.com/lexicon/ai (version 1.0, changed 2026-10-21) ### Generative AI A kind of AI that creates new content (text, images, audio, code). Related: AI-generated, Deepfake Source: https://redfinchadvisory.com/lexicon/generative-ai (version 1.0, changed 2026-10-21) ### Jagged frontier A pattern where a model does some tasks well and fails at others, and the line between them is uneven and hard to predict, even for tasks that seem equally hard. Also called: Jagged edge, Jagged intelligence Often confused with: Frontier model Related: Benchmark Source: https://redfinchadvisory.com/lexicon/jagged-frontier (version 1.0, changed 2026-10-21) ### Model The trained set of numbers (parameters) and structure that takes an input and produces an output; the part of an AI system that does the predicting. The Google for Developers Machine Learning Glossary says: "In general, any mathematical construct that processes input data and returns output. Phrased differently, a model is the set of parameters and structure needed for a system to make predictions." Also called: AI model Source: https://redfinchadvisory.com/lexicon/model (version 1.0, changed 2026-10-21) #### Foundation model A model trained on broad data, mostly with self-supervised learning, that can be adapted, for example through fine-tuning or post-training, for a variety of downstream tasks. Also called: GPAI model, General-purpose AI model Kind of: Model Source: https://redfinchadvisory.com/lexicon/foundation-model (version 1.0, changed 2026-10-21) #### Frontier model A foundation model that is among the current 'best-of-the-best' in the world; the set changes every few months. Also called: Frontier AI, State-of-the-art (SOTA) model Kind of: Model Often confused with: Jagged frontier Source: https://redfinchadvisory.com/lexicon/frontier-model (version 1.0, changed 2026-10-21) #### LLM A very large model trained on huge amounts of text to predict the next word, which lets it write, summarize, answer questions and follow instructions. The Google for Developers Machine Learning Glossary says: "At a minimum, a language model having a very high number of parameters. More informally, any Transformer-based language model, such as Gemini or GPT." Also called: Large language model Kind of: Model The other half: Small language model Source: https://redfinchadvisory.com/lexicon/llm (version 1.0, changed 2026-10-21) #### Open weights A model whose weights (numbers that the model learned in training) are published for anyone to download and run, usually without the training data or code, and under the vendor's license terms. Also called: Open-weight model, Open model Kind of: Model Related: Small language model, Parameters Source: https://redfinchadvisory.com/lexicon/open-weights (version 1.0, changed 2026-10-21) #### Reasoning model A model trained to work through a problem in a sequence of steps and plan its answer before it responds, spending more time and tokens to do better on hard tasks. Also called: Thinking model, Large reasoning model (LRM) Kind of: Model Related: Reasoning effort Source: https://redfinchadvisory.com/lexicon/reasoning-model (version 1.0, changed 2026-10-21) #### Small language model A language model that is small relative to frontier models, usually a few billion parameters, and can be run cheaply, often on local hardware. Also called: SLM Kind of: Model The other half: LLM Related: Open weights Source: https://redfinchadvisory.com/lexicon/small-language-model (version 1.0, changed 2026-10-21) ### Machine learning A way of building software where a model learns patterns from examples in data instead of following rules a person wrote; also the field that studies it. The Google for Developers Machine Learning Glossary says: "A program or system that trains a model from input data. The trained model can make useful predictions from new (never-before-seen) data drawn from the same distribution as the one used to train the model. Machine learning also refers to the field of study concerned with these programs or systems." Also called: ML Source: https://redfinchadvisory.com/lexicon/machine-learning (version 1.0, changed 2026-10-21) ### Multimodal A property of a model that can take in or produce more than one kind of media, such as text, images, audio or video. Not all models are able to generate images and audio. Related: Interface Source: https://redfinchadvisory.com/lexicon/multimodal (version 1.0, changed 2026-10-21) ### Open source AI An AI system released with its weights, its code and enough detail about its training data that anyone can use, study, change and share it for any purpose. The Open Source Initiative says: "An Open Source AI is an AI system made available under terms and in a way that grant the freedoms to: Use the system for any purpose and without having to ask for permission. Study how the system works and inspect its components. Modify the system for any purpose, including to change its output. Share the system for others to use with or without modifications, for any purpose." Many models sold as open source are really open weights under a restrictive license, and don't meet this test. Also called: Open source model Source: https://redfinchadvisory.com/lexicon/open-source-ai (version 1.0, changed 2026-10-21) ## How models work ### Inference The act of running a trained model to produce an output from inputs, as opposed to training it. Also called: Model serving, Serving The other half: Training Related: Cost per token Source: https://redfinchadvisory.com/lexicon/inference (version 1.0, changed 2026-10-21) #### Batch inference A type of inference where many requests are sent together and answered later, usually at a lower price. Also called: Offline inference, Batch API, Batch processing Kind of: Inference Related: Cost per token Source: https://redfinchadvisory.com/lexicon/batch-inference (version 1.0, changed 2026-10-21) #### Reasoning effort A setting which controls how long or how many tokens are spent by a model reasoning before providing an answer. Also called: Thinking budget, Reasoning budget, Thinking level, Test-time compute Kind of: Inference Related: Reasoning model, Cost per task Source: https://redfinchadvisory.com/lexicon/reasoning-effort (version 1.0, changed 2026-10-21) ### Context window The limit of how many tokens a model can hold in one call, counting both what you send and what it writes back. Related: Token Source: https://redfinchadvisory.com/lexicon/context-window (version 1.0, changed 2026-10-21) ### Embedding A numerical representation of the meaning of text, words or phrases (or an image), so that items with similar meaning sit close together and can be found by meaning, not keyword. Also called: Vectors Often confused with: Token Source: https://redfinchadvisory.com/lexicon/embedding (version 1.0, changed 2026-10-21) #### Embedding model A model that creates embeddings from content it is given, usually to fill a vector database and later to turn each search question into an embedding. Also called: Embedder, Text embedding model Kind of: Embedding Source: https://redfinchadvisory.com/lexicon/embedding-model (version 1.0, changed 2026-10-21) ### Knowledge cutoff The latest date used in the data that trains a model. Related: RAG Source: https://redfinchadvisory.com/lexicon/knowledge-cutoff (version 1.0, changed 2026-10-21) ### Parameters The numeric values a model learns in training, which decide how it turns an input into an output; their count, such as 8 billion, is the usual measure of a model's size. Related: Open weights Source: https://redfinchadvisory.com/lexicon/parameters (version 1.0, changed 2026-10-21) ### Token A unit of text that a model reads and writes in, a whole word or a piece of one, on average about three-quarters of an English word. Often confused with: Embedding Related: Context, Context window, Tokenmaxxing Source: https://redfinchadvisory.com/lexicon/token (version 1.0, changed 2026-10-21) ### Training The process of supplying data (often in large quantities) to a model so that it can identify patterns and make predictions. Also called: Model training Often confused with: How AI uses our data The other half: Inference Source: https://redfinchadvisory.com/lexicon/training (version 1.0, changed 2026-10-21) #### Fine-tuning A further round of training that adapts a pretrained model to specific tasks or a particular domain. NIST (AI 100-2e2025, Adversarial Machine Learning glossary) says: “The process of adapting a pre-trained model to perform specific tasks or specialize in a particular domain. This phase follows the initial pre-training phase and involves further training the model on task-specific data. This is often a supervised learning task.” Fine-tuning is one kind of post-training. Post-training usually means what the model maker does; fine-tuning usually means further training that anyone, often a customer, does for a task or domain. Fine-tuning changes how a model behaves and writes. It is a poor way to add facts that change often; RAG handles that better. Kind of: Training The other half: RAG Source: https://redfinchadvisory.com/lexicon/fine-tuning (version 1.0, changed 2026-10-21) #### Post-training The later stage of training that shapes how a model acts and behaves to match its creators' intentions, and builds skills such as following instructions, reasoning, and coding. Kind of: Training Source: https://redfinchadvisory.com/lexicon/post-training (version 1.0, changed 2026-10-21) #### Pretraining The first and largest stage of the training process that involves supplying broad or general data for a model to learn from. Also called: Pre-training Kind of: Training Source: https://redfinchadvisory.com/lexicon/pretraining (version 1.0, changed 2026-10-21) ### Deterministic A property of a system or check where the same input(s) always gives the same output(s). Also called: Rule-based The other half: Probabilistic Related: Deterministic guardrail Source: https://redfinchadvisory.com/lexicon/deterministic (version 1.0, changed 2026-10-21) ### Probabilistic A property of a system or check that works on likelihoods, so the same input(s) can give different outputs, and any single answer can be wrong some of the time. Also called: Non-deterministic, Stochastic The other half: Deterministic Related: Probabilistic guardrail Source: https://redfinchadvisory.com/lexicon/probabilistic (version 1.0, changed 2026-10-21) ## Ways of working ### Agentic A way of working in which an agent plans and carries out the work, decides its own next steps, and uses tools to complete a task or goal. Also called: Process AI, workflows, outcome-driven AI, Agentic AI The other half: AI-assisted Source: https://redfinchadvisory.com/lexicon/agentic (version 1.0, changed 2026-10-21) #### Agentic workflow A workflow with fixed steps, set by people or code, where a model handles some of the steps and uses judgment inside them. Kind of: Agentic The other half: AI-assisted workflow Related: Orchestration, Prompt chaining Source: https://redfinchadvisory.com/lexicon/agentic-workflow (version 1.0, changed 2026-10-21) #### Agentic coding A type of programming where a person gives a goal or loose instructions, and an agent decides its own next steps to plan, write, run and test the code. Also called: Agentic engineering, Agentic software development Kind of: Agentic Often confused with: Vibe coding, AI-assisted coding Related: Software factory, Full-stack builder Source: https://redfinchadvisory.com/lexicon/agentic-coding (version 1.0, changed 2026-10-21) #### Software factory A software development pipeline where people supply the instructions and checks, and agents do most of the building, testing and fixing. Kind of: Agentic Often confused with: Vibe coding Related: Agentic coding, Loop engineering, Human-on-the-loop, Full-stack builder Source: https://redfinchadvisory.com/lexicon/software-factory (version 1.0, changed 2026-10-21) #### Vibe coding A way of building software where a person describes what they want and accepts the code AI tools write without reading it, judging only whether the result works. A way to build software that involves focus primarily on the direction of the result, and little attention to the way the software is being built. Kind of: Agentic Often confused with: Agentic coding, Software factory Related: Cognitive debt, Citizen developer Source: https://redfinchadvisory.com/lexicon/vibe-coding (version 1.0, changed 2026-10-21) ### AI-assisted A way of working where a person decides each step and does the task, and AI tools help with parts of it. Also called: AI-augmented Often confused with: AI-generated The other half: Agentic Related: Cognitive debt Source: https://redfinchadvisory.com/lexicon/ai-assisted (version 1.0, changed 2026-10-21) #### AI-assisted coding A type of programming where a person decides each step and describes what code to write, and AI tools suggest and complete code based on that instruction. Also called: AI pair programming, AI code completion Kind of: AI-assisted Often confused with: Agentic coding Source: https://redfinchadvisory.com/lexicon/ai-assisted-coding (version 1.0, changed 2026-10-21) #### AI-assisted workflow A workflow where a person decides each step and carries it out, and AI tools help with those steps. Kind of: AI-assisted The other half: Agentic workflow Source: https://redfinchadvisory.com/lexicon/ai-assisted-workflow (version 1.0, changed 2026-10-21) ### AI-generated Content whose substance an AI tool produced, even if a person prompted, edited or approved it. Also called: Machine-generated, Synthetic content Often confused with: AI-assisted Related: Deepfake, Generative AI Source: https://redfinchadvisory.com/lexicon/ai-generated (version 1.0, changed 2026-10-21) ### Assistant A general-purpose AI app that you can talk to, such as the Claude app or ChatGPT. Also called: AI assistant, AI chat app Avoid: Chatbot Often confused with: Agent Source: https://redfinchadvisory.com/lexicon/assistant (version 1.0, changed 2026-10-21) #### Custom assistant An assistant set up with fixed instructions, documents and sometimes tools for one role or team. Also called: Custom GPT, Gem, Claude Project Avoid: Chatbot Kind of: Assistant Often confused with: Internal agent Source: https://redfinchadvisory.com/lexicon/custom-assistant (version 1.0, changed 2026-10-21) ### Workflow automation A rule-based workflow that runs a process from pre-set rules or pre-written code. Can include AI as a step. If a model handles some steps and uses judgment inside them, it becomes an agentic workflow. If an agent decides its own next steps, it is agentic. Also called: RPA, Rules-based automation, Process automation Often confused with: Agent Related: Orchestration, Computer use Source: https://redfinchadvisory.com/lexicon/workflow-automation (version 1.0, changed 2026-10-21) ## Agents ### Agent An AI system that works toward a goal by deciding its own next steps and using tools to take actions. What makes something an agent is that the model, not fixed code, decides what to do next. Describe any agent with four questions: what you hand it (autonomy), whose permissions it uses (identity), what it can change (permissions), and what starts it (trigger). Also called: AI agent Often confused with: Assistant, Workflow automation Source: https://redfinchadvisory.com/lexicon/agent (version 1.0, changed 2026-10-21) #### Coding agent An agent whose main job is writing and changing software. It can usually read and edit code, run commands and run tests. Autocomplete in an editor is AI-assisted coding, not a coding agent. The key question is what it can reach: the code repository, a command line, production systems or data. Also called: AI coding agent, AI software engineer Kind of: Agent Related: Full-stack builder Source: https://redfinchadvisory.com/lexicon/coding-agent (version 1.0, changed 2026-10-21) #### Personal agent An agent that acts on your behalf, using your permissions. It comes in two kinds. One uses your login and appears as you, so the logs usually cannot tell you apart. The other has its own name and a share of your permissions, so it can be limited and switched off without touching your account. Ask which kind a tool is before you connect it to email or company systems. Also called: Delegated agent, On-behalf-of agent Kind of: Agent Often confused with: Enterprise agent The other half: Internal agent Related: Acts as you, Acts for you Source: https://redfinchadvisory.com/lexicon/personal-agent (version 1.0, changed 2026-10-21) #### Background agent An agent that runs in the background, often in the cloud, while you do other work, and reports back when it is done. Also called: Cloud agent, Async agent Kind of: Agent Often confused with: Subagent Source: https://redfinchadvisory.com/lexicon/background-agent (version 1.0, changed 2026-10-21) ### Enterprise agent An agent with its own identity and permissions in your company's systems. It is set up and tracked like an account, not borrowed from a person, and it can run with nobody present. Also called: Digital worker, AI employee, Agentic workforce, Digital workforce, Digital labor Often confused with: Personal agent Related: Agent sprawl Source: https://redfinchadvisory.com/lexicon/enterprise-agent (version 1.0, changed 2026-10-21) #### Internal agent One agent that many people in a company can give work to. Shopify's River, which works in public Slack channels, is an example. Because many people direct one identity, decide whose limits apply: the agent's, or those of the person asking. If the agent can do more than the person asking, that person can get it to do what they could not. Also called: AI teammate, AI coworker, Digital colleague Kind of: Enterprise agent Often confused with: Custom assistant The other half: Personal agent Source: https://redfinchadvisory.com/lexicon/internal-agent (version 1.0, changed 2026-10-21) ### Multi-agent system A setup where multiple agents split up one job and coordinate, working simultaneously or in sequence. Also called: Agent swarms, Multi-agent orchestration, Orchestrator-workers, Subagents Source: https://redfinchadvisory.com/lexicon/multi-agent-system (version 1.0, changed 2026-10-21) #### Orchestration The coordination and triggering of agents or steps: deciding what runs when, passing results between them, and handling failures. Also called: Agent orchestration, Workflow orchestration Kind of: Multi-agent system Related: Agent harness, Agentic workflow, Workflow automation, Loop engineering Source: https://redfinchadvisory.com/lexicon/orchestration (version 1.0, changed 2026-10-21) #### Subagent An agent that another agent starts to do one delegated task in its own independent session, returning the result when done. Also called: Sub-agent, Worker agent, Child agent Kind of: Multi-agent system Often confused with: Background agent Related: Session Source: https://redfinchadvisory.com/lexicon/subagent (version 1.0, changed 2026-10-21) #### Router agent An agent which has responsibility to sort incoming requests and pass them to the appropriate agent. Also called: Triage agent, Dispatcher agent Kind of: Multi-agent system Often confused with: Model routing Source: https://redfinchadvisory.com/lexicon/router-agent (version 1.0, changed 2026-10-21) ## Autonomy and identity ### Agent identity The account an agent acts under, which sets who it appears as and whose permissions it uses. There are three patterns. It uses your login and appears as you. It acts for you under its own name, with some of your permissions. Or it has its own identity and permissions, granted by an admin. The pattern decides what the logs show, what it can access, and how IT can switch it off. It is a separate question from how autonomous the agent is. Also called: Non-human identity (NHI), Machine identity Related: Least privilege for AI, Stack, Kill switch Source: https://redfinchadvisory.com/lexicon/agent-identity (version 1.0, changed 2026-10-21) #### Acts as itself An agent using its own identity and its own permission set to take actions. Also called: Service account, Agent account, Own-identity agent Kind of: Agent identity Source: https://redfinchadvisory.com/lexicon/acts-as-itself (version 1.0, changed 2026-10-21) #### Acts as you An agent using your identity and permissions to take actions and appearing as you. When reviewing logs, the agent would appear as you in a system. Also called: Impersonation, Uses your login Kind of: Agent identity Related: Personal agent Source: https://redfinchadvisory.com/lexicon/acts-as-you (version 1.0, changed 2026-10-21) #### Acts for you An agent that acts for you under its own identity, with a limited share of your permissions, so logs usually show both the agent and you. Also called: Delegation, On-behalf-of (OBO) Kind of: Agent identity Related: Personal agent Source: https://redfinchadvisory.com/lexicon/acts-for-you (version 1.0, changed 2026-10-21) ### Levels of autonomy A scale of how much you hand over to an AI system or agent: nothing to act on, an instruction, an outcome, or an area to look after. Autonomy is a setting you choose, not a feature of the model. The same product can run at different levels. As autonomy rises, decide how a person oversees it (see Human oversight). Also called: Autonomy ladder, Levels of agency Related: Human oversight Source: https://redfinchadvisory.com/lexicon/levels-of-autonomy (version 1.0, changed 2026-10-21) #### Autonomous The top level of autonomy, where you hand an agent an area to look after and it acts on its own without a person approving each action. Also called: Autonomous agent Avoid: Fully autonomous Kind of: Levels of autonomy Often confused with: Human-out-of-the-loop Source: https://redfinchadvisory.com/lexicon/autonomous (version 1.0, changed 2026-10-21) #### Advisor A level of autonomy where an AI system or agent only gives advice, and does not take actions or make decisions. Kind of: Levels of autonomy Source: https://redfinchadvisory.com/lexicon/advisor (version 1.0, changed 2026-10-21) #### Task runner A level of autonomy where you give an AI system or agent one specific instruction, and it acts only on that task. Kind of: Levels of autonomy Related: Human-in-the-loop Source: https://redfinchadvisory.com/lexicon/task-runner (version 1.0, changed 2026-10-21) ### Human oversight Human involvement in an AI process to monitor, govern or stop it, before it acts, after it acts, or only for certain actions. Also called: Human review, Human supervision Related: Guardrail, Levels of autonomy, Kill switch, High-risk AI Source: https://redfinchadvisory.com/lexicon/human-oversight (version 1.0, changed 2026-10-21) #### Human-in-the-loop A setup where a person approves each action before it happens. Human in the loop means the person owns the goal and the steps: they tell the agent what to do or approve it at each step (e.g. accounts receivable where the person asks for each reminder draft, reviews it, sends it). Also called: HITL, Human approval Kind of: Human oversight Often confused with: Human-as-judge, Meat proxy The other half: Human-on-the-loop Related: AI brain fry, Task runner Source: https://redfinchadvisory.com/lexicon/human-in-the-loop (version 1.0, changed 2026-10-21) #### Human-on-the-loop A setup where an agent acts on its own while a person watches and can step in or stop it. The person owns the outcome and is responsible for it being done correctly; the agent owns the steps, and the person watches, checks the completed work and can step in or stop it. Also called: HOTL, Supervisory control Kind of: Human oversight The other half: Human-in-the-loop, Human-out-of-the-loop Related: Software factory, Kill switch, AI brain fry Source: https://redfinchadvisory.com/lexicon/human-on-the-loop (version 1.0, changed 2026-10-21) #### Human-out-of-the-loop A setup where an AI system or agent runs with no person watching, reviewing or approving its work. Also called: Fully automated, Unattended Avoid: YOLO mode Kind of: Human oversight Often confused with: Autonomous The other half: Human-on-the-loop Source: https://redfinchadvisory.com/lexicon/human-out-of-the-loop (version 1.0, changed 2026-10-21) ### Kill switch A control that stops an agent or model at once. Related: Agent identity, Human oversight, Human-on-the-loop, Use-case owner Source: https://redfinchadvisory.com/lexicon/kill-switch (version 1.0, changed 2026-10-21) ### Least privilege for AI A rule that an AI system or agent gets only the permissions it strictly needs for the task, and nothing more. Also called: Least privilege, Scoped access Related: Guardrail, Agent identity Source: https://redfinchadvisory.com/lexicon/least-privilege-for-ai (version 1.0, changed 2026-10-21) ## Instructions ### Context Everything the model can see when it answers: your instruction, the conversation so far, and any documents or tool results added to the call. Related: Token, In-context use Source: https://redfinchadvisory.com/lexicon/context (version 1.0, changed 2026-10-21) ### Prompt Text or data given to an AI system, usually by a user and often as an instruction, that is only part of the context the system sees when it answers. Also called: Instruction Often confused with: System prompt Source: https://redfinchadvisory.com/lexicon/prompt (version 1.0, changed 2026-10-21) #### Few-shot prompting A method of prompting where you supply some examples of inputs and the answers you expect from the model. Also called: In-context learning, Few-shot learning Kind of: Prompt Often confused with: In-context use Source: https://redfinchadvisory.com/lexicon/few-shot-prompting (version 1.0, changed 2026-10-21) #### Prompt chaining A simple workflow that uses the output of one prompt as the input of the next. Also called: Chained prompts Kind of: Prompt Related: Agentic workflow Source: https://redfinchadvisory.com/lexicon/prompt-chaining (version 1.0, changed 2026-10-21) #### Prompt template A shared prompt with sections that a person or a program fills in before sending. Also called: Saved prompt Kind of: Prompt Often confused with: Agent skill Source: https://redfinchadvisory.com/lexicon/prompt-template (version 1.0, changed 2026-10-21) #### Role prompting A method of prompting which instructs the model to act as a certain human-like role or persona. Also called: Persona prompt, Persona prompting Kind of: Prompt Related: System prompt Source: https://redfinchadvisory.com/lexicon/role-prompting (version 1.0, changed 2026-10-21) ### Context engineering The practice of choosing and managing everything a model sees on each call, not just the wording of the prompt. Wikipedia (Prompt engineering, 2026) says: "Context engineering is the related area of software engineering that focuses on the management of non-prompt and prompt contexts supplied to the GenAI model, such as system instructions, metadata, API tools and tokens." Also called: Context engineer Source: https://redfinchadvisory.com/lexicon/context-engineering (version 1.0, changed 2026-10-21) ### Memory A note that an AI system saves from past sessions (facts, preferences, decisions) and brings back into later ones to help guide it, since the model itself remembers nothing between calls. Also called: Long-term memory, Persistent memory Often confused with: Knowledge base Related: In-context use Source: https://redfinchadvisory.com/lexicon/memory (version 1.0, changed 2026-10-21) ### Session A continuous stretch of work, such as one chat or one run, in which the AI system carries earlier messages and work forward until it ends. Also called: Chat, Conversation, Thread, Run Related: Subagent Source: https://redfinchadvisory.com/lexicon/session (version 1.0, changed 2026-10-21) ### Prompt engineering The practice of crafting an instruction so that the model executes it well. Also called: Prompt design, Prompting Source: https://redfinchadvisory.com/lexicon/prompt-engineering (version 1.0, changed 2026-10-21) ## Connections and systems ### Agent harness The structure around a model that turns it into one working agent: its interface, agent loop, own context and memory, permissions, and list of allowed tools. The test for what belongs in it: does each agent carry its own, or do they all share one? If each agent has its own, it's harness. If they share it, it's stack. A harness wraps an agent, not a person. One person can run several agents, each with its own harness, and an agent can run with no person attached at all. Your personal assistant is one agent with one harness; that it's yours is about who it acts for, not about the harness belonging to a person. Why the agent is the unit: if the line were drawn around a person, a coding agent and a research agent run by the same person would be lumped into one harness, even though their loops, tools and permissions are completely different. The test stops working. Components that sit on the line: - Tools: which tools an agent may use is harness; the shared connector behind each tool is stack. A tool local to one person's setup is harness; a centrally hosted one is stack. - Permissions: the policy for one agent is harness; the central identity system that enforces it is stack. - Gateway: stack, because it is shared. People disagree on this one, so say it explicitly. Harness vs stack: the stack is the shared foundation: gateway, inference, data and ontology, evals, central identity, and shared connectors. The same harness concept appears in both the build stack (e.g. Claude Code: terminal, files, shell, git, human in the loop) and the product runtime stack (e.g. a support bot: narrow interface, business systems, tight permissions, little oversight). Same components, different tuning. Promotion is how something moves from a harness into the stack. The word grew up around personal coding agents (Claude Code, Codex, OpenCode), but it applies to customer-facing AI too: a product owner picks a harness for a chatbot. Choosing Claude Code vs OpenCode is mostly a harness decision, not an interface or model decision, since they can call the same models. Also called: Scaffold, Scaffolding, Agent runtime, Harness, Agent infrastructure Avoid: Harness engineering; Agent framework Related: Orchestration, Loop engineering Source: https://redfinchadvisory.com/lexicon/agent-harness (version 1.0, changed 2026-10-21) #### System prompt A prompt written by the developer or vendor, not the user, that the app sends alongside every user-supplied prompt to set the model's role, rules, and tone. Also called: Developer prompt, Preamble Kind of: Agent harness Often confused with: Prompt Related: Agent skill, Role prompting Source: https://redfinchadvisory.com/lexicon/system-prompt (version 1.0, changed 2026-10-21) #### Agent skill A package of instructions and files, written in an open text-based format, that an agent loads only when a task needs it. Also called: Skills, SKILL.md, Playbooks Kind of: Agent harness Often confused with: Tool, Prompt template Related: System prompt Source: https://redfinchadvisory.com/lexicon/agent-skill (version 1.0, changed 2026-10-21) #### Agent loop The cycle an agent repeats: decide the next step, take it with a tool, look at the result, and go again until the job is done or something stops it. Google for Developers (Machine Learning Glossary) says: “A cycle that an agent iterates through until a termination condition is met.” Also called: Agentic loop Kind of: Agent harness Often confused with: Loop engineering Source: https://redfinchadvisory.com/lexicon/agent-loop (version 1.0, changed 2026-10-21) #### Client The software that opens the connection and makes the call, to a model or to an MCP server. A role, not a product: ChatGPT with no tools is still a client. Because every AI tool contains a client, the word doesn't tell ChatGPT, Claude Desktop, Claude Code, Codex and OpenCode apart. What separates them is the harness around the model. In MCP's own terms, the app (Claude Desktop, ChatGPT) is the host, and the client is the piece inside it that holds the connection to one MCP server. Not to be confused with the business sense of a customer. Kind of: Agent harness Source: https://redfinchadvisory.com/lexicon/client (version 1.0, changed 2026-10-21) #### Interface The way a person interacts with an AI system: free text, document and image upload, audio, buttons and forms, in a chat window, terminal or IDE. A component of the harness. More isn't always better: a customer-facing chatbot may deliberately offer only fixed options with no free text. Rich features like generated forms and artifacts are hard for in-house builds to keep up with, which is part of the reason to buy a harness rather than build one. Also called: User interface, UI Kind of: Agent harness Related: Multimodal Source: https://redfinchadvisory.com/lexicon/interface (version 1.0, changed 2026-10-21) ### Grounding The practice of tying an AI system's answers to a real source that can be verified and checked. Also called: Grounded generation, Source grounding Related: GEO (generative engine optimization), Retrieval from our data Source: https://redfinchadvisory.com/lexicon/grounding (version 1.0, changed 2026-10-21) ### LLM gateway An intermediary that your company's calls to models route through; when it holds the only API keys, it is both a control point and a log of every call. Keep the two benefits separate: control (key custody, nothing routes around it) and observability (logs of what goes into and comes out of each call). Most gateways also give one API across model providers, with failover and cost tracking. Routing is a capability of the gateway: deciding in real time which model a request goes to. The decision can be a deterministic rule (size, cost, who is asking) or a model that classifies the request. It usually trades off cost, latency and capability: send easy work to a cheap model, escalate hard work to a frontier model. The catch is that the router has to guess how hard a request is before the work is done. Router and gateway can be separate, but splitting them is extra work most companies shouldn't take on. Also called: LLM gateway, AI gateway, model gateway, LLM router, model routing Often confused with: Agent gateway Related: Model-agnostic Source: https://redfinchadvisory.com/lexicon/llm-gateway (version 1.0, changed 2026-10-21) #### Model routing A step that picks, for each request, which model handles it, using a fixed rule or a model that judges the request. Google for Developers (Machine Learning Glossary, 2026) says: “The algorithm that determines the ideal model for inference in model cascading. A model router is itself typically a machine learning model that gradually learns how to pick the best model for a given input. However, a model router could sometimes be a simpler, non-machine learning algorithm.” In practice, many routers are fixed rules. Cascading is a different method: it tries a cheap model first and moves up only if the answer isn't good enough. Also called: LLM routing, Model cascading Kind of: LLM gateway Often confused with: Router agent Source: https://redfinchadvisory.com/lexicon/model-routing (version 1.0, changed 2026-10-21) ### RAG A technique for grounding that finds relevant text in your documents or data and adds it to the model's context, so the model writes its answer from that text without being changed. Also called: Retrieval-augmented generation The other half: Fine-tuning Related: Knowledge cutoff, GEO (generative engine optimization) Source: https://redfinchadvisory.com/lexicon/rag (version 1.0, changed 2026-10-21) #### Vector database A type of database that stores embeddings and can find the closest match to a query by meaning. Also called: Vector store, Vector index, Vector DB Kind of: RAG Often confused with: Knowledge base Source: https://redfinchadvisory.com/lexicon/vector-database (version 1.0, changed 2026-10-21) ### Stack The shared AI infrastructure that every agent and harness in your company draws on, such as gateways, model access, data, evals, central identity and shared connectors. Also called: AI stack, Tech stack Related: Agent gateway, Observability, Agent identity Source: https://redfinchadvisory.com/lexicon/stack (version 1.0, changed 2026-10-21) ### Connector A packaged link to one outside system, such as Google Drive or Salesforce, that gives an agent a set of tools to work with it. Also called: Integrations, Apps, Plugins Related: Retrieval from our data Source: https://redfinchadvisory.com/lexicon/connector (version 1.0, changed 2026-10-21) ### Knowledge base A curated collection of documents that a person or an AI system can access and use as context. Also called: KB, Knowledge store Often confused with: Memory, Vector database Source: https://redfinchadvisory.com/lexicon/knowledge-base (version 1.0, changed 2026-10-21) ### Tool A single action, such as running a task or a function, that an agent can call. Where a tool lives depends on where it can be reached, not who built it. A centrally hosted tool that others can use is part of the stack. A local tool scoped to one person's harness, like an advanced developer's local MCP server, is part of that harness. An agent's permission to use a tool is always harness. Building a central tool is a stack contribution; wiring your agent to it is harness configuration. Today tools are mostly central and shared; local, harness-only tools are the exception. Watch: personal agents may start creating tools on the fly for one person or task, which would shift part of the tool layer into the harness. Also called: Tool use, Tool calling, Function calling Often confused with: Agent skill Related: Model Context Protocol Source: https://redfinchadvisory.com/lexicon/tool (version 1.0, changed 2026-10-21) #### Computer use A tool that lets an agent work a computer screen the same way a person would (view, click, type). Also called: Computer-using agent (CUA), Browser use Kind of: Tool Related: Workflow automation, Sandbox Source: https://redfinchadvisory.com/lexicon/computer-use (version 1.0, changed 2026-10-21) #### Code execution A tool that runs code the model has written, usually in a sandbox. Also called: Code interpreter, Code sandbox Kind of: Tool Source: https://redfinchadvisory.com/lexicon/code-execution (version 1.0, changed 2026-10-21) ### A2A An open standard, created by Google and now run by the Linux Foundation, that lets agents from different vendors communicate and work together. The A2A Protocol project (Linux Foundation) says: “The Agent2Agent (A2A) Protocol is an open standard for seamless communication and collaboration between AI agents.” Also called: Agent2Agent, Agent-to-agent protocol The other half: Model Context Protocol Source: https://redfinchadvisory.com/lexicon/a2a (version 1.0, changed 2026-10-21) ### Agent gateway A control point in front of everything an agent can act on (tools, APIs and MCP servers) that governs what it can reach, enforces least privilege and logs its actions. Used three ways today: the outbound control point for agent actions; an MCP gateway in front of your tool servers; and a registry where agents get their own identities and permissions. The control point and the registry are two views of one job: where rules are enforced vs where they are stored. Once an agent can act, the risk is a real side effect (a write, a payment, a deletion), not just a bad answer. Also called: Tool gateway, MCP gateway, Tool-access gateway Often confused with: LLM gateway Related: Stack Source: https://redfinchadvisory.com/lexicon/agent-gateway (version 1.0, changed 2026-10-21) ### AI operations A discipline concerned with the AI lifecycle and running AI in production. Also called: LLMOps, MLOps Avoid: AIOps Source: https://redfinchadvisory.com/lexicon/ai-operations (version 1.0, changed 2026-10-21) #### Observability The tracking and recording of what an AI system does on each run: inputs, outputs, steps, tool calls, cost and errors. Some tools also record a model's visible reasoning. That is often a summary, and it does not always show what drove the answer. Also called: Tracing, AI observability, LLM monitoring, Agent tracing, AgentOps Kind of: AI operations Often confused with: Explainability, Production monitoring Related: Eval, Stack Source: https://redfinchadvisory.com/lexicon/observability (version 1.0, changed 2026-10-21) ### Loop engineering A practice of building where an agent keeps rerunning and checking its own work, with set rules for when it stops and when it asks a person to step in. Often confused with: Agent loop Related: Agent harness, Orchestration, Software factory Source: https://redfinchadvisory.com/lexicon/loop-engineering (version 1.0, changed 2026-10-21) ### Model Context Protocol An open standard, started by Anthropic in 2024 and now run under the Linux Foundation, for connecting AI systems and agents to outside tools and data in one common way. Wikipedia says: “The Model Context Protocol (MCP) is an open standard and open-source framework introduced by Anthropic in November 2024 to standardize the way artificial intelligence (AI) systems like large language models (LLMs) integrate and share data with external tools, systems, and data sources.” Also called: MCP The other half: A2A Related: Tool Source: https://redfinchadvisory.com/lexicon/model-context-protocol (version 1.0, changed 2026-10-21) #### MCP server A program that uses MCP to offer tools or data to agents, running on your own computer or as a hosted service; it is one way to build a connector. Kind of: Model Context Protocol Source: https://redfinchadvisory.com/lexicon/mcp-server (version 1.0, changed 2026-10-21) ### Sandbox An isolated environment where an agent can execute code or take actions without accessing live systems or data. Also called: Agent sandbox, Isolated environment Related: Guardrail, Computer use Source: https://redfinchadvisory.com/lexicon/sandbox (version 1.0, changed 2026-10-21) ## Evaluation ### Eval A measurement of how well an AI system does something (tone, helpfulness, accuracy, safety), usually as a score, graded by code, a model or a person. An eval doesn't have to be structured or a comparison: a one-off tone check on a single customer-service transcript is an eval. Not limited to models; you can eval a harness, a context setup, or other AI infrastructure. Comparison and benchmarking are uses of evals, not the definition. Edge case: an eval can have a deterministic answer (did the model solve the math problem?) and still be an eval, because it measures capability. Also called: Evaluations, Model evaluation, AI testing Avoid: Vibe check, Vibes-based eval Related: Red teaming, Observability, Reward hacking Source: https://redfinchadvisory.com/lexicon/eval (version 1.0, changed 2026-10-21) #### Benchmark A fixed, standardized eval suite held constant so comparisons are fair: across models, across approaches, or over time. Benchmarks are most commonly thought of as public test sets kept by research groups and AI labs. Your company can and should have its own private benchmarks, built around the tasks you need done, and run evals against them when something new is released or updated. Private benchmarks also avoid a known problem with public ones: the test questions may have leaked into a model's training data. Kind of: Eval Often confused with: Verified test set Related: Jagged frontier Source: https://redfinchadvisory.com/lexicon/benchmark (version 1.0, changed 2026-10-21) #### LLM-as-judge An eval where a model grades the work of another model or system. LLM-as-judge is an efficient way to evaluate quality but can be fraught with misalignment to your own judgment. Judges tend to favor longer answers, the first answer shown, and output from their own model family. It requires careful attention to alignment and appropriate instruction before you should have confidence in its judgment. Also called: LLM judge, AI judge, Model-graded eval Kind of: Eval Related: Probabilistic guardrail Source: https://redfinchadvisory.com/lexicon/llm-as-judge (version 1.0, changed 2026-10-21) #### Regression eval An eval designed to protect a working AI feature against unintended change. Classic software: a regression test, a deterministic check that known-good output stays the same. AI: checking a working feature still performs as well after a change, including the vendor updating or retiring the model. Regression eval: Example: a working chatbot whose tools or documents are rewritten to be more efficient; you need to confirm it works as well as before. Because the output isn't deterministic, exact-match tests aren't enough for the content, so you use eval machinery: a frozen test set, scored, with a threshold you don't drop below, run before and after. Also called: Regression test Kind of: Eval Related: Verified test set, Reward hacking Source: https://redfinchadvisory.com/lexicon/regression-eval (version 1.0, changed 2026-10-21) #### Verified test set A set of test cases that have had correct answers verified by a subject-matter expert. When used for evals: cases should reflect your company's real work; synthetic cases are fine when real data can't be used (common in healthcare, e.g. synthetic patient or customer records), as long as they are representative. The expert sign-off is non-negotiable: the expert owns the answer key and ideally stays involved in reviewing how evals score against it. Also called: Golden Dataset Kind of: Eval Often confused with: Benchmark Related: Regression eval Source: https://redfinchadvisory.com/lexicon/verified-test-set (version 1.0, changed 2026-10-21) #### A/B test An experiment where live users are split between two versions of a product or service and the results are compared. Also called: A/B testing, Split test, Online experiment Kind of: Eval Often confused with: Side-by-side eval Related: Pilot Source: https://redfinchadvisory.com/lexicon/a-b-test (version 1.0, changed 2026-10-21) #### Human-as-judge An eval where a person reviews a system's output and judges its quality. This can be unavoidable when it comes to niche needs like style or preference. Human grades are the standard other judges are checked against, though people often disagree with each other. Human grades are also the starting point to calibrate an LLM-as-judge. Also called: Human evaluation, Human review, Human rating Kind of: Eval Often confused with: Human-in-the-loop Source: https://redfinchadvisory.com/lexicon/human-as-judge (version 1.0, changed 2026-10-21) #### Production monitoring The practice of running evals on a live system to score the quality of its outputs. This type of eval is useful feedback for the owner of the product or system. People use the scores to improve prompts, data or models. Also called: Online evals, Live monitoring Kind of: Eval Often confused with: Observability Related: Production Source: https://redfinchadvisory.com/lexicon/production-monitoring (version 1.0, changed 2026-10-21) #### Side-by-side eval An eval that compares the outputs of two or more models or systems next to each other to judge which is better. Blind comparison, where you hide which model produced which output, is one of the most effective ways to remove your personal bias toward certain models. Side-by-side comparisons also show quickly which output fits your preferences. Also called: Side-by-side evaluation, Pairwise comparison, A/B comparison Kind of: Eval Often confused with: A/B test Source: https://redfinchadvisory.com/lexicon/side-by-side-eval (version 1.0, changed 2026-10-21) ### Red teaming Testing which imitates an adversarial attack on an AI system in order to learn more about the safety and security of the system. Red teaming is not a new concept, but it has some unique applications for AI systems. They can be attacked through plain language (prompt injection, jailbreaks), and they can produce harmful output with no attacker at all, so red teams test for both. Also called: AI red teaming, Adversarial testing Related: Eval Source: https://redfinchadvisory.com/lexicon/red-teaming (version 1.0, changed 2026-10-21) ## Risks ### AI safety The practice of protecting against harm that an AI system may cause or help cause, whether by its own mistakes or by doing what a person asks when it shouldn't. AI systems can sometimes act in a way that is harmful (e.g. deleting the wrong files, giving bad advice) even without an external party that has malicious intent. Systems need to be protected against themselves by having guardrails and safety nets to prevent unintended outcomes. The other half: AI security Source: https://redfinchadvisory.com/lexicon/ai-safety (version 1.0, changed 2026-10-21) ### Guardrail A check, built into the harness or written into the instructions, that screens what goes in, what comes out, or what an agent tries to do. Guardrails are an essential component of a robust AI system that has AI safety and AI security protections. Checks that run in code outside the model are much stronger than rules written in the instructions, which a model can ignore and an attacker can override. They can also be an important part of a product experience, ensuring that the system is behaving as intended and steering it back on track if not. Also called: Safety filters, Content filters, Policy checks, Input/output validation Related: Human oversight, Sandbox, Least privilege for AI Source: https://redfinchadvisory.com/lexicon/guardrail (version 1.0, changed 2026-10-21) #### Deterministic guardrail A guardrail that uses fixed rules or code so that it applies the same way every time. Also called: Hard rule, Hard limit, Rule-based filter Kind of: Guardrail Related: Deterministic, Jailbreak Source: https://redfinchadvisory.com/lexicon/deterministic-guardrail (version 1.0, changed 2026-10-21) #### Input guardrail A guardrail that checks or filters what goes into a model before the model processes it. Also called: Input filter, Input validation, Prompt filter Kind of: Guardrail Related: Prompt injection Source: https://redfinchadvisory.com/lexicon/input-guardrail (version 1.0, changed 2026-10-21) #### Output guardrail A guardrail that checks or filters what comes out of a model before anyone sees it or it takes effect. Also called: Output filter, Content filter, Output validation Kind of: Guardrail Related: Data exfiltration Source: https://redfinchadvisory.com/lexicon/output-guardrail (version 1.0, changed 2026-10-21) #### PII redaction A type of guardrail that removes or masks personal data in text passing through a system, usually before it reaches a model or log. Also called: Data masking, PII masking, De-identification Kind of: Guardrail Related: Data exfiltration Source: https://redfinchadvisory.com/lexicon/pii-redaction (version 1.0, changed 2026-10-21) #### Probabilistic guardrail A guardrail that uses a model or classifier, so its judgment can be wrong: it can miss bad content, block good content, or be tricked. Also called: Model-based check, Classifier Kind of: Guardrail Related: LLM-as-judge, Probabilistic Source: https://redfinchadvisory.com/lexicon/probabilistic-guardrail (version 1.0, changed 2026-10-21) ### Hallucination Output from an AI system that is false or made up but stated with confidence, such as wrong facts, faulty reasoning or sources that don't exist. Hallucinations are what make an AI system especially risky. Even if a hallucination is rare, it can jeopardize the safety of an entire system. Hallucinations are products of the probabilistic nature of models. Grounding in real sources, asking for citations and checking the output reduce them, and guardrails can catch some, but nothing removes them. Newer models hallucinate less on some tasks, but not on all of them. Also called: Confabulation, Fabrication, Making things up Related: AI slop, Sycophancy Source: https://redfinchadvisory.com/lexicon/hallucination (version 1.0, changed 2026-10-21) ### Sycophancy A model's tendency to tell you what you seem to want to hear. When prompting a model, if you ask it to agree with you, it is likely to do so, or at least be agreeable. Models also flatter and agree when you don't ask, and often back down from a correct answer when you push back. This is dangerous for people who tend to prompt in a way that asks for their preconceived notion to be validated rather than questioned. Sycophancy is generally considered a negative attribute of models. Also called: People-pleasing, Flattery, Glazing Related: Alignment, Hallucination Source: https://redfinchadvisory.com/lexicon/sycophancy (version 1.0, changed 2026-10-21) ### AI slop Low-value, likely flawed AI output released without adequate checking by a person or a system. AI slop is defined by a result, not a tool. It's usually obvious due to a pattern that seems wordy, unnatural, or recognizable, and can be repulsive to people. It can be damaging to credibility and confidence in a person or company. Also called: Slop, Workslop, Slop grenade, Slopper, Deckslop Related: Hallucination, Meat proxy Source: https://redfinchadvisory.com/lexicon/ai-slop (version 1.0, changed 2026-10-21) ### Alignment The degree to which a model acts in line with human values and the goals and intent of the people directing it. There are two levels to alignment: Broad: does the model act in line with human values and the intent of an org? Everyday: does this system actually do what you want? Labs train models for alignment, but the useful question is always: aligned to whom? Related: Sycophancy, Reward hacking Source: https://redfinchadvisory.com/lexicon/alignment (version 1.0, changed 2026-10-21) ### AI security The practice of defending an AI system, and the data and tools it can reach, from attackers and deliberate misuse. AI security is concerned with defending a system from misuse, especially intentional misuse that could jeopardize the model, its data, its tools or its outputs. AI safety covers harm with no attacker, including accidents. Also called: Security for AI, LLM security The other half: AI safety Source: https://redfinchadvisory.com/lexicon/ai-security (version 1.0, changed 2026-10-21) #### Data exfiltration A type of AI system attack which attempts to have the system send private data outside your company or to anyone not permitted to see it. Permissions are the strongest defense: an agent that can't access sensitive data, or can't send anything out, can't leak it. Guardrails reduce the risk but can't make this attack impossible. Also called: Data leakage, Data leak Kind of: AI security Related: Output guardrail, PII redaction, Indirect prompt injection Source: https://redfinchadvisory.com/lexicon/data-exfiltration (version 1.0, changed 2026-10-21) #### Indirect prompt injection A type of prompt injection where instructions are hidden in content that the AI system reads (web page, email, file, etc.). Kind of: AI security Related: Data exfiltration Source: https://redfinchadvisory.com/lexicon/indirect-prompt-injection (version 1.0, changed 2026-10-21) #### Jailbreak A type of AI system attack which attempts to 'convince' a system to ignore its rules, guardrails, or system instructions. There are levels of guardrails that can defend against an attack of this sort, but the viability of these attacks is why deterministic guardrails should also be in place. Kind of: AI security Often confused with: Prompt injection Related: Deterministic guardrail Source: https://redfinchadvisory.com/lexicon/jailbreak (version 1.0, changed 2026-10-20) #### Prompt injection A type of AI system attack where text the system reads contains instructions, and the system follows them as if they came from its user or owner. Prompt injection is one of the most basic and most effective attacks that systems need to defend against, and there is no complete fix today. Limiting what the system can access and do reduces the damage. Kind of: AI security Often confused with: Jailbreak Related: Input guardrail Source: https://redfinchadvisory.com/lexicon/prompt-injection (version 1.0, changed 2026-10-21) ### Bias A skew in how an AI system treats people or cases, or in what it produces, inherited from its data or built in by design. Bias is inherently present in an AI system. Data cannot be unbiased, so systems that use it, especially systems that train on it, cannot be unbiased. Not every skew is a bad thing. A deliberate skew, such as favoring recent documents, can improve a system. The harm is bias that treats groups of people unfairly. The challenge for areas like AI ethics is to find that bias, measure it, and decide what is fair to people. Also called: Algorithmic bias, AI bias Related: Explainability Source: https://redfinchadvisory.com/lexicon/bias (version 1.0, changed 2026-10-21) ### Deepfake Image, audio or video made or changed with AI so that it falsely appears to show a real person, place or event. The EU AI Act (Regulation (EU) 2024/1689, Art. 3(60), 2024-06) says: "‘deep fake’ means AI-generated or manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic or truthful;" Related: AI-generated, Generative AI Source: https://redfinchadvisory.com/lexicon/deepfake (version 1.0, changed 2026-10-21) ### Reward hacking A failure where a model or agent games how it is measured, getting a high score without the intended outcome, such as editing tests so they pass instead of fixing the code. Related: Eval, Regression eval, Alignment Source: https://redfinchadvisory.com/lexicon/reward-hacking (version 1.0, changed 2026-10-21) ## Oversight ### Approved tools A set of AI tools that your company has vetted and approved for work use, often only for certain data or tasks. Also called: Sanctioned AI tools, Allowlist Related: AI governance, Vendor training Source: https://redfinchadvisory.com/lexicon/approved-tools (version 1.0, changed 2026-10-21) #### Shadow AI The use of unapproved AI tools, or personal accounts on approved ones, for work tasks with work data. Also called: BYOAI (bring your own AI), Unsanctioned AI Kind of: Approved tools Often confused with: Agent sprawl, Citizen developer Related: AI inventory, Vendor training Source: https://redfinchadvisory.com/lexicon/shadow-ai (version 1.0, changed 2026-10-21) ### How AI uses our data The ways an AI system can use your company's data: reading it during a task, which leaves the model unchanged, or training on it, which changes the model. Also called: Training on our data Often confused with: Training Related: Moat Source: https://redfinchadvisory.com/lexicon/how-ai-uses-our-data (version 1.0, changed 2026-10-21) #### Fine-tuning on our data Adjusting the weights of a model by training it further on your own data. Also called: Private fine-tune, Custom model training Kind of: How AI uses our data Source: https://redfinchadvisory.com/lexicon/fine-tuning-on-our-data (version 1.0, changed 2026-10-21) #### In-context use Supplying company-specific data to a model in the prompt so that it can provide a more relevant response, without changing the model. Also called: Pasted context, Context stuffing Kind of: How AI uses our data Often confused with: Few-shot prompting Related: Memory, Context Source: https://redfinchadvisory.com/lexicon/in-context-use (version 1.0, changed 2026-10-21) #### Retrieval from our data RAG applied to your company's data: the AI system fetches the relevant data as needed to inform a response or action. Also called: Grounding on our documents, RAG on our data Kind of: How AI uses our data Related: Connector, Grounding Source: https://redfinchadvisory.com/lexicon/retrieval-from-our-data (version 1.0, changed 2026-10-21) #### Vendor training A vendor using your company's data, such as prompts and outputs, to train or improve its AI models. Also called: Training on your chats, Model improvement Kind of: How AI uses our data Related: Shadow AI, Approved tools Source: https://redfinchadvisory.com/lexicon/vendor-training (version 1.0, changed 2026-10-21) ### Responsible AI The practice of AI ethics: turning principles like fairness, safety, privacy and accountability into actual controls. Also called: Trustworthy AI, Ethical AI, RAI Related: ISO/IEC 42001 Source: https://redfinchadvisory.com/lexicon/responsible-ai (version 1.0, changed 2026-10-21) #### AI ethics A field of inquiry into what is right and fair in how AI is built and used: the principles and the 'what should we do' questions. Also called: Ethics of AI Kind of: Responsible AI Source: https://redfinchadvisory.com/lexicon/ai-ethics (version 1.0, changed 2026-10-21) ### AI governance A structure of accountability, oversight, and policy for how your company builds, buys, and uses AI. Also called: AI risk management, AI compliance Often confused with: Model risk management Related: Approved tools Source: https://redfinchadvisory.com/lexicon/ai-governance (version 1.0, changed 2026-10-21) ### AI governance committee A standing group of leaders from legal, risk, security, IT and the business that sets your company's AI policy and approves higher-risk AI uses. Also called: AI council, AI steering committee, AI governance board Related: Chief AI Officer, Use-case owner Source: https://redfinchadvisory.com/lexicon/ai-governance-committee (version 1.0, changed 2026-10-22) ### AI inventory A list of the AI systems that your company uses or builds, each with an owner, purpose, and risk level. Also called: AI register, Model inventory, AI system inventory Related: Shadow AI, Agent sprawl, Citizen developer Source: https://redfinchadvisory.com/lexicon/ai-inventory (version 1.0, changed 2026-10-21) ### Agent sprawl A governance problem where agents multiply without control, so no one knows which agents exist, who owns them, what they are for, or what they can access. Often confused with: Shadow AI Related: AI inventory, Enterprise agent, Citizen developer Source: https://redfinchadvisory.com/lexicon/agent-sprawl (version 1.0, changed 2026-10-21) ### Deployer A company or other body that uses an AI system under its own authority, as opposed to the provider that makes it. The EU AI Act (Article 3(4)) says: "a natural or legal person, public authority, agency or other body using an AI system under its authority except where the AI system is used in the course of a personal non-professional activity." Many people assume a deployer is the builder. Under the Act, a company that uses an AI system at work is a deployer; the company that makes it is the provider. Also called: AI deployer Often confused with: Use-case owner The other half: Provider Related: High-risk AI Source: https://redfinchadvisory.com/lexicon/deployer (version 1.0, changed 2026-10-21) ### Explainability The ability to understand how or why a model reached a conclusion and produced a given output. Also called: Explainable AI (XAI), Interpretability Often confused with: Observability Related: Bias Source: https://redfinchadvisory.com/lexicon/explainability (version 1.0, changed 2026-10-21) ### High-risk AI An AI system used where errors can seriously affect people's health, safety or rights, such as hiring, credit, education or critical infrastructure, which brings strict legal duties. The EU AI Act sets out which systems count in Article 6 and Annex III. Providers and deployers of these systems have extra duties, such as risk management, human oversight and record keeping. Some US state laws, such as Colorado's, use their own definitions, and these have changed recently, so check the current text. Also called: High-risk AI system Related: Deployer, Human oversight Source: https://redfinchadvisory.com/lexicon/high-risk-ai (version 1.0, changed 2026-10-21) ### ISO/IEC 42001 An international standard, published in 2023, that sets requirements for the policies, roles, risk checks and reviews your company uses to manage AI, and that an outside auditor can certify against. Also called: ISO 42001, AI management system (AIMS) Related: Responsible AI Source: https://redfinchadvisory.com/lexicon/iso-iec-42001 (version 1.0, changed 2026-10-22) ### Model risk management A risk discipline, set by bank regulators, that inventories, independently validates and monitors the models a bank relies on, so that a wrong or misused model does not cause losses. Also called: MRM, AI risk officer Often confused with: AI governance Source: https://redfinchadvisory.com/lexicon/model-risk-management (version 1.0, changed 2026-10-22) ### Provider A company or other body that builds an AI system or model, or has one built, and offers it under its own name, paid or free. The EU AI Act (Article 3(3)) says: "a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge." A company that only uses a system is a deployer. It becomes the provider of a high-risk system if it puts its own name on it or makes a substantial change to it (Article 25). Also called: Developer, AI provider, Model provider The other half: Deployer Source: https://redfinchadvisory.com/lexicon/provider (version 1.0, changed 2026-10-21) ## People and roles ### AI champion A person driving AI adoption and enablement inside a team or function. The local person who helps colleagues actually use the tools, surfaces use cases, and connects the floor to the central AI effort. Embedded and usually part-time and unofficial, unlike a central, dedicated center of excellence (CoE). Also called: AI ambassador, AI advocate Often confused with: Forward deployed engineer Related: AI center of excellence, AI-pilled, Chief AI Officer Source: https://redfinchadvisory.com/lexicon/ai-champion (version 1.0, changed 2026-10-21) ### Chief AI Officer A senior executive who is accountable for your company's AI strategy, adoption and governance. Also called: CAIO, Head of AI, AI leader Related: AI governance committee, AI champion, AI center of excellence Source: https://redfinchadvisory.com/lexicon/chief-ai-officer (version 1.0, changed 2026-10-22) ### Evals engineer A person or team whose job is to measure AI systems using verified test sets, benchmarks and regression evals. Also called: Eval engineer, AI evaluation engineer, AI quality engineer Related: AI product manager, Product engineer, Agent manager, Full-stack builder Source: https://redfinchadvisory.com/lexicon/evals-engineer (version 1.0, changed 2026-10-21) ### Agent manager A person whose job is to oversee a group of agents at work, by setting their tasks and standards, checking the quality of their output, and fixing them when they fail. Also called: Agent supervisor, Agent operator Related: Evals engineer, Use-case owner Source: https://redfinchadvisory.com/lexicon/agent-manager (version 1.0, changed 2026-10-22) ### AI center of excellence An enterprise org that is dedicated to improving the usability of AI within the company. Also called: AI CoE, Center of excellence, Hub-and-spoke model, AI enablement team, AI platform team Related: AI champion, Chief AI Officer, Citizen developer Source: https://redfinchadvisory.com/lexicon/ai-center-of-excellence (version 1.0, changed 2026-10-22) ### AI product manager A person responsible for AI features built into products. Also called: AI PM Often confused with: Use-case owner Related: Full-stack builder, Evals engineer, Product engineer Source: https://redfinchadvisory.com/lexicon/ai-product-manager (version 1.0, changed 2026-10-22) ### Citizen developer An employee who sits outside the formal IT team and uses tools available to build software to support their team or company. Also called: Agent builder, Maker, Business technologist Often confused with: Shadow AI Related: Agent sprawl, Vibe coding, AI inventory, AI center of excellence Source: https://redfinchadvisory.com/lexicon/citizen-developer (version 1.0, changed 2026-10-22) ### Forward deployed engineer An engineer who works inside a business unit to build a working solution for the company through to production. An FDE often identifies the goals and steps that make up a job, which is the groundwork Agentic needs. Also called: FDE, Deployment engineer, Embedded engineer, Deployment strategist, Applied AI engineer Often confused with: Product engineer, AI champion Related: Use-case owner Source: https://redfinchadvisory.com/lexicon/forward-deployed-engineer (version 1.0, changed 2026-10-22) #### External FDE An engineer from a software or AI vendor who works inside a customer's company to build a working solution on the vendor's product and get it into production. Also called: Vendor FDE Kind of: Forward deployed engineer The other half: Internal FDE Related: Use-case owner Source: https://redfinchadvisory.com/lexicon/external-fde (version 1.0, changed 2026-10-21) #### Internal FDE An engineer on your company's own staff who sits inside a business unit to build and run AI systems for that team's work. Also called: Embedded engineer Kind of: Forward deployed engineer The other half: External FDE Source: https://redfinchadvisory.com/lexicon/internal-fde (version 1.0, changed 2026-10-21) ### Full-stack builder A person who uses AI tools to take a product from idea to launch alone, doing the work of a product manager, a designer and an engineer. LinkedIn made this a formal title in late 2025 and opened it to people from any function. Elsewhere, the word builder is mostly a label product managers use when they build working software with coding agents instead of handing specs to engineers. It overlaps with product engineer, which reaches the same scope from the engineering side. It matters because it signals that some companies now plan teams around one person per product idea instead of three specialists. Also called: Associate product builder Often confused with: Product engineer Related: AI product manager, Agentic coding, Coding agent, Evals engineer, Software factory Source: https://redfinchadvisory.com/lexicon/full-stack-builder (version 1.0, changed 2026-10-22) ### Product engineer A role which entails responsibility to design and build features through production to meet a user need. Often confused with: Full-stack builder, Forward deployed engineer Related: AI product manager, Evals engineer Source: https://redfinchadvisory.com/lexicon/product-engineer (version 1.0, changed 2026-10-22) ### Use-case owner A named person, usually a business leader, who is accountable for one use of AI in your company, including its results, its risks and its compliance. Also called: AI system owner, Business owner Often confused with: Deployer, AI product manager Related: Forward deployed engineer, External FDE, AI governance committee, Kill switch, Agent manager Source: https://redfinchadvisory.com/lexicon/use-case-owner (version 1.0, changed 2026-10-22) ### AI brain fry Mental fatigue from guiding, steering and overseeing more AI tools and output than you can handle. Related: Human-in-the-loop, Human-on-the-loop Source: https://redfinchadvisory.com/lexicon/ai-brain-fry (version 1.0, changed 2026-10-21) ### AI literacy The ongoing work of understanding how AI works, where it fails, and how to use it well. Also called: AI fluency, AI upskilling Source: https://redfinchadvisory.com/lexicon/ai-literacy (version 1.0, changed 2026-10-21) ### AI-pilled A person, team, or company that is fully on board with working AI-first. Also called: AGI-pilled Related: AI champion Source: https://redfinchadvisory.com/lexicon/ai-pilled (version 1.0, changed 2026-10-21) ### Cognitive debt The growing gap between what you have produced with AI and what you understand, which you pay for later when you need to fix, explain or build on it. Related: Vibe coding, AI-assisted, Meat proxy Source: https://redfinchadvisory.com/lexicon/cognitive-debt (version 1.0, changed 2026-10-21) ### Entry-level gap A drop in hiring for early-career roles in jobs exposed to AI, as AI takes over the routine work that junior people used to learn on. Also called: Missing rung, Canaries in the coal mine Avoid: Broken rung (McKinsey and LeanIn use it for women passed over for their first promotion to manager) Related: The great flattening Source: https://redfinchadvisory.com/lexicon/entry-level-gap (version 1.0, changed 2026-10-22) ### Meat proxy A person who approves or passes along AI output without verifying or reading it. Often confused with: Human-in-the-loop Related: AI slop, Cognitive debt Source: https://redfinchadvisory.com/lexicon/meat-proxy (version 1.0, changed 2026-10-21) ### The great flattening A wave of restructuring that cuts layers of middle management and gives each remaining manager more people to lead, often with AI named as the reason. Also called: Flattening, Delayering, Unbossing Related: Entry-level gap, AI washing Source: https://redfinchadvisory.com/lexicon/the-great-flattening (version 1.0, changed 2026-10-22) ### Tokenmaxxing The practice of measuring AI adoption by the number of tokens used, which pushes people to run up usage to look productive. Related: Token, Cost per task Source: https://redfinchadvisory.com/lexicon/tokenmaxxing (version 1.0, changed 2026-10-21) ## Value case ### Cost per token The posted price of input and output tokens consumed or generated. Also called: Token price, Price per million tokens Related: Inference, Batch inference Source: https://redfinchadvisory.com/lexicon/cost-per-token (version 1.0, changed 2026-10-21) ### AI washing A claim that overstates how much AI a product, service or company uses, or how capable it is, to attract customers or investors. Usually marketing. Also called: Agent washing Related: Wrapper, The great flattening Source: https://redfinchadvisory.com/lexicon/ai-washing (version 1.0, changed 2026-10-21) ### Cost per task The total cost of completing one task with an AI system, counting every model call, retry and tool call, and ideally the human review time. Also called: Tokenomics, Cost per outcome, Cost per resolution Related: Reasoning effort, Tokenmaxxing Source: https://redfinchadvisory.com/lexicon/cost-per-task (version 1.0, changed 2026-10-21) ### Model-agnostic A design that lets your company, product or service swap one model or vendor for another with little rework, so it is not locked in to one provider. Also called: Multi-model, LLM-agnostic, Vendor-neutral Related: LLM gateway, Sovereign AI Source: https://redfinchadvisory.com/lexicon/model-agnostic (version 1.0, changed 2026-10-21) ### Sovereign AI An approach where a country or company keeps its AI (models, data and computing) under its own control and laws, insulated from foreign vendors and jurisdictions. Also called: AI sovereignty, Sovereign cloud Related: Model-agnostic Source: https://redfinchadvisory.com/lexicon/sovereign-ai (version 1.0, changed 2026-10-21) ### Usage-based pricing A pricing model based on consumption, such as per token, per call or per task, so cost rises and falls with use. Also called: Consumption pricing, Pay-as-you-go, Metered pricing, API pricing The other half: Per-seat pricing Related: Outcome-based pricing Source: https://redfinchadvisory.com/lexicon/usage-based-pricing (version 1.0, changed 2026-10-21) ### Wrapper A product that is mostly valuable because of basic and simple features that add onto an existing model. Also called: Thin wrapper, GPT wrapper, LLM wrapper Related: AI washing Source: https://redfinchadvisory.com/lexicon/wrapper (version 1.0, changed 2026-10-21) ### GEO (generative engine optimization) A marketing practice to shape content so that AI search and answer tools are more likely to find it and cite it. Related: RAG, Grounding Source: https://redfinchadvisory.com/lexicon/geo (version 1.0, changed 2026-10-21) ### Moat A lasting advantage, hard for competitors to copy, that makes a company, product, or service less likely to lose business. Also called: Defensibility, Competitive advantage, AI moat Related: How AI uses our data Source: https://redfinchadvisory.com/lexicon/moat (version 1.0, changed 2026-10-21) ### MVP The first real, usable version of a product: the smallest scope that delivers value to real users, with planned near-term improvements. The original startup meaning is the smallest version that tests whether people want the product at all. In large companies it has come to mean the first production release. Also called: Minimum viable product Often confused with: Proof of concept (POC), Pilot Source: https://redfinchadvisory.com/lexicon/mvp (version 1.0, changed 2026-10-21) ### Outcome-based pricing A pricing model which is linked to the result of using a product or service. Also called: Outcome pricing, Pay per resolution, Results-based pricing The other half: Per-seat pricing Related: Usage-based pricing Source: https://redfinchadvisory.com/lexicon/outcome-based-pricing (version 1.0, changed 2026-10-21) ### Per-seat pricing A pricing model with a fixed fee for each person per month or year, no matter how much they use it. Also called: Per-user pricing, Seat-based pricing, Per-user licence The other half: Outcome-based pricing, Usage-based pricing Source: https://redfinchadvisory.com/lexicon/per-seat-pricing (version 1.0, changed 2026-10-21) ### Pilot A release of a product to a limited group of real users doing real work, for a limited time, with a planned short-term review of its viability or effectiveness. Also called: Pilot program, Trial, Limited rollout Often confused with: MVP Related: A/B test Source: https://redfinchadvisory.com/lexicon/pilot (version 1.0, changed 2026-10-21) ### Production The state of a process or product that real work depends on and expects to be reliable. Also called: Prod, Live, General availability (GA) Related: Production monitoring Source: https://redfinchadvisory.com/lexicon/production (version 1.0, changed 2026-10-21) ### Proof of concept (POC) A small test of whether an idea or approach can work at all, usually on sample data and without real users. Also called: POC, PoC Often confused with: MVP Source: https://redfinchadvisory.com/lexicon/proof-of-concept (version 1.0, changed 2026-10-21)