The white paper glossary

The 50 terms used in the white paper How Agentic AI Is Entering the Enterprise Now, with the book's definitions (Appendix B), in the book's order.

A

A2A (Agent2Agent)

Open protocol released by Google in April 2025, governed since by the Linux Foundation, that lets AI agents discover one another and coordinate actions.

ACP (Agentic Commerce Protocol)

Open standard co-developed by OpenAI and Stripe that structures the exchanges between an AI agent, a buyer and a merchant to complete a purchase.

Agentic agent

A system that receives a goal, plans, calls tools, observes, corrects and drives the task to a terminal state.

Agentic RAG

Architecture that combines agentic reasoning with document retrieval: the agent decides what to search for, where, how many times, and judges whether the evidence is sufficient.

AP2 (Agent Payments Protocol)

Extension of A2A released by Google that represents every payment executed by an agent as cryptographically signed mandates, proving the user's authorization.

Apertus (EPFL, ETH Zurich, CSCS)

Multilingual open-weight model family developed by a Swiss academic consortium (EPFL, ETH Zurich and the Swiss National Supercomputing Centre). Documentation: apertus.ai and huggingface.co/swiss-ai.

Tool call

A structured invocation of an external function by the agent: searching a document, reading a calendar or creating a CRM opportunity.

Multi-model architecture

An architecture able to draw on several providers or model families without locking the company's knowledge and processes into a single engine.

B

Knowledge base

A governed set of validated, attributed, dated sources, accessible under explicit permissions, used by one or more AI systems.

C

Conversational channel

The interface through which the user talks with the agent: web, WhatsApp, phone, email, SMS or another messaging app.

ChatGPT and GPT (OpenAI)

ChatGPT is OpenAI's consumer application (United States); GPT is the model family that powers it. Documentation: openai.com/index and platform.openai.com/docs.

Chunking

Splitting documents into coherent fragments before vectorization.

Claude (Anthropic)

Language model family developed by Anthropic (United States), available as Opus, Sonnet and Haiku. Documentation: anthropic.com/news and docs.anthropic.com.

Copilot (Microsoft)

Microsoft's line of AI assistants (United States) built into its products, backed by Azure AI Foundry. Documentation: microsoft.com/copilot and learn.microsoft.com/azure/ai-foundry.

Intelligence layer (LLM)

The component that understands, reasons, plans, synthesizes and phrases; it replaces neither the knowledge base, nor memory, nor permissions.

D

DeepSeek

Chinese developer of open-weight models, known for its V and R families and its performance-to-cost ratio. Documentation: api-docs.deepseek.com.

Doubao (ByteDance)

Model family and consumer assistant developed by ByteDance (China), widely used in the Chinese domestic market. Documentation: volcengine.com.

E

Embedding

A numerical representation of a text that enables semantic comparison.

Escalation (human handoff)

A controlled handoff from the agent to a person, along with the identity, the history, the sources consulted and the recommended next action.

F

Faithfulness

A measure of how closely the generated answer matches the sources provided.

G

Guardrail

A deterministic control that bounds the agent's behavior, independently of the model.

Gemini (Google)

Multimodal model family from Google and Google DeepMind (United States), available in Pro and Flash versions. Documentation: blog.google and ai.google.dev.

GLM (Zhipu AI)

Open-weight model family developed by Zhipu AI (China). Documentation: bigmodel.cn and z.ai.

GraphRAG

A variant of RAG that uses a knowledge graph to link entities to one another.

Grok (xAI)

Model family developed by xAI (United States), built into X and available through an API. Documentation: x.ai/news and docs.x.ai.

H

Hallucination

The production of plausible but false information, in the absence of a source.

K

Kimi (Moonshot AI)

Long-context model family developed by Moonshot AI (China). Documentation: moonshot.ai and platform.moonshot.ai.

L

Llama (Meta)

Open-weight model family released by Meta (United States), widely used for self-hosted deployments. Documentation: ai.meta.com/llama.

LLM

Large language model, the engine for understanding and generating text.

M

MCP (Model Context Protocol)

Open standard for connecting AI systems and tools, governed since December 2025 by the Agentic AI Foundation.

Persistent memory

Retention of context and history beyond a single session.

Mistral AI

French developer of open-weight and proprietary language models, with the Le Chat assistant and an agents API. Europe's leading foundation model company. Documentation: mistral.ai/news and docs.mistral.ai.

Single-tenant

An architecture in which each customer has a dedicated infrastructure and database.

Multi-hop

Reasoning that requires several successive searches, each informed by the previous one.

N

Nova and Bedrock (Amazon)

Nova is Amazon's model family (United States); Bedrock is the AWS service that gives access to several model families. Documentation: aws.amazon.com/bedrock.

O

Opt-in (prior consent)

A regime in which a sales solicitation is lawful only after the explicit, prior consent of the person contacted, as opposed to opt-out, where it is lawful unless the person has refused in advance.

Orchestration

The layer that coordinates reasoning, rules, memory, document retrieval, tools and the end of the process.

P

Perplexity

Search engine built on generative AI (United States) that produces answers along with their citations. Documentation: perplexity.ai.

First useful response

The first reply that actually moves the request forward; distinct from an acknowledgment with no operational value.

Q

Qwen (Alibaba Cloud)

Open-weight model family developed by Alibaba Cloud (China), among the most widely adopted in the world for self-hosting. Documentation: qwen.ai and alibabacloud.com.

R

RAG

Retrieval-Augmented Generation: generating answers grounded in a document base queried at the moment of the question.

Dedicated RAG

A knowledge base and retrieval pipeline isolated for one company or entity, with no data mixed with other customers.

Reranking

Fine-grained re-ordering of candidate passages by a specialized model, after the initial search.

Model routing

The orchestrator's selection of the model best suited to a task based on language, latency, depth of reasoning, cost and data constraints.

S

Authoritative source

The document, system or content owner recognized as the reference when several versions of a piece of information exist.

Speed-to-lead

The time between the arrival of an inbound inquiry and the first useful response.

Human oversight

A human validation or intervention mechanism built into the agent's workflow.

T

Token

The basic unit an LLM uses to process information. APIs generally bill input and output tokens; the volume depends on the context passed in, the reasoning steps and the generated answer. The token is a major unit of LLM cost, but not the only cost of an agentic system.

U

Agentic unit of work

A discrete task completed by an agent: a decision made, a record updated, a process triggered.

W

Agentic web

An emerging evolution of the Internet in which AI agents discover, authenticate and exchange directly with one another, beyond human browsing of websites.

The full white paper is available as a PDF: in French · in English.