Glossary
Data and AI terms in plain words
The terms you will hear most often when a group puts its data in order, one short definition each.
- AI agent
- An AI model that can use tools, such as running a query or reading a document, and take steps toward a goal.
- Attribution
- Deciding which marketing touch gets the credit for a sale.
- Churn
- Customers who stop buying or cancel. A churn model scores each customer’s risk of leaving.
- Conformed dimension
- A shared definition of customer, date, location or product used the same way across companies. It makes companies comparable.
- Cross-border data transfer
- Sending data to servers or recipients in another country. Many laws allow it only with consent or to countries with adequate protection.
- Data contract
- An agreement with a source system’s owner on what the data looks like and when it arrives.
- Data localization
- A legal requirement to store certain data, usually personal data, on servers inside a country.
- Data processing agreement
- A contract that sets where a client’s data is stored, who may access it and when it is deleted.
- Data quality checks
- Automated tests for missing values, duplicates, accepted values, freshness and volume that run before anyone opens a dashboard.
- Data warehouse
- One database built for analysis. Data from every source system is loaded into it, cleaned and kept with its history.
- dbt
- A tool for writing data transformations as tested, versioned SQL.
- Driver tree
- A key metric broken into the parts that drive it, for example revenue equals active clients times visits per client times price.
- ELT and ETL
- ELT loads raw data first and transforms it inside the warehouse; it keeps history and makes changes cheap. ETL is the older order: transform, then load.
- Forecast error (MAPE, WAPE)
- How far a forecast was from the actual result, in percent. WAPE weights errors by volume.
- Hallucination
- A confident AI answer that is wrong or made up. Grounding in certified data and showing sources reduce it.
- Large language model (LLM)
- A model trained on large amounts of text that can answer questions and write, such as Claude, GPT or Gemini.
- MCP
- Model Context Protocol, an open standard for connecting tools and data to AI agents.
- Metric dictionary
- A list of key metrics with the definition, formula, level of detail, filters, owner and source of each.
- Orchestration
- Running data jobs in the right order on a schedule, with retries and alerts. Typical tools are Airflow, Dagster and Prefect.
- Personal data
- Any information that identifies a person directly or indirectly, such as a name, a phone number or an ID number.
- Prompt injection
- Hidden instructions in an email, document or web page that try to make an AI misbehave.
- RAG
- Retrieval-augmented generation: before answering, the AI receives the most relevant passages from trusted documents.
- RFM segmentation
- Grouping customers by recency, frequency and monetary value of their purchases.
- Row-level security
- Each user sees only the rows they are allowed to see. For example, a school director sees only their own school.
- Semantic layer
- The place where business metrics are defined once, so dashboards, models and AI agents all use the same logic.
- Single source of truth
- An agreement that each metric has one definition, one owner and one source, so every report shows the same number.
- Star schema
- A data model with a fact table, such as orders or visits, in the middle and descriptive tables, such as customer or product, around it.
- Text-to-SQL
- The AI turns a question into a database query and answers from the result.
- Weekly business review
- A standing weekly meeting on the same pages of metrics: what moved, why, and who does what next.
Start with one question
A pilot takes four weeks, covers one question your leader cares about and has a written measure of success.