Analysing HR data with AI

A safe workflow

1. Start with a question. “Why is warehouse turnover high, and what would reduce it?” is a question. “Analyse this data” is not.
2. Prepare the data. Remove names and identifiers. Aggregate where possible. Remove or merge groups below your minimum size. Confirm the tool is approved for this data.
3. Orient the tool. Upload the file and ask the assistant to describe what it sees: columns, row count, missing values, anything odd. This catches import problems early.
4. Analyse. Ask specific questions. Ask it to show the calculation.
5. Verify. Recalculate the key numbers yourself in a spreadsheet. Spot-check categories and quotations against the raw data.
6. Interpret. Decide what it means with people who know the context. The AI does not know that the warehouse changed its shift pattern in March.
7. Communicate. Lessons 4.3 to 4.5.

A language model asked to add up a column “in its head” can get the answer wrong, because it is generating likely-looking text. Tools with a data-analysis feature avoid this by writing and running code on your file. ChatGPT’s data analysis feature, for example, runs Python on uploaded spreadsheets and lets you inspect the code. Use such a feature for any calculation, ask to see the method, and verify the important numbers independently.

I've uploaded a CSV of engagement survey results by department. There is no personal data.

1. Describe the file: columns, number of rows, any missing or unusual values.
2. Using code, calculate for each department: response rate, turnover rate (exits / average headcount), and favourable engagement.
3. Calculate the company-wide turnover rate from totals, not by averaging the department rates. Show the formula you used.
4. Present the results in one table, sorted by turnover rate, highest first.
5. List three observations and, for each, say how confident we can be given the group sizes and response rates. Do not claim causes.

Analysing open-text comments Grouping comments into themes is called thematic analysis or coding. It is slow by hand and a good fit for AI, provided you check the result.

Below are anonymous comments from warehouse employees answering "What one thing would make Brightwater a better place to work?"

1. Propose 4 to 6 themes that cover the comments. Give each a short name and a one- sentence definition.
2. Assign each comment to its single best-fitting theme. Show a table: comment number | theme | a note if it could fit a second theme.
3. Count the comments per theme.
4. For each theme, give one representative quotation, copied exactly, with its comment number.
5. Do not invent, merge or paraphrase quotations. Do not guess who wrote any comment.

[PASTE NUMBERED COMMENTS]

Then verify: check that every quotation appears word for word in the data; read a sample of the assignments (at least 10%, or all of them if the set is small) and note where you disagree; adjust the theme definitions and re-run if needed.

Typical failures and how to prevent them