← All use cases
Use case 09

Remuneration health check

Load a workforce file and get a quick snapshot of how healthy pay is, for the whole organisation or any area of it. Six checks, the same every time, each with a clear status.

Vibe-coded tool No AI needed to run it Works offline
Try itUse the real tool. Watch itFive steps, animated. Build itPrompts you can paste. Check the dataWhat is safe, and where. Take itDownload the finished file.

What it does and why it matters

Most reward teams can answer "how is pay in Technology?" only after days of spreadsheet work. So the question gets asked once a year. A health check that runs in seconds can be run for any area, at any time, and before a problem turns into resignations or a pay gap figure.

It runs six checks: where people sit against their reference point; pay compression (differences too small to reflect seniority or experience), across managers and their reports, newer and experienced people, and the steps between levels; whether pay follows performance ratings; how widely pay spreads within the same job; which people are paid far from their peers; and a gender snapshot.

A status is a prompt to look, not a verdict. The thresholds behind Healthy, Watch and Act are examples that you change to suit your organisation.

This is not use case 06 (the remuneration analyser). This one runs fixed checks as ordinary code on your own computer. That one answers free-form questions and is built with IT. For the gender detail, use the pay gap analyser (use case 01).

1Try it

This is the real tool, running in your browser on an invented workforce. Pick an area, change a threshold, or load your own file.

Your file is read by your own browser. Nothing is uploaded. All sample data is invented.

Try this

  1. Read the six statuses for the whole organisation.
  2. Switch the department to Technology and compare it with the whole organisation.
  3. Open the thresholds and change one. Watch a status move.
  4. Open "People behind the flags" and filter by one reason.
  5. Copy the summary prompt and read what is, and is not, in it.

2Watch it

Five steps from a workforce file to a shareable picture. Press Play, or step through at your own pace.

Step 1 of 5Invented data
Choose an area
DepartmentTechnology
LevelAll
Job familyAll

Technology · 80 people

Whole organisation · 320 people

Load the file, pick an area. One row per person, with an ID, not a name. Choose a department, a level, a job family, or leave them all on "All".

Whole organisation: 3 healthy, 2 watch, 1 act
Position against the reference point✓Healthy
Pay compression!Watch
Pay and performance✓Healthy
Spread within the same job✓Healthy
Outliers!Watch
Gender snapshot✕Act

No overall score. Each status has an icon and a word, so it never depends on colour alone.

Six checks, one status each. Every check ends in Healthy, Watch or Act. Thresholds are examples that you replace with your own.

Share of people outside 80% to 120% of their reference point
Technology
11%
Whole organisation
3%
Status
✕ActTechnology ✓HealthyWhole organisation

Compare with the whole organisation. The same figure sits beside the whole organisation's, so "is this unusual?" is answered on screen.

Managers and the people they manage
ManagerManager's baseHighest-paid reportRatio
Manager A, Level 4$118,000$116,50098.7% !Compressed
Manager B, Level 4$104,000$109,500105.3% ✕Inverted

Compressed: the best-paid report earns 95% or more, but less than 100%, of the manager's base. Inverted: 100% or more.

See what is behind a flag. The "People behind the flags" list shows who, and why, for your own working use. It sits closed until you open it.

One-page summary (printed)
  • Area and headcount
  • Date
  • Six statuses and one figure each
  • Whole organisation beside it
  • Thresholds used
Grouped figures only
People list (working data)
Not in the printed summary

Share the picture, keep the people private. The printed summary and the AI summary prompt carry grouped figures only. No employee ID appears in either.

All figures in this walkthrough are invented, for illustration only.

3Build it

You can build your own version with any capable AI assistant your employer allows, for example Microsoft Copilot, Claude or ChatGPT. You describe what you want, it writes the code, and you check the answers.

Build it with invented data, never your real file. Expect two or three rounds of "that's not quite right, change X".

Step 1: Describe the tool

Say what the tool is, what goes in, and what comes out.

I work in reward (remuneration) in Australia and I am not a developer. Please build me a remuneration health check as a single HTML file that I can save to my computer and open in a browser by double-clicking it. Requirements: - It must run offline. Use no external libraries, fonts, images or links, and send no data anywhere. Save nothing. - The user loads a CSV file with one row per employee. Required columns: employee_id, gender, department, level, job_family, base_salary, reference_point. Optional columns, each of which switches on a check: rating, manager_id, years_in_role. - The user chooses an area using three dropdowns (department, level, job family), each defaulting to "All". The checks run on the people in that area, and the same figures for the whole organisation are shown beside them. - It runs six checks, the same every time, and each ends in one status: Healthy, Watch or Act. If a check needs a column that is missing, it says "Not run" and does not fail. There is no overall score. - Build in an invented sample workforce of about 300 people, so the tool shows results as soon as it opens, with a button to download the sample as a CSV. - The audience is reward practitioners, not developers. Use plain English and Australian spelling. Show each status with an icon and a word, never colour alone. Before you write any code, tell me in a few sentences what you plan to build, and ask me about anything that is unclear.

Step 2: Define every check exactly

This is where vibe coding goes wrong. If you write "check for pay compression", the AI will invent its own definition and you will not know which one. Give it the full definition of every check, then adapt these to your own policy. Compa-ratio means a person's pay divided by the reference point for their role.

Now add the six checks. Follow these definitions exactly. Do not substitute your own. Compa-ratio (CR) = base_salary divided by reference_point. 1. Position against the reference point. Show headcount in six bands of CR: below 80%, 80 to 90, 90 to 100, 100 to 110, 110 to 120, above 120 (a CR exactly on a boundary belongs to the higher band). Show the median CR and the share of people below 80% or above 120% (exactly 80 or 120 counts as inside). Status: Healthy if that share is at most 5%, Watch if at most 10%, otherwise Act. 2. Pay compression (pay differences too small to reflect seniority or experience), in three parts. The status is the worst of the parts that ran. 2a. Managers and the people they manage (needs manager_id). A manager is anyone with at least one direct report anywhere in the file. ratio = highest base_salary among their direct reports divided by the manager's base_salary. Compressed if the ratio is at least 0.95 and below 1.00. Inverted if it is 1.00 or more. Status from the share of managers who are compressed or inverted: Healthy if at most 5%, Watch if at most 15%, otherwise Act. 2b. Newer and more experienced people (needs years_in_role). Newer = under 2 years in role. Difference = median CR of newer people minus median CR of the rest, in percentage points. Healthy if the difference is at most 0, Watch if at most 3, otherwise Act. If either group has fewer than the minimum group size, do not run it. 2c. Steps between levels. For each pair of neighbouring levels that both have at least the minimum group size: step = median base of the higher level divided by median base of the lower level, minus 1. Status from the smallest step: Healthy if at least 10%, Watch if at least 5%, otherwise Act. 3. Pay and performance (needs rating). For each rating, lowest to highest: headcount and average CR. Spread = average CR of the highest rating minus average CR of the lowest, in points. Act if the spread is below 0. Watch if it is below 5 points, or if any higher rating has an average CR more than 0.5 points below the rating beneath it. Otherwise Healthy. 4. Spread within the same job. A job group is the same level and job family, with at least the minimum group size of people. Spread = (90th percentile of base minus 10th percentile) divided by the median base, using linear interpolation between closest ranks (position = p x (n minus 1)). A group is wide if its spread is above 30%. Status from the share of job groups that are wide: Healthy if at most 10%, Watch if at most 25%, otherwise Act. 5. Outliers. Within each job group, a person is an outlier if their base_salary is more than 20% above or below the group's median. Status from the share of people in job groups who are outliers: Healthy if at most 3%, Watch if at most 8%, otherwise Act. 6. Gender snapshot. The average gap = (men's average base minus women's average base) divided by men's average, times 100. The like-for-like gap = the headcount-weighted average of the gaps within each level and job family group. Status from the size of the like-for-like gap, in either direction: Healthy if at most 1%, Watch if at most 3%, otherwise Act. Show a figure only when there are at least the minimum group size of women and of men. Write each check as a separate function that only does the calculation, so I can test it without the screen.

Step 3: Make the thresholds yours

The numbers above are invented examples. Your organisation will have its own view of what is "healthy".

Now put every threshold in a collapsed "Thresholds" panel. Give each one a number input and a plain-English label, and add a "Reset to examples" button. Head the panel "Example thresholds. Replace them with your own." Everything on screen, in the printed summary and in any text the tool generates must call the thresholds examples. When I change a threshold, every status must update straight away.

Agree the thresholds with your team before anyone outside it sees a red status. If a number appears on screen as "Act" and nobody has agreed what that means, the conversation becomes about the label and not about the pay.

Step 4: Protect small groups and individuals

Hiding a small group is not enough on its own. If a total is shown and every other group is shown, anyone can subtract the others from the total and work out the hidden one. So when a small group is hidden, one more group may need hiding too.

Now add privacy protections. - Add a "Minimum group size" setting, default 5. Hide the figures for any group smaller than that, and do not run a check when an area has fewer than twice that many people. Say "This area is too small to report on without pointing to individuals." - A hidden figure must not be recoverable by subtracting. If hiding one small group would let someone work it out from the totals and the other groups, hide one more group so that the hidden groups together are at least the minimum size. Tell the user when this has happened. - Put the people behind each flag in a section that is closed by default, with a line above it saying the list names individuals by ID. Never include that list in the printed summary. - Make a "Print summary" button with a print style sheet that prints one page: the area, headcount, date, the six statuses with one headline figure each, the whole organisation beside it, and the thresholds used. No controls and no people. - Make a "Copy a summary prompt" button that builds a prompt for an AI assistant using grouped figures and statuses only. It must never contain an employee_id or any figure for a hidden group. It must tell the AI: "Use only the figures below. Do not invent causes. If a figure is not given, say so." - Do not save anything. Do not use cookies, local storage or any network request.

Step 5: Test it against cases you can check by hand

Ask the AI for a tiny case with round numbers, work the answer out yourself, then compare. Here is one for check 4 (spread within the same job) and check 5 (outliers).

Please help me check that the calculations are right. 1. Make a tiny test CSV with one job group of exactly five people, all in the same level and job family, with base salaries of $80,000, $90,000, $100,000, $110,000 and $150,000. 2. Add a "Show your working" panel that lists, for that group: the median, the 10th percentile, the 90th percentile, the spread, and which people are outliers. 3. Before I load the file, tell me the answers you expect, so I can compare them with what the tool shows. 4. If any result does not match my hand calculation, say so plainly and fix the calculation. Do not change the test data to make it fit. Then make similar tiny test files for compression (a manager and three direct reports) and for pay and performance (three ratings, two people in each).

A check you can do by hand

Invented example
PersonBase salary
1$80,000
2$90,000
3$100,000
4$110,000
5$150,000
Median$100,000
10th percentile$84,000
90th percentile$134,000
Spread: (134,000 − 84,000) ÷ 100,000. Above 30%, so "Wide".50%
Outliers, more than 20% from the median: $150,000 only1 of 5

$150,000 is 50% above the median. $80,000 is exactly 20% below it, which is not more than 20%, so it is not an outlier. If the tool disagrees, it is wrong. Tell the AI what you expected and ask it to fix the calculation.

What you would add next

  • Trend over time. That needs a tool that remembers each run, which this one deliberately does not. See the review tracker for reward teams for a tool that keeps its data.
  • Market position, if your survey licence allows you to load survey data into your own tool. Check the licence first.

Good habits

  • Write down the definition of every check before you ask for code, and keep it with the tool.
  • Test with a file where you already know the answer.
  • Agree the thresholds, and what "Act" means, before anyone sees a red status.
  • Get someone else to check the numbers before anything goes to an executive or a board.

4Check the data

What this touches

Individual pay, ratings, reporting lines and time in role for everyone in the file.

What is safe

The tool runs on your own computer, saves nothing and is technically blocked from sending anything. Use IDs, not names. The printed summary and the summary prompt carry grouped figures only. The people list is your working data, so treat a downloaded copy like any confidential pay file.

Ask before you start

  • Who is allowed to see a health check for an area?
  • Who agreed the thresholds?
  • What happens when a status is red: who is told and who acts?
  • Which AI assistant, if any, may receive the grouped summary?

Privacy rules for employee records differ between private employers and government agencies, so check with your privacy officer. Start with the OAIC's guidance on the Australian Privacy Principles and on employment and privacy. General information, not legal advice.

A status is a prompt, not a verdict

  • Check the data before you check the people. A missing reference point or a wrong level can turn a healthy area red.
  • A red status in a small area can come from two or three individuals.
  • Agree what "Act" means before the first red appears.

5Take it

Download the tool Open it in a new tab

It is a single file. Save it anywhere and double-click to open it. The sample file can be downloaded from inside the tool.

What next: Pay gap and pay equity analyser (use case 01) · Remuneration analyser (use case 06) · Back to all use cases