The vocabulary of enterprise AI, without the jargon

A model trained on vast text corpora to predict the next word — and, in doing so, to write, summarise, code or reason.

A large language model learns the statistical regularities of language from vast corpora. GPT (OpenAI) and Claude (Anthropic) are two such families. It does not "know" anything in the sense of a database: it generates the most likely continuation — hence the need to ground it on your data (see RAG) and to keep it within guardrails.

The elementary unit the model reads and produces — a fragment of a word, not a character nor a whole word.

A token is worth three to four characters of English on average. API billing and context-window size are both counted in tokens: controlling their consumption (caching, chunking, model choice) is a direct lever on cost.

The amount of text the model can "hold in mind" at once — question, supplied documents and answer included.

An extended context window lets you analyse a framework contract or an entire annual report without artificially chunking it. It is one of the strengths of recent models (Claude, GPT) and a selection criterion for document-heavy use cases.

A plausible but false answer, produced with confidence by the model.

Because an LLM generates the most likely continuation, it can invent a reference or a figure. You reduce this — never to zero — by grounding answers on verifiable sources (RAG), citing those sources, and keeping human validation on actions that commit the organisation.

The instruction given to the model — and the craft of phrasing it to obtain a reliable, reproducible result.

A good prompt frames the role, the expected format, the constraints and the examples. In the enterprise, you don't stop at the hand-crafted prompt: you version prompts, test them against evaluation sets and treat them as code.

Re-training a model on your own examples to specialise its style or its task.

Fine-tuning adjusts the model's weights on an in-house dataset. It is useful for a very specific format or tone, but often more costly and less flexible than RAG for bringing in knowledge: the call is made case by case, on evidence.

Turning a text into a vector of numbers that captures its meaning, so it can be compared by similarity.

Two texts close in meaning have close embeddings. This is the mechanism that lets you retrieve, from your document base, the passages relevant to answering a question — the foundation of RAG and of semantic search.

A model able to process several types of input — text, image, sometimes audio.

A multimodal model reads a scanned document, describes a photo of a production line or analyses a chart. In the enterprise, this opens up the analysis of heterogeneous attachments without a prior conversion pipeline.

Making the model answer from your documents: you retrieve the relevant passages, supply them to it, and it writes grounded on them.

RAG (Retrieval-Augmented Generation) grounds answers on your corpus rather than on the model's memory alone: answers cite their sources, stay up to date when your documents change, and hallucinate less. It is the default architecture for a trustworthy enterprise assistant.

An open standard that connects models to your tools and data through reusable, governed servers.

The Model Context Protocol replaces fragile, ad hoc integrations with a maintainable layer: an MCP server cleanly exposes a CRM, an ERP or a document base to an agent, with the access rights that go with it. It is one of Hunter BI's signature areas of expertise.

A system that doesn't just answer: it plans steps and acts in your tools to accomplish a task.

An agent chains actions — reading a file, querying a database, drafting a document, triggering an operation — under control. Best practice is not the autonomous demo but the governed agent: sourced answers and human validation on whatever commits the organisation.

A database designed to store embeddings and retrieve those closest to a query.

It is the meaning-based search engine behind a RAG: it indexes your documents as vectors and returns, in a few milliseconds, the passages most relevant to a given question.

The interface through which your applications call a model programmatically, without going through a chat interface.

The API (OpenAI, Claude) is the foundation of your products and automations: tool-augmented calls, batch processing, prompt caching, per-task model choice. This is where production integration is designed — with observability, budgets and continuous evaluations.

The controls that bound what an AI system is allowed to say or to do.

Input and output filtering, an agent's action scope, escalation to a human, logging: guardrails turn a raw capability into a system deployable within a regulated organisation.

Routing each use case to the model best suited in quality, cost and latency, instead of imposing a single one everywhere.

A frontier model for complex tasks, lighter models for volume — all behind an abstraction layer that preserves your reversibility. It is an architecture trade-off, not a fashion.

The framework that makes AI use controllable: who decides, who is accountable, what gets traced.

A registry of models in production, risk classification, a named owner, evidence captured automatically: without this foundation, pilots die because no one can approve them. Governance is not the brake, it is the runway.

The first international standard for an artificial-intelligence management system.

ISO/IEC 42001 describes how to steer AI responsibly: risk management, roles, documentation, continual improvement. It is a structuring path for an organisation that wants to prove control over its uses — not a stamp obtained in a day.

Control over where your data lives and which law applies to it.

Data residency, a dedicated tenant, retention rules, reversibility: for a regulated or public-sector player, knowing where data is processed and under which jurisdiction is often an entry condition. Sovereign deployment answers that requirement.

The contractual guarantee that your content is not used to train the provider's models.

On the enterprise offerings from OpenAI and Anthropic, your data is not used to train the models. We document these commitments in your compliance file and complement them with appropriate access controls and logging.

The Commission Nationale de contrôle de la protection des Données à caractère Personnel — Morocco's data-protection authority.

In Morocco, any processing of personal data falls under Law 09-08 and the oversight of the CNDP. An enterprise AI deployment must comply with it: a legal basis, informing data subjects, framing data transfers. We factor this in from the scoping stage.

OpenAI's enterprise offering: ChatGPT with organisation-grade administration, security and confidentiality.

SSO, centralised administration, security controls, extended context, and the commitment that your data does not train the models. It is one of the two flagship products we deploy — from scoping to connectors, governance and adoption.

Anthropic's enterprise offering: Claude with administration, very large context and confidentiality guarantees.

Shared spaces, SSO, extended context, data not used for training. Suited to large corpora — legal, finance, technical — and to engineering via Claude Code and MCP. The second flagship product we put to work in your teams.

Models that no longer just write text: they carry out tasks by driving software.

Agentic AI plans and acts — navigating an application, filling in a form, producing a deliverable — within a supervised environment. Claude Cowork (Anthropic) and "Computer Use" capabilities are examples: value moves from the developer's desk to every business function, under supervision.

The full cost of an AI platform over time — not just the sticker licence price.

Licences, token consumption, integration, operations, training: the three-year TCO reveals the true cost of a scenario. It is the basis for a platform choice built like an architecture case, with the numbers to back it.