AI glossary: the terms, defined plainly.

The terms that come up when a business is deciding how to run AI, defined plainly. Where a term has a technical meaning and a marketing meaning, the entry says which is which. Each definition links to the page on this site that goes deeper.

On-premise (on-prem, on-premises)

Software or hardware that runs at your own site, on machines you control. On-premise AI runs the model, your documents and the index on a unit in your building. See on-premise AI explained.

Cloud AI

AI delivered as a service from a provider’s data centre. Your prompts and documents are transmitted to the provider’s servers to be processed, and for the major providers those servers are outside Australia.

Public cloud

Cloud infrastructure shared among many customers, each renting a slice. Consumer AI chatbots run this way.

Private cloud

Cloud infrastructure provisioned for one organisation’s exclusive use. It can be hosted in a provider’s data centre or run on your own premises, so “private” says who it serves rather than where it is. See private cloud vs on-premise AI.

Hosted private cloud

A private cloud run for you in a provider’s data centre. Ermos Edge is a hosted private AI instance in an Australian data centre.

Sovereign AI

AI whose data, processing and control stay under Australian jurisdiction. The term is used loosely in the market; the precise questions are where the model runs, who holds the keys, and which law the operator answers to.

Data sovereignty

The principle that data is subject to the laws of the country where it is held and processed. An Australian business whose documents are processed offshore has placed them under another country’s law as well.

Data residency

Where data is stored. A provider can offer Australian data residency for storage while still processing questions elsewhere, which is why residency alone does not settle the sovereignty question.

Resident AI

AI that lives at the business: the model, the documents and the index run on hardware inside the business’s own premises, the business holds the keys, and a provider keeps it running. See what is Resident AI.

Tenant AI

Our term for multi-tenant cloud AI, where many customers’ prompts share the provider’s infrastructure and the provider holds the keys.

Frontier model

One of the largest and most capable AI models, trained by a major provider and delivered through the cloud. ChatGPT and Claude are built on frontier models. They lead at open-ended reasoning and cannot be installed on your premises.

Large language model (LLM)

The kind of AI model that reads and generates text. Every AI assistant, cloud or on-premise, is built on one. Open LLMs can be downloaded and run on your own hardware.

Inference

The act of running a model to produce an answer. Where inference happens is where your question and the excerpts sent with it are processed, which is the fact that matters for confidentiality.

Retrieval (grounding)

Finding the relevant parts of your own documents and giving them to the model along with your question, so the answer is grounded in your material rather than the model’s training. The parts retrieved are the excerpts.

Excerpts

The pieces of your documents retrieved to answer a question. On an on-premise unit they never leave it. On a hosted instance they are what travels to the model, so which excerpts a person is permitted to retrieve is the boundary that matters.

Access boundaries (permissions)

Rules about which people may see which documents, applied before a question reaches the model, so that nobody can obtain an answer drawn from material they are not allowed to open.

Fine-tuning

Adjusting a model to a particular business’s language, documents and standards. On Ermos Dominion, ErmosIQ does this on the unit itself, so the improvement belongs to the business.

Training data

The material a model learned from. Consumer AI services can use what users enter to train future models unless the user opts out; that is one of the ways information entered into a chatbot can leave a business’s control.

Hallucination

A confident, fluent answer that is wrong: an invented citation, a figure that does not exist, a rule misstated. Every generative AI can do it, which is why answers should show their sources and people must check them.

AI data leakage

Information leaving an organisation’s control through an AI tool: transmitted to a provider, retained on its terms, used for training, or reachable by the provider’s staff. See what leaks to AI, document by document.

APP 8

Australian Privacy Principle 8, on cross-border disclosure of personal information. Before disclosing personal information to an overseas recipient, an entity must take reasonable steps to ensure the recipient will not breach the APPs, and remains accountable if they do. See APP 8 and AI.

Common questions