Aureus Sovereign › AI glossary

An AI glossary for executives

Twenty words that turn up in board papers and vendor decks, what each one actually means, and the question it warrants the next time it appears. No technical background is assumed.

Agent

Where it appears: In a vendor deck, or a note from a division head

A system that is allowed to take several steps on its own, in sequence, towards an objective, without stopping to ask between each one. An ordinary assistant answers one question at a time; an agent is given a goal and works through it.

Ask: What is it allowed to do without checking with a person first?

Agentic

Where it appears: In board papers, describing a platform

Describes software that acts in steps of its own choosing rather than responding once per instruction. It is an adjective applied to a system, not a technology in itself.

Ask: Between the point where somebody asks and the point where we see a result, what actions has it taken?

Human in the loop

Where it appears: In the paragraph immediately after a risk is raised

An arrangement where a person is meant to approve something before it proceeds. The phrase says nothing about who that person is, what they are shown, or whether they have time to read it.

Ask: Which named person, at which step, and what exactly appears on their screen?

Model

Where it appears: Everywhere, usually without being defined

The engine underneath. It is bought or licensed from somebody else, and which one it is determines where the client's material is processed and under whose terms it is processed.

Ask: Whose model, running where, and under what contract?

Large language model, or LLM

Where it appears: In technical appendices and vendor comparisons

The kind of model behind most current AI assistants. It was trained on very large quantities of text and works by predicting what should come next, which is why it is fluent, fast, and occasionally confidently wrong.

Ask: What does it do when it does not know the answer?

Fine-tuned

Where it appears: In a proposal, as evidence of bespoke work

Adjusted on a body of the organisation's own past material so that output conforms more closely to its conventions. It is a real technique and it is also the easiest word in this list to use loosely.

Ask: Fine-tuned on which of our files, approved by whom, and where do those files now sit?

Prompt

Where it appears: In training material and casual use

The instruction given to the system. Most of the difference between a useless answer and a good one is in how much context the instruction carries, which is why standing instructions written once are worth more than clever phrasing each time.

Ask: Nothing. This one is straightforward.

Standing instructions, or system prompt

Where it appears: In setup discussions

A set of instructions that applies to every request without being retyped: the holder's position, their conventions of writing, what to assume and what never to do. Writing these properly is most of what makes a general tool feel bespoke.

Ask: Who wrote ours, and when were they last reviewed?

Context window

Where it appears: When somebody explains why a document was truncated

The amount of material the system can hold in view at once. A pack that exceeds it is not read in full, and the system will usually not say so.

Ask: Did it read the whole document, or the first part of it?

Token

Where it appears: In pricing schedules

The unit these systems are metered in, roughly a short word or part of a word. It matters only because usage is billed in tokens, so a long document is a bigger charge than a short one.

Ask: What does a typical month look like in real currency?

Hallucination

Where it appears: In the risk section, briefly

A confident, fluent, entirely invented answer. It is not a malfunction to be patched out; it is a property of how these systems work, and it is why anything consequential is checked.

Ask: What is the check, and who performs it?

Retrieval, or RAG

Where it appears: In proposals about using an organisation's own documents

Retrieval-augmented generation. The system retrieves passages from the organisation's own documents and answers from them rather than from memory. It is the usual method by which a general tool is made to answer on an organisation's own material.

Ask: Which repository is it retrieving from, and who can see what it returns?

Copilot

Where it appears: As a product name and as a category

A tool that operates inside software already in use and offers assistance in place. It is a category description as much as a product, so the name alone indicates little about what it is permitted to do.

Ask: Which one, sitting inside what, with access to which files?

Guardrails

Where it appears: Where a boundary is being described

Restrictions on what a system may do or say. They range from a genuine technical limit to an instruction the model is asked politely to follow.

Ask: Is this enforced outside the model, or is it an instruction the model can be talked out of?

On-premise, or air-gapped

Where it appears: In discussions about confidential material

Running the system on equipment the client controls. Air-gapped means that equipment has no network path out at all, so material physically cannot be transmitted.

Ask: Is there any route by which this material reaches the internet?

Inference

Where it appears: In technical costings

The act of running the model to produce an answer, as distinct from training it. Almost all ongoing cost is inference.

Ask: Nothing. It is a cost line, not a decision.

Training data

Where it appears: In due diligence and in contracts

The material a model was built from. The separate and more urgent question at director level is whether the organisation's material is being added to it.

Ask: Is what we send used to train anybody's model? Get it in writing.

Multimodal

Where it appears: In product announcements

Able to handle more than text: images, audio, documents as they are scanned rather than retyped. For executive work the practical meaning is that a photographed page or a scanned annex can be read.

Ask: Can it read our scanned documents, or only clean text?

Benchmark

Where it appears: In comparison slides

A standard test used to rank models. Rankings move monthly and a benchmark score rarely predicts whether a tool performs on the work actually in hand.

Ask: How did it do on our documents, not on the test?

Pilot

Where it appears: In the recommendation at the end of a deck

A limited trial. The word is doing a lot of work: a pilot that is never scoped to end becomes a permanent arrangement nobody decided on.

Ask: What would make us stop, and who decides?

That is most of the vocabulary. The rest is not harder. It has simply never been put in front of a board by anybody with nothing to sell in the room.