A column called region can hold the customer’s region, the sales rep’s territory, or the warehouse that shipped the order. The values look the same in all three cases. Context is what tells a person, or a model, which one it is.
The evidence for context sits outside the table:
- a data dictionary or catalog entry
- business rules, such as which orders count as revenue
- examples that show how edge cases are recorded
- the field owner’s confirmation of what the field means today
When that information is missing, people fill the gap with assumptions. An AI model or agent does the same, and may not flag the assumption.
How is data context different from technical metadata?
- Technical metadata describes a field’s type, size, source table, and refresh schedule.
- Business metadata, such as glossary definitions and owners, is part of data context. Context for a task also needs the units, relationships, and exceptions the task depends on.
- A catalog can hold both; whether it holds the business side for the fields a task uses has to be checked. Data lineage shows where a field came from; it does not say what the field means.
Where does data context break in AI work?
An agent answering questions about overdue invoices reads due_date. For most customers that is the payment deadline. For customers on extended terms, the finance team tracks the real deadline in a separate field and treats due_date as a placeholder. Nothing in the table says so, and the agent flags extended-terms customers as overdue once the placeholder date passes.
The fix is to capture meaning, units, relationships, and exceptions with the data, and to keep them when the data is prepared, so the task and the model read the same definitions. In Syntitan, context is one of the six axes Ready for Raw Data diagnoses, with data dictionaries, business rules, examples, and owner confirmation as the evidence. Field mapping is one place where the meaning of each field gets confirmed.