Resources · Glossary
AI governance, defined plainly.
The terms that come up when enterprises connect their data to AI: what each one means, and why it matters for keeping data inside the boundary it belongs to.
Governance & access
Who may do what, and how it's enforced.
| Term | Definition |
|---|---|
| AI governance | The policies, controls and evidence that determine which people, agents and models may use which data and tools, and how that use is recorded and reviewed. |
| Policy decision point (PDP) | The component that evaluates an access request against policy and returns a decision. In Data.co Control it is deterministic and runs inside the customer's infrastructure. |
| Policy enforcement point (PEP) | Where a decision is applied before data moves: an AI gateway, an MCP gateway or a connector. One decision point can serve many enforcement points. |
| Deterministic policy | A rule set that always returns the same decision for the same request and policies. No language model participates in the decision. |
| Fail closed | If the policy service is unavailable or a policy is invalid, requests are denied rather than allowed. |
| RBAC / ABAC | Role-based access control grants permissions by role; attribute-based access control evaluates attributes of the subject, resource and context, such as classification, environment or purpose. |
| Human approval workflow | A policy outcome that pauses a risky request, such as a destructive database operation, until a named person or group approves it. |
| Access simulator | A tool that evaluates hypothetical requests against a policy set and shows the decision, the matching policy IDs and the explanation, before anything changes in production. |
| Audit log | An append-only record of requests, decisions, approvals and administrative actions, exportable as evidence for auditors. |
Identities & agents
Every actor gets an identity.
| Term | Definition |
|---|---|
| AI identity | An identity assigned to a model or AI agent so its requests can be evaluated and audited separately from the human it acts for. |
| Agent scope | The ceiling on what an agent may ever touch — specific datasets, tools and actions — regardless of who is using it. |
| Service account | A non-human identity for an application or automation, with its own scoped credentials. |
| MCP (Model Context Protocol) | An open protocol for connecting AI applications to tools and data sources through MCP servers. |
| MCP gateway | An enforcement point in front of MCP servers that identifies the caller, checks policy for each tool call and records the result. |
| Prompt injection | Instructions hidden in content a model reads — a retrieved document, a ticket, a web page — that try to change its behavior. Deterministic access control limits the damage by keeping unauthorized data out of context entirely. |
Data
Knowing what data you have and where it goes.
| Term | Definition |
|---|---|
| Data classification | Labels such as PHI, PII, financial or privileged applied to fields and documents at ingest, so policies can refer to sensitivity rather than to individual tables. |
| Permission-aware retrieval | Retrieval that evaluates the source system's permissions at query time, so an AI answer never includes content the requester could not open directly. |
| Data lineage | The trace from an AI answer back to the source records, versions and sync jobs that produced it. |
| Data loss prevention (DLP) | Controls that stop classified data from reaching unapproved destinations, including external model endpoints. |
| Data sovereignty | Keeping data, processing and control within a specific jurisdiction or organizational boundary. |
| Change data capture (CDC) | Synchronizing changes from a source system as they happen rather than by periodic full copies. |
| Provenance | Where a record came from: provider, source identifier, canonical URL, version and license. Derived data, such as AI summaries, is marked as derived. |
Private AI
Running models on infrastructure you control.
| Term | Definition |
|---|---|
| Private LLM | A language model, typically open-weight, served on infrastructure the organization controls rather than through a third-party API. |
| Open-weight model | A model whose trained weights are published so it can be run locally, under the model's own license terms. |
| Inference server | Software that serves model predictions, such as vLLM for GPU servers or Ollama for smaller deployments. |
| AI gateway | A single endpoint — often OpenAI-compatible — in front of private and external models that adds identity, routing, quotas and policy enforcement. |
| RAG (retrieval-augmented generation) | Answering with a model after retrieving relevant documents, so responses are grounded in the organization's own sources. |
| Vector index | A store of embeddings used for semantic search; pgvector in PostgreSQL is a common self-hosted option. |
| Air-gapped deployment | An environment with no connection to the internet, which requires offline licensing, local updates and no external telemetry. |
See it against your own use case.
A 30-minute technical walkthrough, policy simulation on your scenarios, and a deployment recommendation for on-prem, private cloud, or hybrid.