Glossary

Agent — An AI model that carries out a task with tools: reading metadata, checking quality, measuring a proposal. An agent runs in an isolated container and never gets a model vendor key or a database password.

AI gateway — The point every request to an AI model passes through, with exactly one policy decision per request. See The AI gateway.

Cohort threshold — Counts below ten become "<10", so no individual can be inferred from a count.

Connector — A link to a system that is not a database, such as Power BI, Azure Data Factory, SharePoint or Databricks. See Sources and connectors.

Customer — The shielded space where sources, rules, knowledge and permissions come together. Can also be a department.

Customer profile — A layer between platform and customer, for skills and knowledge that apply to a kind of customer.

Data access — The mode per source that determines what an AI model sees: metrics only, encrypted values or open. See Data access.

Dataset — A reusable definition of data for dashboards: one query on one source, or a derived dataset that combines other datasets. See Datasets.

Encrypted values — The data access mode in which every cell is a stable token. The model can count and join without seeing the value.

Environment — Production, acceptance, test or development. Every source belongs to an environment.

Finding — What is created when a monitor sees a deviation. At most one open finding per monitor.

First responder — The part of the platform that answers simple questions in a conversation directly and hands heavier work to an agent itself.

Ground rules — The agreements that apply to all tasks for a customer: models, budget, data access, guidelines. Every change is a new version. See Ground rules.

Knowledge document — Background information that goes along with an agent, such as a data dictionary. See Skills and knowledge.

Landscape — The map of an environment: systems, databases, pipelines, datasets, reports and their connections. See Environments.

MCP — Model Context Protocol, the open standard that lets an AI tool use tools from a server. With MCP you connect your own AI tool to the platform. See Connecting via MCP.

Metrics only — The strictest data access mode: structure, counts and statistics, no rows. The default.

Monitor — A scheduled check by an agent that gives a verdict. See Monitors.

Open — The data access mode in which literal values go to the model. Only for sources the customer designates.

Proposal — A change an agent proposes that is only carried out after a human approves it, for example a script on a source or a new skill.

Pseudonym — The replacement name under which a table, column or value goes to the model, such as t_0056. See Pseudonymisation.

Sandbox — A disposable copy of the database in which an agent runs and measures its proposal. See The verification sandbox.

Skill — A work instruction that goes along with an agent. See Skills and knowledge.

Solo — The way of working in which all work happens in your own AI tool and the platform only supplies tools, rules and storage. See Solo.

Source — A customer database or system the platform works on, for example an Azure SQL database, a PostgreSQL server or a Power BI workspace.

Task — A piece of work on the platform, with a goal, a work mode, allowed sources and a history. A conversation is a task too.

Work mode — What an agent may do in a task: look only, measure on the source, or prove on a disposable copy. See Work modes.