Choose one of three approaches. The steps for that option appear below.
Choose an option above. On a phone, swipe the tabs left or right to see all options →
Gemini proposes the profile; you recompute figures, record column decisions and write the report.
Collect the dataset and transformation
Draft the per-column profile
Check figures and record column decisions
Use: You
Have ready: dataset extract and a note documenting the transformation used for profiling.
Permit only the reduced extract into Gemini.
Use: Gemini
Import the extract into Gemini; ask for explicit counting definitions.
What to paste into the tool
Task: Candidate per-column profile and calculation notes. Use the reduced extract only; report type, null percentage with row-count denominator, distinct-value count, range and frequent values with counts for each column; distinguish blank from zero; state whether analysis uses all rows or a sample; flag unsupported calculations rather than fabricate numbers
Use: You
Recompute null rates and distinct counts for three columns; check coverage, ranges and frequent counts. Record fix, ask owner or accept.
Explain choices without changing the dataset. Retain the report without a fitness-for-use decision; keep transformation limitations and unanswered decisions. Hold the candidate profile for analyst audit. Advance only the analyst-audited report content.
Hex drafts code from schema metadata; you inspect and run it locally, then write the report.
Write the schema-only worksheet
Draft profiling code without running
Inspect code, run locally
Use: You
Have ready: local dataset, schema-only worksheet with column IDs, intended types and blank conventions; no data rows.
Authorise only schema metadata for the notebook agent.
Use: Hex
Ask Hex for local profiling code and calculation definitions in an empty project without connecting a warehouse or running cells.
What to paste into the tool
Task: Proposed profiling code and calculation definitions. Use only the schema; draft local Python code to read an analyst-chosen extract and calculate type, null rate, distinct values, range and frequent values per column; define denominator and blank handling; include mixed-type handling; return code and definitions without running cells
Use: You
Inspect file access and counting logic; adapt and run locally. Recompute null rates and distinct counts for three columns; write column decisions.
Allow only inspected local code to process the retained extract. Record failed calculations rather than substituting guessed values. Check coverage, mixed types and ranges; record fix, ask owner or accept.
You compute statistics locally; ChatGPT suggests concerns and drafts the report from aggregates and your recorded decisions.
Compute and check the local profile
List candidate column questions
Check candidate questions
Draft the profile report
Check every figure and finish
Use: You
Have ready: local dataset. Compute types, null percentage, distinct counts, ranges and frequent values; document blanks and recompute three columns.
Check every profile row against the local extract. Keep the source extract within the analyst environment. Send only the approved aggregate worksheet to ChatGPT.
Use: ChatGPT
Upload only the aggregate worksheet; ask ChatGPT for candidate decision columns using row IDs without changing statistics.
What to paste into the tool
How to use: upload the aggregate worksheet as a CSV file, then paste this request. Task: Candidate column questions from aggregates. Use only the aggregate profile in the uploaded file; list questions about mixed types, missingness and unusual ranges; quote only supplied non-identifying aggregates; do not invent underlying rows or decide fitness for use; preserve each profile row ID. Return the questions as a numbered list with the profile row ID each refers to.
Use: You
Compare questions with the full local profile, add omissions, discard unsupported concerns; record fix, ask owner or accept.
Conclude without applying a dataset correction. Keep candidate questions separate from analyst decisions.
Use: ChatGPT
Ask ChatGPT to draft the report from only the aggregate worksheet and your recorded decisions, preserving all statistics.
What to paste into the tool
Task: Draft profile report. Assemble the profile report using only the uploaded aggregate worksheet and my recorded decisions: per-column type, null rate, distinct values, range and top values, plus the list of columns that need a decision with my recorded fix, ask owner or accept. State the redaction limits. Do not change any statistic and do not invent underlying rows.
Use: You
Compare the report with the full local profile and decisions; fix mismatches from local figures before release.
Every figure in the report matches the local profile and every listed column carries your recorded decision before you release it.
Best when: You want spreadsheet profiles with a separate human metric audit.
What it does: Gemini profiles; you recompute metrics and record decisions
Strong at: Shows per-column profiles with calculation definitions.
Limit: Candidate calculations need independent checking.
Free plan: Yes
Best when: You want schema-based code proposals without uploading the dataset.
What it does: Hex writes code; you inspect, run and author findings
Strong at: Drafts Python profiling code and explicit calculation definitions in a notebook.
Limit: You must inspect, adapt and run the code locally.
Free plan: Yes, limited AI credits
Best when: You want local statistics with AI-assisted concern and report drafting.
What it does: ChatGPT drafts questions and report; you compute and decide
Strong at: Turns supplied aggregates and recorded decisions into a readable profile report.
Limit: It cannot invent underlying rows or decide fitness for use.
Free plan: Yes, limited uploads and data analysis
Before you paste into any AI tool, remove any confidential information. Put it back at the end.
Deciding between ChatGPT and Gemini for this? See the comparison.
Each column receives the requested profile fields or an explicit missing-calculation note.
Null and distinct checks are repeated locally for three columns.
Concerns remain separate from a decision that the dataset is fit for a downstream purpose.
Check the result against your original inputs. Approve it before sharing or acting.