← All use cases
Use case 03

Pay recommendation

A step-by-step process that gets to a defensible offer, not a guess. It checks what is missing before it gives any number, and shows its reasoning every time.

Skill Works in Copilot, Claude or ChatGPT A person still decides
Try itReplay three real exchanges. Watch itFive steps, animated. Build itPrompts you can paste. Check the dataWhat is safe, and where. Take itDownload the skill.

What it does and why it matters

Pay recommendations tend to go wrong in three ways. Steps get skipped. The logic changes from one case to the next. And reward is brought in late, then asked to approve a number when an offer is already on the table.

This skill holds the line on process. It checks what is in hand before it says anything about money. If something essential is missing, it stops and lists the gaps. Then it works six steps in a fixed order and shows its reasoning at each one.

A skill is a written set of instructions, saved as a file, that an AI assistant follows every time it does a particular job. It turns “a good first draft, if you ask the right way” into “the same method, every time, whoever is asking”.

For your team, that means the same logic whoever handles the case, and something to point to when a manager pushes.

1Try it

Three real exchanges with an AI assistant using this skill, on an invented case at Banksia Example Co (an invented organisation). Press Play to watch each one, or Show all to read it.

The gate holds

Replay Captured from Claude following this exact skill, November 2026. Invented data.

A full recommendation

Replay Captured from Claude following this exact skill, November 2026. Invented data.

Outside the range

Replay Captured from Claude following this exact skill, November 2026. Invented data.

These are real responses, saved and replayed. Nothing you do here is sent to an AI.

2Watch it

Five steps from a manager's request to a reasoned position. Press Play, or step through at your own pace.

Step 1 of 5Invented data
Gate check: what is in hand?
  • In handJob size
  • In handJob family
  • MissingMarket data for the specific role
  • MissingInternal reference point or range
  • In handPositioning guidelines
  • In handPeers' positions
  • In handCandidate experience
  • In handFigures offered or implied: none
Stop: cannot recommend yet. Missing: market data and the reference point. No figure is given.

The gate. Before any recommendation, the skill checks what is in hand. In this example two items are missing, so it stops and says why they matter. Once they are supplied, the run continues.

Size, family, market role
Job size
Level 4
From your job sizing
Job family
Technology
Placed in a family
Market role
Senior Data Analyst
Specific, not generic

Not “Technology” or “Data” in general. The market match is the specific role.

Size, family, market match. Steps a, b and c fill in one at a time. The skill uses your job sizing result. It does not size the job itself.

Market median, Senior Data Analyst
$128,000median

18 people in the sample, from 11 organisations.

Example thresholds. Replace them with your own.

Market median and how far to trust it. A median is only as good as the sample behind it. Here 18 people lands in the medium band, so confidence is medium.

Range around the reference point (plus or minus 10 percent, an example)
Five current peers (invented): 94%, 97%, 100%, 103%, 106%

Reference point and range. The skill matches the role to your reference point, then draws the range around it. Peers appear as positions, with no names and no salaries.

Where the candidate sits
  • 6 years of relevant experience, against an entry expectation of about 3.
  • Fully meets Level 4, but has not yet led work with senior stakeholders, and has no performance record with us.
  • Sits between the peer with 5 years (103%) and the peer with 8 years (106%).

Position the person, show the reasons. This step is judgement, not a formula. The skill's job is to make the reasons explicit. A person decides.

All figures in this walkthrough are invented, for illustration only. The thresholds and the plus or minus 10 percent range are examples to replace with your own.

3Build it

You can build your own version with any capable AI assistant your employer allows, for example Microsoft Copilot, Claude or ChatGPT. The aim is to write down how your team really makes a pay recommendation, then turn it into a skill.

The method on this page is generic. Your thresholds, ranges and rules will differ, so build with your own process and invented data, never real pay data. Expect a few rounds of “that's not quite right, change X”.

Step 1: Write down your real process

Most teams have never written the process down. Let the AI interview you, one question at a time.

I work in reward in Australia. I want to turn how my team makes a pay recommendation into a written set of instructions (a skill) that an AI assistant can follow every time. Please interview me about our process. Ask one question at a time and wait for my answer before you ask the next one. Start with what triggers a request for a pay recommendation. Then cover: - what we check before we give any number - how we size a job and where that result comes from - how we match a role to the market, and how we judge whether the market data is reliable - how we set a reference point or pay range for a role - how we decide where an individual sits in that range - how we treat internal equity - who approves the final figure, and where the reasoning is recorded Ask me for an example of a case that went well and one that went badly. Do not ask for real names or real pay figures. Use roles and levels only. When the interview is finished, summarise our process as numbered steps in my own words and ask me what is wrong or missing. Do not write the skill yet.

Step 2: Add the gate

The gate is the most valuable part. It stops the AI from guessing when information is missing.

Now add a gate at the start of the skill. Before it gives any recommendation, it must check that all of these are in hand: 1. Role description or job size, from our own job sizing 2. Job family 3. Market data for the specific role: the median, the sample size and the number of organisations 4. Our internal reference point or pay range for the role 5. Our pay positioning guidelines 6. Internal equity: how people in the same role are paid relative to the reference point (a compa-ratio is actual pay divided by the reference point) 7. The candidate's relevant experience and capability 8. Whether any figure has already been offered or implied If anything essential is missing, the skill must STOP. It must not give a figure, a range, or a view on whether a proposed figure is high or low. It must list what is missing, say why each item matters, and say what to send next. It must never fill a gap with its own estimate. Show me the gate wording before you add it to the skill.

Step 3: Add the hard rules

These are the rules you want held every time, however much pressure the person asking is under.

Add these hard rules to the skill, in plain language: 1. Out-of-range rule. If a proposed figure is outside the range, say so plainly with the arithmetic and explain what would be needed to justify it. Going above the range needs an exceptionally strong, written reason. Do NOT suggest a figure just inside the range, because that anchors the decision and could still be too high. 2. A figure that has been implied counts as an offer. If a manager says the candidate has been told to expect around a number, treat that as an offer and declare it in the summary. 3. Never store pay values in the skill. Ask for them every time and do not reuse them between cases. 4. Work with anonymised data by default. Ask for role and level, not names. 5. Always end the summary with a line saying the skill recommends and a person decides. 6. Output format: a short summary on top (recommendation or "cannot recommend yet", confidence, the one biggest risk), then the audit trail for each step in order, then what was assumed or missing. Then show me the full skill and ask whether anything contradicts anything else.

Step 4: Test it with awkward cases

Test with invented cases built to make the skill fail. If it holds up, then trust it a little.

Act as a difficult hiring manager and test the skill I have just built. Use invented data only. Run each case below as a separate conversation. After each one, tell me plainly whether the skill followed its own rules, and quote the part of the skill that applies. Cases: 1. Market data is missing. 2. The market sample is tiny. 3. The manager asks for a figure above the range. 4. The manager is in a hurry and says the candidate has another offer. If the skill gives a figure when it should have stopped, or suggests a figure when a proposed one is outside the range, say so, and suggest a change to the wording of the skill.
  • Missing market data. It should stop and say why it matters, with no figure.
  • Tiny sample size. It should carry on, but rate confidence as very low and say so in the summary.
  • A figure above range. It should say so plainly, list what would be needed, and suggest no figure.
  • A manager in a hurry. It should stay calm, declare any implied offer, and still run the gate first.

Step 5: Put it where your team works

A skill is only useful if people can reach it. Where you put it depends on the tool your employer has approved.

  1. In Claude: add the file as a skill. Claude's help pages explain how to do this for your plan.
  2. In Microsoft 365 Copilot: create an agent in Agent Builder and paste the instructions into it.
  3. In any other assistant: paste the instructions at the start of each conversation, then add the case details.

Microsoft's documentation says agent instructions can be up to 8,000 characters, and a skill file up to 20,000 characters (Microsoft Learn, checked November 2026). We have not yet tested this skill inside Copilot.

Good habits

  • The skill recommends, a person decides. Keep the approver's name and reasoning with the case.
  • Replace the example thresholds and range with your own before you rely on it.
  • Keep a library of awkward test cases and re-run them whenever you change the skill.
  • Give it role and level, not names, and treat any pay data you paste in as sensitive.

4Check the data

What this touches

An individual's proposed pay, peers' pay positions, and market data that is usually licensed. Each one needs care, for different reasons.

What is safe

Work with role and level, not names. Give peer pay as positions against the reference point, not as names and salaries. Do not paste licensed survey data into a tool your licence does not cover. Use tools your employer has approved.

Ask before you start

  • Is this AI tool approved for pay data?
  • Does our survey licence allow this use?
  • Who approves the final figure?
  • Where is the reasoning kept?

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.

5Take it

Download the skill Download the sample case View the captured output

Add SKILL.md to your assistant as a skill or as instructions, then try the sample case before you use a real one. Replace the example thresholds and the plus or minus 10 percent range with your own first.

What next: Back to all use cases