Glossary

AI glossary.

The vocabulary of an AI conversation, in the language of the person buying. Each entry says what the term is, then what it changes for the company — without the second half a definition is of no use.

Terms17

MCP (Model Context Protocol)RAG — grounding in your own data Embeddings and semantic search Context engineering Agent Tool use Structured output Human-in-the-loop Policy layer (guardrails)Evaluation set (evals)Observability and audit trail Hallucination Tokens and usage cost Model routing and provider independence Prompt management and versioning Data residency, KVKK and training exclusion Fine-tuning or grounding?

Connection and data

The terms about how a model reaches the company’s own information. This is the first and most decisive part of an AI project.

MCP (Model Context Protocol)

The open standard by which AI connects to company systems.

A protocol published as open source by Anthropic and since adopted as common ground across the industry. It lets an AI client connect to external systems — files, a database, an ERP, a service desk — through one standard interface. Instead of writing a separate integration per tool, a system is exposed once as an MCP server.

What it changes

For a company it means this: every operation the model can reach is defined individually. Which role may read what, and change what, is written at that interface and can be audited. That is the difference between AI having access to your data in general and reaching it through a defined door.

RAG — grounding in your own data

Answers produced from the company’s documents rather than the model’s memory.

Retrieval-Augmented Generation. When a question arrives the model does not answer directly: the relevant passages are first retrieved from the company’s documents, handed to the model, and the answer is produced only from those. The answer states which document it came from.

What it changes

This is the technical answer to the fear of invention. A model that does not know your procedure, your contract or your product catalogue will answer generically; grounding ties it to the company’s own reality. Citation then makes the answer checkable.

Embeddings and semantic search

Search that looks for meaning rather than words.

Text is converted into numerical vectors and search runs over those vectors. A question like “how do we cancel an invoice” then also finds a procedure clause that covers it without containing those words.

What it changes

Keyword search requires the user to know the right word. In corporate documents that rarely holds: staff ask in plain language while the document is written in regulatory language. Semantic search closes that gap.

Context engineering

Designing exactly what the model is given each time.

A model sees only what it is handed. Which document, how much of it, in what order and with which instruction is a design decision, and most of the quality is settled there.

What it changes

The same model gives very different answers depending on the context it receives. Most of what gets called “the model is bad” is in fact context that was assembled badly.

Execution

Where the model stops talking and starts doing. Both the value and the risk rise here.

Agent

AI that carries a multi-step job through on its own.

Unlike an assistant that answers one question at a time, an agent takes a goal, breaks it into steps, calls the tools it needs and returns a result. A request like “find this month’s overdue dealers, age the balances and prepare the report” runs from a single instruction.

What it changes

The value is here and so is the need for limits. Because an agent performs real operations, which steps require human approval and how a wrong step is undone must be designed upfront.

Tool use

The model calling a real operation in your system.

The model is given a set of operations it may call: “check stock”, “create a quote”, “produce a report”. The model decides which tool to call and when; the system performs the operation.

What it changes

This is what turns AI from a chat box into a working part of the business. Because each tool carries its own authorisation, access can be controlled one operation at a time.

Structured output

The model returning data your system can consume, not prose.

The answer is constrained to a predefined schema: fields, types and permitted values. The model cannot step outside it.

What it changes

It is the precondition for integration. Prose is enough for a person and not enough for a system; without structured output what the model produces cannot enter an ERP or a CRM.

Human-in-the-loop

The model drafts; a person decides.

The model prepares the work and the authority to publish it or start the transaction stays with a person. Approval points are written at the start of the project, and every irreversible action falls inside them.

What it changes

It settles both risk and accountability. Behind anything sent outside, any movement of money and any deletion, there is a named person.

Reliability and measurement

The part that turns a demonstration into a system somebody signs for — and the part an enterprise buyer asks about first.

Policy layer (guardrails)

The rules defining what the model may not say or do.

Forbidden topics, forbidden forms of steering, an obligation to speak only from supplied data, personal-data masking and validation of output against business rules are all defined here, along with the fallback behaviour when a rule is hit.

What it changes

This is what stops AI from creating corporate risk. In a finance application it is the ban on inventing figures and on steering an investor to buy or sell; in a health or legal context it is the boundary on advice.

Evaluation set (evals)

The fixed test set by which output is measured.

A set of real questions with expected answers is assembled. Every prompt or model change is scored against it, and a version below the agreed threshold does not ship. Failures seen in the field are collected and folded back into the set.

What it changes

Without measurement a quality discussion collapses into opinion. Keeping an AI system in production depends on being able to prove that a change did not make it worse.

Observability and audit trail

A record of which question got which answer, on what basis.

For every request the question, the context handed to the model, the tools called, the answer returned, the latency and the cost are recorded. Retention and who may read those records are defined.

What it changes

When a customer says “the system told me this”, this is where you look. Audit, debugging and cost tracking all run off the same record, and in enterprise procurement its absence is often disqualifying on its own.

Hallucination

The model stating something it does not know, with confidence.

Language models produce the most likely continuation, so where they do not know they can still write fluently and with conviction. No method removes this entirely today.

What it changes

The right question is not how to eliminate it but how to reduce and catch it. Grounding, citation, a policy layer, output validation and an evaluation set do that work together. A supplier who builds none of them has left the risk with you.

Operation, cost and compliance

What gets discussed once the system is live — and what decides both the budget and the legal exposure.

Tokens and usage cost

The unit AI cost is measured in.

Models process text in pieces called tokens, and pricing generally follows the number processed. Long context and long answers are cost.

What it changes

This is what makes a monthly bill predictable. Without measuring at real volume during a pilot, a budget is only a guess — and without a cost per transaction, a decision to scale cannot be made.

Model routing and provider independence

A different model per task, and no lock-in to one supplier.

A small, inexpensive model handles simple classification while a stronger one handles complex reasoning, and failover to a secondary model is defined for when a provider degrades.

What it changes

It lowers cost and keeps you from depending on one supplier’s pricing or policy changes. Because this field moves quickly, an architecture where the model can be swapped is what keeps it valid two years from now.

Prompt management and versioning

The system instruction as a managed artefact rather than code.

The instruction the model follows is edited from an admin panel, versioned, and rolled back when needed, with changes tried in a staging environment before they ship.

What it changes

It puts the tone and the limits of the text in the hands of the person who knows the business rather than the developer. Waiting for a release to change a sentence is where AI systems most often stall.

Data residency, KVKK and training exclusion

Where data goes, where it stays, and whether it trains the model.

Which country’s servers a request reaches, how long it is retained, whether it contains personal data and whether the provider may use it for training are each defined separately. Where data must not leave the premises at all, the model runs on your own servers.

What it changes

It is the first thing legal and IT ask in enterprise procurement, and a supplier without a written answer is usually eliminated there. Where there are European clients, the GDPR processing inventory follows from these same definitions.

Fine-tuning or grounding?

Training a model on your own data is unnecessary for most companies.

Fine-tuning changes a model’s behaviour using your examples; it costs data, time and money, and has to be repeated whenever the information changes. Grounding keeps the knowledge outside the model, so the system updates the moment a document does.

What it changes

The right answer to “let’s train an AI on our data” is usually grounding rather than fine-tuning. Projects that start without making the distinction end up with an expensive model nobody can update.

We build these

Most of what is defined here is a structure to be built rather than a phrase. The AI layer page sets out module by module how we build each one.

AI layer Readiness assessment

Looking to build something with AI?

Tell us about the process — and if it is not a fit, we will say so plainly.

Get in touch