Book a discovery call
Reference

The Plain-English AI Glossary

Every term an organisation meets on the way to adopting AI, defined in plain English — no vendor spin, honest about limitations. From private AI, sovereign AI and RAG to agents, data residency and hallucination. Each definition stands alone, so you can link colleagues straight to the term that settles the meeting.

Private & sovereign AI

Private AI

Artificial intelligence deployed inside infrastructure an organisation controls — its own servers, a local data centre, or a dedicated cloud tenancy — rather than on a vendor’s shared public platform. Data is processed under the organisation’s own governance rules, in a jurisdiction it chooses, with the organisation controlling the model version, access rules and update schedule.

Sovereign AI

AI capability kept under the control of a jurisdiction, sector or people: the data, the infrastructure it runs on, and the rules that govern it. Where "private AI" describes one organisation’s architecture, sovereign AI widens the lens — a country keeping citizen data onshore, or iwi keeping Māori data under Māori governance. A private AI deployment is the practical mechanism that makes an organisation’s AI use sovereign in fact.

Private LLM

A large language model running under an organisation’s own control instead of as a public service — typically a strong open-weight or licensable model deployed on-premise or in a dedicated tenancy. "Private GPT" and "private ChatGPT" are informal names for the same idea; the deployment provides equivalent capability without sending data to a public platform.

Self-hosted AI

AI that an organisation runs on infrastructure it operates itself — its own servers or its own cloud environment — rather than consuming as a vendor-hosted service. Often used interchangeably with "local AI". Self-hosting is one ingredient of private AI; a full private deployment adds retrieval over the organisation’s knowledge, access controls and logging.

On-premise AI (on-prem)

AI running on an organisation’s own physical hardware, in its own server room or racks. The strongest form of physical control over where data is processed. Compare private cloud AI, where the deployment runs in a dedicated but provider-operated environment.

Air-gapped AI

A deployment with no outbound internet connection at all: the model, data and logs live on an isolated network. Used where sensitivity justifies the operational overhead — defence, health, critical infrastructure. An air-gapped system trades convenience (no live web access, manual update process) for the strongest available isolation.

Confidential AI

AI protected by confidential-computing techniques — hardware-level enclaves that shield data even while it is being processed, so not even the infrastructure operator can inspect it. Complements, rather than replaces, the architectural controls of a private deployment.
Models & mechanics

Large language model (LLM)

The class of AI model behind modern AI assistants: trained on vast amounts of text to understand and generate language. An LLM supplies general language ability; making it useful and safe for an organisation requires the architecture around it — retrieval over approved knowledge, access controls and logging.

Open-weight model

A model whose weights — the trained parameters — are published, so organisations can download and run it on their own infrastructure. Open-weight models are what make private AI practical: strong capability without dependence on a vendor’s hosted service. "Open weight" is not the same as "open source"; licences vary and should be checked per model.

Model weights

The numerical parameters a model learned during training — effectively the model itself. Whoever holds the weights can run the model. Controlling where the weights run is the core of private AI.

Retrieval-augmented generation (RAG)

The technique that grounds an AI’s answers in an organisation’s own documents: relevant passages are retrieved from an approved knowledge base and supplied to the model alongside the question, so answers reflect what the organisation actually knows and can cite sources. RAG is how a deployment stays accurate about your policies and procedures without retraining the model.

Fine-tuning

Additional training that adjusts a model’s weights on an organisation’s examples — useful for style, format or narrow specialist tasks. Often confused with RAG: fine-tuning changes how the model behaves, RAG changes what it knows in the moment. Most organisational needs are met by RAG; fine-tuning is the exception, not the default.

Inference

Running a trained model to produce output — every question answered is inference. Inference is where operating cost and data exposure happen, which is why private AI focuses on where inference runs, not on training models from scratch.

Context window

The amount of text a model can consider at once — the question, retrieved documents and conversation so far. Content beyond the window is invisible to the model, which is why retrieval quality (choosing the right passages) matters more than raw window size for organisational deployments.

Token

The unit models read and produce — roughly three-quarters of an English word. Public AI pricing is usually metered per token, which is how costs scale unpredictably with adoption; private deployments typically replace per-token pricing with fixed running costs.

Hallucination

When a model states something false with confidence — the inherent failure mode of language models answering from general memory. Mitigated, not eliminated, by grounding answers in retrieved documents (RAG), constraining scope, requiring citations, and keeping humans accountable for consequential decisions.
Agents & automation

AI agent

An AI system that acts, not just answers: it can search systems, draft documents, update records and carry multi-step tasks toward a goal, pausing for human approval where required. The practical difference from a chatbot is system access and the ability to complete work.

Agentic AI

The broader pattern of AI systems that plan and execute multi-step work with a degree of autonomy — deciding what to do next, using tools, and checking results. In organisational deployments, agentic behaviour operates inside explicit boundaries: which systems, which actions, which approvals.

AI receptionist / AI voice agent

An AI system that answers phone calls and chat like a front desk: handling routine enquiries, taking messages, booking appointments and escalating what matters to a person — around the clock. Deployed privately, conversations and caller details stay inside the organisation’s own environment rather than an offshore SaaS platform.

Human-in-the-loop

The governance pattern where a person owns every consequential decision: the AI drafts, retrieves and recommends, and a human approves before anything that matters happens. Not a disclaimer — an architectural decision about where accountability sits, enforced by design.

Robotic process automation (RPA)

Traditional rules-based automation: software following fixed scripts against predictable inputs. Breaks when inputs vary. AI automation complements it by handling the judgement-shaped steps — reading meaning rather than matching patterns — with governance replacing script maintenance as the operating discipline.
Sovereignty, privacy & governance

Data sovereignty

The principle that data is subject to the laws and governance of the place — or people — it belongs to, and that its owner should control where it lives and who can access it. In practice for AI: knowing, and being able to prove, that regulated or sensitive information is processed within a jurisdiction and under rules you choose.

Data residency

Where data is physically stored and processed — the geographic fact. Related to, but narrower than, data sovereignty: residency says "the data is in New Zealand"; sovereignty adds "and it is governed by rules we control". A vendor can offer residency while still controlling the platform; sovereignty requires control of the boundary itself.

Māori data sovereignty

The principle that data about Māori is taonga and should be subject to Māori governance — a distinct, well-developed body of thinking in Aotearoa New Zealand that shapes public-sector procurement and data practice. For AI deployments, it strengthens the case for architectures where the data owner, not a platform vendor, controls access and use.

AI governance

The policies, boundaries and oversight that keep an organisation’s AI use safe, lawful and accountable: what tools are approved, what data they may touch, who owns decisions, how output is reviewed, and how it is all monitored. Good governance is designed into deployments from the start, not added after an incident.

Shadow AI

Staff using unapproved AI tools with organisational data — pasting client information into free chatbots because no sanctioned alternative exists. The realistic fix combines a clear use policy with a governed tool that is genuinely better than the shadow option, removing the incentive rather than just the permission.

AI compliance

Meeting the legal and regulatory obligations that apply to AI use — in New Zealand, chiefly the Privacy Act 2020, sector codes like the Health Information Privacy Code, and public-sector frameworks; overseas, regimes like the EU AI Act. Compliance attaches to how a system is deployed and governed, not to a product label — no tool is "compliant" in the abstract.

Vendor lock-in

Dependence on a supplier that makes leaving impractical — workflows built on a platform whose pricing, models or terms can change under you. Private AI reduces lock-in structurally: the organisation owns the deployment, the knowledge base and the data, and models can be swapped as better ones emerge.

Ready to talk it through?

Book a free discovery call. No preparation required — just tell us what you’re trying to solve.

Book a free discovery callOther ways to reach us