FORWARDED THIS? Subscribe here to get the next one.
ISSUE 11 · AI + COMPENSATION
Before you build anything with AI in comp
TL;DR
Benchmarking is where nearly all of us start, and vendors already do it better than a weekend build will. Martin Smit thinks that's good news, because it leaves you the parts of reward nobody has built for yet. Put a hard spend cap and a kill switch in before you run anything.
A NOTE FROM GIAC
This one is late and the reason is pretty simple. Yesterday spent the entire day browsing not one but two IKEA stores in Madrid so you can imagine what I’ve been up to. But the apartment is coming out nicely and I can’t wait to host a Range house warming party! I’m sure nobody actually noticed about the delay, but if you did… reply to this email, you’re a legend.
Enjoy this week's edition.
Giac
THE BREAKDOWN
The best AI build in comp is the one nobody in your function has tried
Martin Smit spent 25 years in HR and total rewards inside large pharma and oil and gas businesses. He started coding in 1984, he spends about $500 a month on his own tools now, and he builds on most days and weekends. I asked him what people get wrong with AI in reward, and he went somewhere I wasn't expecting.
Almost everyone starts with pay benchmarking. Can AI do our market pricing. Martin's answer is that of course it can, he could build you something in a weekend that pulls competitor sites and consolidates what it finds, and then you're holding a pile of numbers with all the hard work still in front of you. Are the jobs actually aligned. Where did the value come from. Vendors solve that with better data and far more validation than a weekend build will ever carry, and paying a bit more for higher quality is a reasonable trade.
That sounds like a door closing until you follow where it goes. If we all develop our own AI tools for benchmarking, and then all develop AI for pay equity, we may end up with thousands of different versions of two problems that already have good answers. The consequence is that we may all hold optimal tools while never pointing AI at other activities that could be worth far more.
Everything else is wide open ground. Martin's way into it is a greenfield question. Drop your constraints, at least mentally, for ten minutes, ignore the governance model, and ask what the company or the function actually needs or wants. Then ask where AI plays a meaningful role in that. Done in that order, you stop shaving time off work you already do and start on work you couldn't do at all.
He also thinks small teams have the advantage here, which is not what a room of small biotech reward leaders expected when they asked him. To develop usable proof of concept products you don't need a huge budget. You need a few hundred dollars a month and, if you have a sandbox environment, none of the governance drag that tends to slow innovation down.
The second half of our conversation went somewhere I hadn't planned for. One day he left an agent running overnight in auto mode, expected one document to be processed, and the agent processed thousands of them, which left him with a bill he never expected. His learning was to be more precise in what he instructs an agent to do, and to put in hard spend caps rather than budget alerts, because most billing alerts arrive after the money is already gone. Plus a kill switch, so that an unattended job, or an API key used by the wrong actor, can't set you back thousands of dollars. He also recommends against using people data unless you really know what you're doing, because there's plenty of room to use AI without touching it at all.
I've put the whole thing on a card. Five questions to answer before you start, four guardrails to have in place before you run anything.
PODCAST
 |
Martin's episode went live today. He walks through the executive compensation database he built with Claude Code, 10,000 of the world's largest companies with four years of annual reports and proxy filings pulled apart and made searchable, and the pay equity tool he built in a weekend that runs entirely on your own machine so no employee data ever leaves it. He'd also like the annual compensation review killed off, which felt like a good place to stop.
ON THE RADAR
The New York Fed has AI adoption nearly doubling in a year. Published 5 August. Service firms using AI went from 25% to 40% in twelve months and manufacturers from 16% to 26%, with both expecting to climb again. Very few AI driven layoffs so far. What's moving instead is hiring plans, especially for college educated workers. Read → the analysis
If nobody is being laid off but entry level hiring quietly slows, the first thing that breaks is your job architecture, and it breaks somewhere nobody is looking.
A third of US adults have already asked AI what they should be paid. The Harris Poll surveyed 2,131 US adults and found 33% have asked AI about a salary, a raise, a bonus or how to negotiate, and 47% would let AI run the negotiation itself. More than 75% had no idea AI can recommend lower salaries for women and people of colour. Read → the findings
That last figure is the one I'd sit with. Our employees are taking pay advice from a tool they've never questioned, and bringing it to a manager who has no idea where the number came from.
Martin reckons almost all of us start in the same place. Let's find out whether he's right.
What was the first thing you tried to build with AI in comp?
RECENT EDITIONS
When you're ready, here are two ways Range can help you.
1. 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, all in the open. Browse the Comp AI Vault →
2. Join the community
Find your people. Inside Range, reward pros learn to build with AI together, the prompts, the setups, the band workflows that actually hold up. Free is the signal, the room is the substance. Apply to join →
Reading this because a friend forwarded it? Get Range Weekly in your own inbox every Sunday. Subscribe free |
 That's me for this week. The Madrid flat is nearly a flat, so next week's edition should land on time. Until next week, Giac | |
|
 |