Data profiling answers one question before anyone builds on a dataset: what is actually in it? Teams run it at the start of a migration, an integration, or a quality program.
A profile reports, column by column:
- the data type and format, and how many values are missing or distinct
- the minimum, maximum, and distribution of numeric fields
- the most frequent values of categorical fields, and the patterns in text, such as identifiers that follow two different formats
- across columns, the candidate keys, the fields that join two tables, and dependencies between values
The output becomes the basis for data quality rules. A profile shows that 24% of an income column is empty; a rule then decides whether that is acceptable.
For AI work, data profiling is the first step of readiness and not the last. A profile describes the data. It does not say whether the data fits the task. A table can profile cleanly, with consistent formats and no duplicates, and still be missing the one exception a model has to handle, or carry a field whose values mean something different from what the task assumes.
So the profile has to be read against the task: which fields the task depends on, which records are exceptions, and what result counts as correct. In Syntitan, Ready for Raw Data begins with that kind of diagnosis across six axes (usability, integrity, context, consistency, reproducibility, and traceability). It extends a conventional profile with the context and traceability an AI task needs, and records the preparation steps applied as a project Version. AI readiness assessment describes how the six axes are read, and data quality dimensions covers the measures a profile usually reports.
How is data profiling different from an AI readiness assessment?
A profile measures the data as it is. An AI readiness assessment judges those measurements against one AI task. The profile reports that a column is a quarter empty; the assessment asks whether that column is one the task depends on, and whether the empty rows are the exact cases the model has to handle.
- Profiling: types, ranges, missing and duplicate values, keys and dependencies, with no task in mind.
- Readiness assessment: usability, integrity, context, consistency, reproducibility, and traceability, read for one use case.
- Relationship: a profile is an input to the assessment, and a clean profile does not pass a task on its own.
Where does data profiling break in AI work?
A lending team profiles a loan table before training a default model. Every column has the right type, duplicates are gone, and missing values sit under a tolerable share. The model still fails in review, because the income field mixes monthly and annual figures from two source systems, and the profile reported one clean numeric range.
The profile did its job; the task check was missing. After profiling, read each field the decision depends on against the task: what the value means, which unit it carries, which records are exceptions, and what result counts as correct. In Syntitan that reading is the Ready for Raw Data diagnosis, repeated when the next extract arrives.