Definitions of the AI terms that come up in executive conversations, vendor pitches and governance work, written in plain language.
These 47 terms are additions to a 53-term rewrite that is not in the copy folder yet. The live glossary stays at its current address until that rewrite is ready.
New sections
- Agents and how they are built
- Model behavior and capability
- Regulation and assurance
- Adoption, cost and the market
Agents and how they are built
Foundation model
A large, general-purpose model trained on broad data and adapted to many tasks by prompting or fine-tuning. The term separates these from smaller models built for one job.
Frontier model
The most capable models available at a given moment, from the handful of labs training at that scale. The label moves; what counted as frontier eighteen months ago is now mid-tier and considerably cheaper.
Small language model (SLM)
A compact model built for a narrow task, cheap and fast enough to run on modest hardware or on a device. Often the right answer for high-volume, low-complexity work where a frontier model is overkill.
Subagent
A child agent spawned by a main agent to handle one scoped task in its own context, returning a condensed summary rather than everything it read. It is how agent systems handle work too large for a single context window.
Orchestration
Coordinating multiple models, agents, tools and steps into one working process. As organizations move past single-tool use, orchestration is where most of the engineering effort now goes.
Multi-agent system
Several agents working on one problem, usually a lead agent directing specialized subagents. It buys parallelism and isolation at the cost of far higher token use and real coordination difficulty.
Agent memory
What an agent retains: short-term memory within a session, and persistent memory carried across sessions. Persistent memory is what separates a tool that starts cold every time from one that accumulates context about the work.
Context engineering
The discipline of deciding what information occupies a model's context at any moment, and what gets retrieved, compressed or dropped. Anthropic's applied team gave it a canonical definition in September 2025, and it has largely displaced prompt engineering as the term for serious work with agents.
Compaction
Summarizing an ongoing conversation or task history so work can continue past the context limit without losing the thread. Necessary for any long-running agent, and lossy by nature.
Tool call
A model invoking an external function, search, database or system rather than answering from its own knowledge. Tool calls are what turn a model from something that talks into something that does.
MCP server
A service that exposes tools and data to any AI system speaking the model context protocol. The vocabulary settled during 2026 around server, client, transport, primitive and sampling; older framework-specific terms such as plugin are being retired.
Agent skills
Packaged instructions and scripts that teach an agent how to do a specific task, loaded when relevant rather than held permanently in context. An increasingly common alternative to building a dedicated integration for every capability.
Computer use
A model operating a computer directly, reading the screen and controlling keyboard and mouse, rather than working through an application programming interface. It opens up systems that were never built to be automated, and it raises the governance stakes considerably.
Model behavior and capability
Chain of thought
A model working through intermediate steps before giving an answer. It improves accuracy on multi-step problems and is the mechanism behind reasoning models.
Test-time compute
Spending more processing at the moment of answering, rather than more training beforehand, to get a better result. It is why a model can be told to think longer on a hard problem, and why that costs more.
Mixture of experts (MoE)
An architecture where only part of the model activates for any given request. It allows very large models to run at the cost of much smaller ones, and explains why parameter counts have become a poor comparison tool.
Distillation
Training a smaller model to reproduce the behavior of a larger one. It is how capable, inexpensive models keep arriving months after the frontier model they learned from.
Grounding
Tying a model's output to a verifiable source, usually retrieved documents or live data, rather than letting it answer from training alone. Grounding is the practical defense against confident invention.
Context rot
Degradation in a model's attention to material as the context grows, particularly for information sitting in the middle. It is why filling a large context window with everything available often performs worse than supplying less, better-chosen material.
Benchmark saturation
The point at which models score so highly on a published test that it no longer separates them. Vendor benchmark claims should be read with this in mind; a near-perfect score often means the test is finished, not the problem.
Synthetic data
Data generated by a model rather than collected from the world, used for training or testing. Useful where real data is scarce or sensitive, and risky when its flaws compound quietly.
Model routing
Automatically sending each request to the model that fits it, reserving expensive models for work that needs them. One of the few cost levers that does not require giving anything up.
Nondeterminism
The property that the same prompt can produce different output on different runs. It is why AI systems need evaluation rather than testing, and why a demo that worked once proves less than it appears to.
Regulation and assurance
EU AI Act
The European Union's comprehensive AI law, in force since August 2024 and applying in phases. Prohibited practices and AI literacy duties began in February 2025, general-purpose AI obligations in August 2025, and most remaining obligations including high-risk rules were scheduled for August 2026, with enforcement sitting with the AI Office and member state authorities from that date. A provisional Digital Omnibus agreement in spring 2026 would defer several high-risk deadlines; it was not final as of this writing and should be confirmed before anyone plans around it.
High-risk AI system
Under the EU AI Act, a system in a category where failure carries serious consequences, such as employment, credit, education or essential services. High-risk classification brings documentation, human oversight, data quality and record-keeping obligations that other systems do not carry.
General-purpose AI model (GPAI)
A model capable of a wide range of tasks and used as the basis for downstream applications. The EU AI Act places transparency and documentation duties on GPAI providers, and requires them to give downstream builders what those builders need to comply.
Conformity assessment
The process of demonstrating that a high-risk AI system meets its legal requirements before it goes to market. It is the mechanism that turns an AI law from a statement of principles into a gate a product has to pass.
ISO/IEC 42001
The certifiable international standard for an AI management system, published in late 2023. Certification is becoming the practical way multinational organizations demonstrate governance maturity across jurisdictions rather than proving it separately to each.
NIST AI Risk Management Framework (AI RMF)
A voluntary United States framework organized around four functions: govern, map, measure and manage. Widely used as the backbone of internal AI governance programs, including by organizations with no regulatory obligation at all.
AI inventory
A maintained record of every AI system an organization uses, who owns it, what data it touches and what it is rated. It is the first artifact nearly every framework asks for, and the one most organizations cannot produce.
Algorithmic discrimination
Unlawful disparate treatment produced by an automated system. Definitions differ meaningfully between states, which is the practical problem for any organization operating in more than one.
Model card and system card
Published documentation of what a model or system is, how it was evaluated, what it is not suited for and what is known about its failure modes. Increasingly expected in vendor review and required in some regimes.
AI disclosure
Telling people when they are interacting with an AI system or seeing AI-generated content. Requirements vary by jurisdiction and are among the most likely to apply to organizations that consider themselves low risk.
AI literacy obligation
The EU AI Act duty requiring organizations to ensure staff working with AI systems have adequate understanding of them. It is the rare regulatory requirement that training actually satisfies, and it applied from February 2025.
Serious incident reporting
The duty to notify authorities when a high-risk AI system causes or nearly causes significant harm. It requires an organization to be able to detect such an incident in the first place, which is a capability question, not a paperwork question.
Graduated autonomy
Classifying AI agents by how much they decide and act without a person, then applying oversight proportional to that level. It is the emerging approach to governing agents, and it maps poorly onto policies written for chat tools.
Data provenance
A documented account of where data came from, what rights attach to it and how it has been handled. It underpins both regulatory defensibility and any honest answer about whether a model should have been trained on something.
Content credentials
Tamper-evident metadata attached to media recording how it was made and whether AI was involved, based on the C2PA standard. Adoption is uneven, and it is currently the most credible technical answer to provenance in images and video.
Adoption, cost and the market
AI center of excellence
A central team that sets standards, evaluates tools and supports AI work across an organization. It works when it enables the edges and fails when it becomes a queue everyone routes around.
Champion network
Named people across departments who are further along with AI and support their colleagues directly. Peer proof spreads through an organization in a way that a mandate from the executive floor does not.
Agentic organization
An operating model in which agents carry a meaningful share of routine work and people move to direction, judgment and exception handling. Currently more forecast than practice, and the term is increasingly used in vendor pitches where automation would be the honest word.
Cost per task
What a completed unit of work costs in model usage, rather than what a license costs per seat. It is the number that makes AI spending comparable to the work it replaces, and almost nobody tracks it.
Token budget
A deliberate allocation of how many tokens a task or system may consume. Agent workloads can consume many times what a chat conversation does, which turns token budgeting from a technical detail into a cost control.
LLMOps
The operational practice of running AI systems in production: versioning prompts, evaluating output, monitoring cost and quality, and managing model changes. The discipline that turns a working pilot into something dependable.
Observability
Being able to see what an AI system actually did: what it was asked, what it retrieved, which tools it called and what it returned. Without it, debugging an agent is guesswork and governance claims are unverifiable.
AI slop
Low-quality AI-generated content produced at volume because producing it is nearly free. It is a reputational risk for any organization publishing at scale and a growing reason audiences discount what they read.
Generative engine optimization (GEO)
The practice of structuring content so AI assistants and AI search features cite it. Also called answer engine optimization, large language model optimization and AI search optimization; as of 2026 no settled distinction separates the terms, and they are used interchangeably.