## Column classification (23 columns in `raw.AI_Usage_and_Impact_on_Students_and_Professionals`) These classifications come from the column profiles I pulled (type, semantic role, distinct counts, ranges). I did not run SQL to check the value formats. ### Identifiers - **`User_ID`** is VARCHAR with ~287 distinct values across 300 rows, ranging U0001 to U0288. It is tagged a "near key". It is not perfectly unique, so some IDs repeat, and I did not check whether those rows are true duplicates. ### Categorical dimensions - **Low cardinality:** - `User_Type` (2 values: Student or Professional) - `Would_Recommend` (~4 values, including Yes) - `AI_Purpose` (~7) - **Medium cardinality:** - `AI_Tool` (~16) - `Education_Level` (~17) - `Gender` (~12). That is far more than expected, which suggests inconsistent spellings or casing (e.g. an "other" value). - **Geographic or occupational:** - `Country` (~20) is tagged spatial. - `Profession` (~110) is also tagged spatial, but it is really a high-cardinality categorical label, not a place. - **Free text:** `Tasks_Performed` has ~288 distinct values across 300 rows. It is a descriptive text field, not a dimension to group by. ### Numeric metrics - **Typed as numbers:** - `Productivity_Score` (DOUBLE, 1.6–10.0, 16 nulls) - `Satisfaction_Score` (BIGINT, 3–10, 7 nulls) - `Accuracy_Rating` (BIGINT, 2–5, 15 nulls). With only ~4 distinct values this behaves like an ordinal scale. - **Numeric in meaning but stored as text (VARCHAR), so they need cleaning before use:** - `Age` - `Monthly_Income`. This is personal data, so aggregate it and never quote individual values. - `Monthly_AI_Cost` - `AI_Usage_Hours_Per_Day`. Some values carry a unit suffix such as "9.9 hrs". - `Time_Saved_Hours_Per_Week` - `Work_or_Study_Hours_Per_Day` ### Dates and times - **There is no business date in the data.** Nothing records when a survey response was given, so trend-over-time analysis isn't possible. - The only temporal column is **`_ingestion_timestamp`** (TIMESTAMP). It has a single distinct value, 2026-10-02 09:58:27, because the whole file was loaded in one batch. It describes the load, not the users. ### Pipeline metadata (not analytical) - `_batch_id`, `_source_file` and `_source_system` each have one distinct value and can be ignored. ### Practical takeaway Only three metrics (`Productivity_Score`, `Satisfaction_Score`, `Accuracy_Rating`) are properly typed. Six other numeric fields are stored as text and need `CAST` after stripping units or symbols. The three typed metrics also have missing values.