Free preview. This lesson is part of a paid course that earns the ETA Specialist: Data Analysis with AI credential. About this course
Preparing data so AI reads it right
In this lesson you'll
- Shape a spreadsheet so AI tools read it correctly
- Remove or mask sensitive fields before analysis
- Keep an untouched original and a record of every change
Welcome
Welcome to Data Analysis with AI from the Enterprise Technology Association (ETA). AI assistants can read spreadsheets, write formulas, run code and draw charts in seconds. Most of the errors they make come from two places: messy data and vague questions. This module fixes both.
- Step 1: Assess
Is this a good task for AI, and what's at stake?
- Step 2: Direct
Brief the AI clearly: goal, context, constraints, done criteria.
- Step 3: Assure
Check the output before you rely on it.
- Step 4: Protect
Keep people, data and your organization safe.
- Step 5: Tune
Iterate, then save what works for next time.
Use the steps in order on every task, and loop back to Tune when results fall short.
In ADAPT terms, this course is mostly about Assure (checking the numbers) and Protect (keeping data safe), with Direct (the analytical brief) in between.
Shape the data
AI tools read tidy tables best. Before you upload:
| Do | Avoid |
|---|---|
| One table per sheet, starting in cell A1 | Several tables stacked on one sheet |
| One header row with clear, unique names | Merged headers, blank header cells, two header rows |
| One kind of value per column (dates, numbers or text) | "12 (est.)" or "N/A" in a number column |
| One row per record | Subtotal rows mixed in with data rows |
| Consistent labels ("Ohio," not "OH," "ohio" and "Ohio ") | Colors or bold carrying meaning |
Then ask the AI to describe the data back to you before any analysis: "How many rows and columns? What type is each column? How many blanks in each? Any values that look inconsistent?" Compare its answer to what you know. If it says 900 rows and you expected 1,200, stop and find out why.
Protect: strip what you don't need
Passwords, access codes, security keys and other credentials
- Personal information
Social Security numbers, birth dates, home addresses, personal phone numbers
- Health information
Diagnoses, medical leave details, patient records
- Financial account details
Bank and card numbers, individual pay records
- Student records
Grades, discipline and other student information
- Confidential business and client information
Unreleased results, pricing, contracts, legal matters, anything under a confidentiality agreement
Some organizations approve specific tools for some of this data. That decision belongs to your organization, not to you. If you're unsure, treat it as off-limits and ask. You rarely need real details: redact, substitute or generalize.
Ask: does this analysis need people's names, IDs, contact details or account numbers? Usually not. Delete those columns, or replace them with a code, before uploading. Check your organization's policy for which data can go into which AI tools; health, student, financial and HR data often has extra rules.
Keep the original
Save an untouched copy of the source file. Do all cleaning on a copy, and keep a short change log: what you changed, why and how many rows it touched. When someone asks "why does your total differ from finance's?", the log answers it.
Workplace example: a county health department
An epidemiology assistant needed monthly counts of food safety inspections by outcome. The export had two header rows, "Pass w/ conditions" spelled four ways, and inspector names. She deleted the names, merged the headers into one row, standardized the outcome labels, logged each change, and only then asked the AI tool for counts. The counts matched the state system on the first try.
Try it
Ask your approved AI assistant to create a messy sample to practice on:
Create a sample table of 40 rows of made-up volunteer hours with columns: Name, Site, Date, Hours. Include realistic problems: inconsistent site names, a few blank hours, one duplicate row, and one row where hours is "4 (est.)".Then ask it to describe the data (rows, types, blanks, inconsistencies), fix the issues one at a time, and keep a change log. Remove the Name column before any analysis.
Key takeaways
- Tidy tables (one header row, one value type per column, one row per record) prevent most AI analysis errors.
- Ask the AI to describe the data first and compare it to what you expect.
- Remove fields you don't need, keep an untouched original and log every change.