Custom GPT: Build a Compensation Policy Assistant

What This Builds

Instead of fielding the same 40 manager questions every merit cycle, you build a custom AI assistant loaded with your company's actual compensation philosophy, pay grade structure, and merit guidelines, then share it with HR business partners and managers. They get instant, accurate answers at 11pm on a Sunday; you get your time back.

Prerequisites

• {{tool:ChatGPT.plan}} subscription ({{tool:ChatGPT.price}}) — Custom GPTs require a paid account
• Your company's compensation philosophy document (or the key points written down)
• Your current salary grade structure (grade bands and midpoints; you can anonymize if needed)
• Current merit cycle guidelines (budget %, eligibility rules, performance rating criteria)
• 30–60 minutes to gather and organize documents; 60–90 minutes to build and test the GPT

The Concept

A Custom GPT is like training a new HR colleague who has read every document you give them and can answer questions about those documents 24/7, without forgetting what they learned or giving inconsistent answers. You set it up once. Every future conversation starts from that shared foundation.

The key difference from using a regular ChatGPT conversation: the Custom GPT's "instructions" and uploaded documents persist. A manager who asks "how do I explain to my team member why they got 3% instead of 6%?" gets an answer grounded in YOUR company's actual philosophy, not generic HR advice.

Build It Step by Step

Part 1: Gather and organize your source documents

Documents to collect (15–20 minutes):

1. Compensation Philosophy: The one-page document explaining your company's approach. If you don't have one, write the key points: What percentile do you target? Do you pay for performance? How are ranges set? What's your philosophy on pay transparency?

2. Pay Grade Structure: A table of grade levels and midpoints. You can anonymize salary amounts if your organization restricts sharing actual numbers; the GPT needs the structure and ranges to answer grade-related questions.

3. Merit Cycle Guidelines: Current year's budget percentage, eligibility rules (who is eligible, minimum months of service), performance rating definitions, merit increase matrix.

4. Common FAQ: Write down the 10–15 questions managers ask you most often during merit season. You'll use these to test the GPT later.

What to exclude: Don't upload individual employee salary data, pending comp decisions, or anything marked confidential beyond the comp team.

Part 2: Create the Custom GPT

1. Go to chatgpt.com → click your profile icon → My GPTs → Create a GPT
2. You'll see a "Create" tab and a "Configure" tab. Start with Configure for more control.

Fill in the Configure fields:

Name: [YourCompany] Compensation Guide (or just "Comp & Benefits Assistant")

Description: Answers compensation and benefits questions using [Company]'s policies, pay philosophy, and merit guidelines. For HR Business Partners and managers during merit planning season.

Instructions (paste this, then customize):

You are the [Company Name] Compensation and Benefits Guide. Your purpose is to help HR Business Partners and managers understand our compensation philosophy and make informed pay decisions.

WHAT YOU KNOW:
- [Company]'s compensation philosophy and pay positioning strategy
- Our pay grade structure and salary ranges
- This year's merit cycle guidelines and budget
- Common benefits questions and plan basics

HOW TO RESPOND:
- Always answer based on [Company]'s policies in your uploaded documents
- If the question involves an individual employee's specific situation, provide general guidance and recommend they work with their HR Business Partner for the final decision
- Never disclose salary information about specific employees
- If you don't know the answer from the documents provided, say so and recommend contacting the compensation team at [email]
- Keep answers concise and practical — managers are busy during merit season
- For FLSA questions, always recommend legal review before reclassifying

TONE:
- Helpful and clear
- Direct — don't hedge every answer with "it depends" if the policy is clear
- Professional but not stiff

SCOPE LIMITS:
- This tool covers compensation and benefits policy questions
- Redirect recruiting, employee relations, or performance management questions to the appropriate HR team

1. Upload your documents in the Knowledge section. Upload each file individually (PDF or Word format). ChatGPT will reference these files when answering questions.

2. Under Capabilities: Keep Web Browsing OFF (you don't want the GPT pulling in external comp data that might contradict your policies). Keep Code Interpreter ON if you want it to help with basic Excel formulas.

Part 3: Test and refine

Before sharing with managers, test your GPT with the FAQ questions you collected in Part 1.

Test scenarios to run:

1. "My employee got a 3.2% merit increase. They expected more because they got an 'Exceeds' rating. How do I explain this?"

2. "Can I give someone an off-cycle adjustment outside of merit season?"

3. "One of my team members just got a competing offer. They're asking for a counter-offer above our grade maximum. What are our options?"

4. "What's the difference between a market adjustment and a merit increase?"

5. "My new hire in California is asking about our salary ranges because of the pay transparency law. What do I tell them?"

For each question, evaluate:

• Is the answer accurate based on your policy documents?
• Is it specific enough to be useful, or too generic?
• Does it hedge inappropriately, or give clear guidance where your policy is clear?

Refine the instructions based on gaps. Common fixes:

• Too generic → add more specific policy details to the instructions or documents
• Wrong tone → adjust the tone section of instructions
• Missing a key topic → add a document or add a section to the instructions

Real Example: Merit Cycle Launch

Setup: You've built the GPT with your merit philosophy, grade structure, and 2026 merit guidelines. You launch it to 45 HR Business Partners the week before merit cycle opens.

Typical week 1 interactions:

HRBP asks: "One of my directors is currently at grade 7, step 4 — which is 110% of midpoint. Are they eligible for a merit increase?"

GPT responds: "Based on our merit guidelines, employees above 105% of midpoint are still eligible for merit increases but at the reduced rate in our matrix: maximum 1.5% for 'Meets' and 2.5% for 'Exceeds.' The full budget applies to employees between 85–105% of midpoint. Since your director is at 110%, the reduced rate applies. If they're a strong performer you feel should receive more, that's a conversation for an off-cycle equity review after merit season closes."

Time saved: That answer took 8 seconds. Without the GPT, the HRBP would have emailed you, you'd have answered it in 10 minutes (factoring in context-switching). At 45 HRBPs asking an average of 8 questions each during merit season = 360 questions → from 60 hours of your time to 60 seconds.

What to Do When It Breaks

• GPT gives wrong information from documents → Go to the GPT configuration and add a clarification to the instructions: "Note: the policy on [topic] was updated in 2026. The correct answer is [X]."
• GPT is too cautious/adds too many caveats → In the instructions, add: "When our policy is clear, give a direct answer. Don't add 'consult HR' to every response — add it only when the situation genuinely requires HR judgment."
• Users ask questions outside scope → Strengthen the scope limits section of instructions with specific redirects: "For questions about equity compensation or RSU vesting, redirect to [name/contact]."

Variations

• Simpler version: Skip the Custom GPT setup and instead create a Claude Project (claude.ai) with the same documents uploaded. Slightly less shareable but easier to set up and maintain.
• Extended version: Add a "Benefits FAQ" document covering the top 20 benefits questions. Now the same GPT handles both comp and benefits inquiries during open enrollment.

What to Do Next

• This week: Build a basic version with just your merit guidelines and test it with 2–3 trusted HRBPs before the full launch.
• This month: Add your job description templates to the GPT so managers can ask it to generate a first draft of a JD in your company's format
• Advanced: Add your company's job leveling criteria so the GPT can help managers self-assess whether a role should be at grade 5 vs. grade 6 before submitting a formal reclassification request

Advanced guide for Compensation and Benefits Analyst professionals. This guide requires a paid ChatGPT subscription.