Forwarded this? Range Weekly is free. Subscribe and it lands in your inbox.
Every other function has its public AI builders. Compensation still mostly doesn't, so the only way to know what people actually run is to ask them one by one.
This is six patterns from over 100 conversations with compensation and total rewards leaders.
THE SIX PATTERNS, AND WHERE EACH ONE GOT TO
Here's the whole map at a glance, then the story behind each one underneath.
WORKING NOW
Three of these four never leave the chat window.
1. The spreadsheet nobody could unpick. A multi country profit sharing calculation across fifteen entities and three countries, the kind that lives in macros and formulae nobody wants to inherit. Prototyped with ChatGPT and Claude open side by side, after two years of trying to get it automated through IT, Finance and a vendor (most likely, the wrong one).
2. Two models reading your survey exports. Reports downloaded, then two models working over them together to surface range discrepancies, title proliferation and overlap, ending in a ranked view of what to fix first.
3. The second comp person in the room. Comp comms rewritten so they land for a finance audience rather than a manager, and comp theory stress tested against cost scenarios before it reaches anyone. This is the most common use by a distance.
4. Custom GPTs for the steps of a design project. One per major step, each carrying its own context so the brief doesn't get retyped every time. They work separately and they stay separate, because joining them into a single workflow that holds context throughout brings hallucination and drift with it.
WHERE IT STOPS
Neither of these is a story about the model being too slow or too stupid.
5. Confidence that isn't accuracy. A job ad matching run in Codex returned errors inside results it rated above 85% confident. The person who ran it thinks that's down to feeding in untidied job adverts rather than the model, which is a fair caveat and also the point. Nobody found those errors except by reading every row.
6. What happens when managers need it. The job levelling tool that drew the most attention in Range this summer was deliberately not built on a custom GPT. It runs on a deterministic script, which gives role based access control and keeps salary data bounded and auditable. The architecture is future proofed for an AI API later, and the decision was about what happens the day somebody other than the builder opens it.
WHERE THE WORK LIVES
Four surfaces, and the wall sits between the third and the fourth.
THE THREE THINGS THAT DECIDE WHETHER IT LEAVES YOUR LAPTOP
The map above is the easy part but what keeps a working build on one laptop shows up across all six.
You can't see where the error is. A confidence score is not an accuracy score, and on this evidence the gap between them is wide enough to sit under 85%. Checking meant reading every row, which is the cost nobody prices in when they estimate the time saved.
Joining the steps up breaks them. Four custom GPTs that each hold their own context work. One workflow that holds context across all four drifts. Whoever builds the second version of this will hit the same wall.
Access control is a product decision. The moment a tool faces managers it has to know who is allowed to see which salary, and it has to leave a trail. That is why the levelling tool is a script.
THE LINE THAT SEPARATES THEM
Items one to four are one person working faster. Items five and six are what happens the moment the output has to be checked or handed over to a hiring manager. Nobody in this list stopped because the AI model wasn't clever enough.
 | That's the map as it stands in September 2026. If you're the only comp or reward person trying to figure this out, come and do it in good company. Giac |
When you're ready, here are three ways Range can help you.
1. Join the community
Find your people. Range is a private, vetted and capped community of comp and total rewards pros learning how to build with AI together. Apply to join →
2. Grab the free resources
Everything free we've made now lives in one place, the Comp AI Vault. The guides, the prompt library and the frameworks. Browse the Comp AI Vault →