AI glossary for executives: 15 terms your supplier assumes you understand

Token

A token is the basic unit that AI models handle. Roughly: 1 token ≈ 4 characters in French, or about 0.75 word. A standard A4 page represents 500 to 600 tokens. Everything is billed by the token, in two directions: input tokens (what you send) and output tokens (what the model generates). Output tokens cost 3 to 5 times more than input tokens.

April 2026 prices, per million tokens (input / output):

Sources: Anthropic, OpenAI, Google AI, Mistral AI.

Inference

Inference is the operation the model performs to produce a response to a request. A crucial difference from a traditional SaaS: every inference consumes tokens and generates a variable cost. If your agent handles 100 requests a day, you pay 100 times. If it handles 10,000, you pay 10,000 times.

Cache (prompt caching)

The cache stores the repeated part of a request (system instructions, document base) so that it is not billed at full price on every call. At Anthropic, OpenAI and Google, a cache read costs 10% of the standard input price, i.e. a 90% saving on the cached portion. It is the lever most under-used by French suppliers.

For an agent handling 10,000 requests a month with 5,000 tokens of fixed context each time: without cache, 50 million tokens billed at full price. With cache (read at $0.30/M on Claude Sonnet), only 5 million at full price, the rest at $0.30/M. Saving: up to 90%.

Batch (asynchronous processing)

Batch mode lets you submit thousands of requests at once and receive the responses within 24 hours. Price: 50% of the standard price at Anthropic, OpenAI and Google. Relevant for mass document analysis, ticket classification, monthly newsletter generation, prospect database enrichment.

Combined with the cache, the cache + batch pairing can reduce the bill by up to 95% on compatible workflows. Cumulative and confirmed by the Anthropic documentation.

Model / LLM

A language model (LLM, Large Language Model) is the heart of a generative AI system. In 2026, the market is dominated by five main families: Claude (Anthropic), GPT (OpenAI), Gemini (Google), Mistral (FR), Llama (Meta). Each has its strengths, limits, prices and data policies. They are not interchangeable.

Context window

The context window is the maximum quantity of tokens a model can process in a single request, instructions and response included. It determines whether your agent can analyse an 80-page contract in one go, or whether it has to be split. In 2026: Claude Sonnet 4.6 reaches 1 million tokens (about 1,500 pages), GPT-5 200,000 to 400,000 depending on the version, Gemini 2.5 Pro 2 million.

Beyond the window, the model has no direct access to the information. That is why an agent "forgets" what was said at the start of a session after several hours.

Multimodality

A multimodal model can receive and generate several types of content at the same time. Claude 4.7, GPT-5 and Gemini 2.5 handle text + image natively. Specialised versions add voice (ElevenLabs, OpenAI Realtime) and video (Veo, Runway, Kling). In 2026, the voice agent on a switchboard or the analysis of scanned documents have become accessible to SMBs.

Hallucination

A hallucination is when a model produces a plausible but factually false answer, without signalling its uncertainty. According to the Stanford AI Index 2026, the average hallucination rate on large models has fallen from 38% in 2021 to 8.2% in 2026. The best systems reach 0.7% on simple tasks such as summarisation. But high-stakes fields (legal, medical) remain critical: 69 to 88% hallucinations on specialised legal queries.

RAG (Retrieval Augmented Generation)

RAG is the most common architecture in business in 2026. The principle: your documents (PDF, Word, Notion, database) are indexed in a vector database, then each question triggers a search in that database before the relevant passages are passed to the model. Result: an answer anchored in your sources, citable, updated within minutes after a new document is indexed.

According to 2026 benchmarks, RAG covers 80% of enterprise use cases. It is quicker to deploy, cheaper to maintain, and far more traceable than fine-tuning.

Fine-tuning

Fine-tuning consists of retraining a model on a specific dataset so that it learns a particular style, vocabulary or behaviour. It is not the addition of knowledge (RAG does that), it is the modification of behaviour. Useful for high-volume structured extraction, compliance with a proprietary format, or marked stylisation.

Cost of the operation: typically €1,000 to €50,000 depending on the size of the model and the quantity of data. Maintenance is non-trivial: every business change requires a new training cycle.

AI agent

An AI agent is not an enriched chatbot. It acts. It calls external tools (calendar, CRM, database, browser, telephone), executes chains of actions, makes intermediate decisions. According to Gartner, 40% of enterprise applications will have integrated task-specific agents by the end of 2026. The first ROI studies cited by AI players show an average return of around 171% in the first year on well-targeted deployments.

MCP (Model Context Protocol)

MCP is an open standard, launched by Anthropic in November 2024 and handed over to the Linux Foundation in December 2025. It lets an AI agent connect to any data source (Notion, Slack, GitHub, Google Drive…) or any business tool in a uniform way. No need to code a bespoke integration for each combination.

In March 2026, Anthropic counted more than 10,000 public MCP servers and 97 million SDK downloads a month. It has become the de facto protocol for agents in production.

Prompt and prompt engineering

The prompt is the instruction given to a model. A good prompt frames the role, the task, the expected format and the constraints. Prompt engineering is the discipline of optimising these instructions to obtain reliable, reproducible answers. It is not a trade of the future in the sense of a full-time position in every company: it is a cross-cutting skill that good suppliers master.

Open-weight vs closed

A closed model (GPT-5, Claude, Gemini) is only accessible via the supplier's API. An open-weight model (Mistral, Llama 4, DeepSeek V4) can be downloaded and hosted on your own, on your infrastructure or a sovereign cloud. Implications: data sovereignty, independence from a single supplier, the ability to fine-tune freely.

Mistral Large 3 at $2/M input and $6/M output remains the cheapest flagship model on the market on output: 2.5 times cheaper than GPT-5.4 or Claude Sonnet, with EU hosting by default.

AI Act

EU Regulation 2024/1689, known as the AI Act, classifies AI uses by level of risk and imposes graduated obligations. Prohibited practices (social scoring, subliminal manipulation) have applied since 2 February 2025. High-risk systems (HR, scoring, credit, biometrics) have been subject to reinforced obligations since 2 August 2026.

Penalties according to the level of breach:

• Prohibited practices: up to €35 million or 7% of worldwide turnover
• Breaches of high-risk system obligations: €15 million or 3% of turnover
• Inaccurate information to the authorities: €7.5 million or 1% of turnover

These penalties are cumulative with those of the GDPR. For an SMB with €5 million in turnover, the theoretical maximum exceeds €350,000 in the most severe category.

Source: European Commission.

Leaderboard / Bench

Rankings (LMArena, MMLU, ARC-AGI, HHEM for hallucinations) compare model performance on standardised tasks. Useful for spotting rapid changes, but limited: a model ranked no. 1 on LMArena can be mediocre on your use case. The most relevant benchmark remains the one you build on your own data.