FORWARDED THIS? Subscribe here to get the next one.

ISSUE 09 · AI + COMPENSATION

New this week? Welcome. Most people arrive here from one of Range AI in comp guides and today’s newsletter includes two major updates for teams using Microsoft Copilot and Google AI.

THE PERMISSION GAP

Most compensation people aren't held back by what they can do with AI. They're held back by what their employer lets them do. Nobody has measured that, so we did.

It takes under 3 minutes to participate to the poll and your result comes back the moment you finish. Contributors get the full benchmark before it goes public. Deadline August 31st.

THE BREAKDOWN

Greg Laney has over 20 years across compensation and HR systems. Hay point factor analysis first, then VBA, then Access databases a mentor talked him into, then Workday, and a psychology degree underneath all of it.

About 18 months ago he tried to get ChatGPT to do job matching for him. He took the survey spreadsheet from his consultant, turned it into a file the model could read, and built an agent.

That simple? Learning a new skill is all about trial and error. He hit a ceiling at some point, but that didn’t stop him from making progress.

So he paid 20 dollars for a Udemy class, spent 3 or 4 hours learning enough n8n to move around in it, and then spent 4 hours rebuilding the thing properly.

He shares his screen and runs a job match in front of you. A hiring manager emails over a couple of sentences to describe what a job does. Greg pastes it into a form and hits submit. The workflow reads it, compares it against the survey data, compares it against the internal job catalogue, returns a primary match and an alternate, works out which grade it lands in, and drafts the email back to the manager.

Work that took him 4 hours a day now takes 5 to 10 minutes.

He didn't write a 20 page prompt.

Every box in that workflow is one step, and he built them one at a time. Ask AI to write the code for the first box, run it, look at what came out, and only then start the second. Some of them took him under 5 minutes.

When he wanted his compensation philosophy in there, the percentile the midpoint anchors to, the range spread, how much overlap he allows between grades, he didn't bury that in a prompt either. He asked the AI to write it into the code inside that one box, then tested the box on its own before moving on.

Greg has since sent over the workflow itself and the sample data behind it. The real canvas runs to 16 nodes, and the workbook he tested it on is in there too, six sheets of made up numbers you can open and run your own job titles through.

Building it that way is what makes the output defensible.

You can see the input and the output at every step, so when something looks wrong you know which box to open. Every run also writes itself back into the spreadsheet with the match, the grade and the reasoning, so in 6 months somebody can still see why a job was evaluated the way it was.

Greg's own position is that AI hasn't caught up with our professional expertise, so a human still has to look at the grades coming out of it. The workflow does the work you’d delegate to a junior analysis and hands you the decision.

By the way, he has no idea how to write JSON, i.e. a text-based data format used to store and send data. He asked the model to write it and then checked it did what he wanted.

If you’ve never heard about n8n, Make or Zapier, watching Greg walking through his job matching workflow box by box will help you decide whether it’s something worth investing your time in.

Listen on the podcast →

AI IN COMP GUIDE UPDATES

The Copilot guide was rewritten on 5 August, two days after GitHub Copilot's harnesses went generally available. A harness decides how much of a job the AI takes on before handing back to you, and the three of them bill differently. The catch nobody puts on a slide is that Copilot Credits burn while you're building, before the thing has run once for real. The Workday and SuccessFactors connectors are rewritten too. They exist now, and they still can't produce a merit population.

Read the Copilot guide →

The Google guide got a bigger change on 13 August, and Antigravity now has a chapter of its own. Everything else in that guide works on the one document you have open. Antigravity works on a whole folder. Point it at twelve survey exports, all shaped differently, tell it in plain English to clean them into one master sheet, and it reads, edits and writes the files itself. Google's own tool will also happily run Anthropic's models, which we weren't expecting.

Read the Google guide →

ON THE RADAR

RECENT EDITIONS