FORWARDED THIS? Subscribe here to get the next one.
TL;DR
OpenAI shipped ChatGPT Work, an agent that runs for hours and hands back finished spreadsheets. What it means for comp, how to test it safely this week, plus Ryan Buhrke's live payroll agent and my own case for why a human still owns the call.
A NOTE FROM GIAC
Earlier this year, before community building became my full time job, I took on a pay benchmarking project for a client, around 200 employees. A job like that used to be ten grand and the best part of three weeks. This time it took me about a week spread over two months.
The AI did the tedious middle. The survey matching, the aging, the data cleanup, the first cut of the salary ranges. I read every number myself because it was my name on the deliverable and real people were going to be paid against it.
This week OpenAI shipped ChatGPT Work, an agent that runs for hours and hands you finished spreadsheets. What got shrunk is the processing part, things like aggregating datasets, visualising the data, identifying gaps and outliers, etc. The decision still becomes a payslip or a contract, and you can't unsend that without mutual consent so a human still signs off.
That's the thread running through this whole issue. Ryan Buhrke built an agent that consolidates 140 payroll files in 7 minutes, and he's the loudest voice in the room on why humans stay in charge. My own episode makes the same case from the other chair.
Grab a coffee. This one's worth the 10 minutes.
Giac
PS: Five Range founding members went public this week! More at the bottom.
THE BREAKDOWN
ChatGPT Work lands. What actually changes for comp.
On 9 July OpenAI merged ChatGPT and Codex into one desktop app and called the agentic half ChatGPT Work. You give it an outcome and it works on its own for minutes or hours, then hands back finished spreadsheets, decks and dashboards. It runs on every plan, including the free one.
For comp the uses are obvious. It can refresh a merit cycle dashboard, clean up survey benchmarking that used to eat a week, or build a board deck straight from the analysis instead of rebuilding it by hand at 11pm. Ryan walks through his own version further down.
Open the app and ask it to generate a realistic dummy comp dataset, roughly what you were working with 5 years ago, then point it at a merit cycle task. No live employee data. You'll learn more in 20 minutes of watching it plan than in any launch video.
Then keep the approval gate manual. The tool can plan and model all day, but the moment its output becomes a payslip or a contract, a human name goes on that decision.
A human name still goes on every pay decision.
This news only just landed. I'll put ChatGPT Work through real compensation and benefits work over the coming weeks and come back with a proper guide on how to use it, the same way I did with Claude, Microsoft and Google.
RANGE PODCAST
Ryan Buhrke built an AI agent that runs his rewards team's grunt work
Ryan Buhrke runs total rewards and people ops at Edmentum. He's also one of the first people I found building real AI tools in comp, with zero engineering background.
His agent is called ARIA. Built in Claude Code, it holds about 20 reusable skills, writing job descriptions, leveling roles, market pricing, moving files. In the episode he runs it live, consolidating 140 payroll registers across a messy HRIS migration in 7 minutes. The only thing that ever leaves his machine is file names and paths, never the banking data or the employee names. His rule is simple. If you can see it with your eyeballs, so can AI.
Then he draws a hard line. He calls it radical agency, the idea that you no longer need permission or a dev team to build what you imagine. In the same breath he says the real job is having the guts to make the call the models won't. AI is great at optimising options. It can't care.
AI is great at optimising options. It can't care.
ELSEWHERE
I went on We Need to Talk About HR to argue the opposite of the hype
Back in April, when Range was still just an idea, I sat down with Louise Fraser on We Need to Talk About HR. The episode is titled Why You Can't Automate a Pay Decision, so you know where I stand.
When an employee turns up with an AI salary number and wants to argue, don't ban the chatbot. Tell them to open Claude, Gemini and ChatGPT side by side and ask each the same question. They'll get a different answer from every model, and another if they come back an hour later. That volatility is the lesson. The number is scraped from Reddit and self reported data, and it moves. It's the best way I know to show someone why the output can't be trusted.
ON THE RADAR
Comp teams have the data. Almost none trust AI with the call. Most comp teams have solid data foundations. Almost none let AI near an actual pay recommendation yet. A new survey of over 500 comp pros puts real numbers on the gap. Read →
Real wages are shrinking. June payrolls added just 57,000 jobs and hourly earnings rose 3.5% over the year, below the latest inflation reading of 4.2%. Pay is losing ground, and that moderating merit budget is a real cut in disguise. Read →
Europe's pay transparency deadline came and went. The EU Pay Transparency Directive's transposition deadline passed on 7 June with only 4 member states ready. If you run pay across Europe, the patchwork just got messier. Read →
Benefits are getting squeezed from both ends. GLP-1 weight loss drugs are straining budgets so hard that some employers are dropping the coverage, and mental health benefits are getting cut too. Mental health offerings fell 6 points in a year, down to 82%. Read →
Reading this because a friend forwarded it? Get Range Weekly in your own inbox every Sunday. Subscribe free |
FROM RANGE
5 founding members went public this week.
They work at different companies across different industries, and every one of them thought they were the only person building AI into comp.
Are you the only one building this? No. You're just the only one in your company.
That's the whole point of Range. A private, vetted and capped room where we practice AI together.
How far would you let an AI agent into your comp work?
RECENT EDITIONS