FORWARDED THIS? Subscribe here to get the next one.

ISSUE 07 · AI + COMPENSATION

A NOTE FROM GIAC

Two things kept running into each other this week.

The first was recording with Theresa Cortese. She runs total rewards at Nirvana Insurance and she built her company's variable compensation system herself, in Claude Code, in the evenings. It's live now.

The second is the conversation I keep having on calls. People with the skill and the appetite, sitting behind a policy nobody has written down, or a licence for a tool that only does chat.

The difference between them is what their employer lets them touch, and who pays for the hours they spend learning it.

So I built something to measure that. It's at the bottom of this issue.

Giac

THE BREAKDOWN

Theresa Cortese is Senior Manager, Compensation and Benefits at Nirvana Insurance. Twenty plus years across retail, semiconductors, ed tech and HR tech. Earlier this year she sat on a 45 minute call with her business operations team, working out how to calculate the quarter's variable compensation. The same call they have every quarter.

The problem is boring, which is why it survives. Business operations own the metrics and the attainments. They can't see pay data. So the two teams keep separate spreadsheets and check them against each other every quarter, five senior people, roughly a week of their time. She asked whether Google Sheets could hide a single column from someone with edit access. It can't. Anyone with edit rights can unhide it.

The call ended with everyone agreeing to go and look at buying a tool.

She has nothing against buying software. A compensation planning tool she bought is what finally retired the multi tab merit spreadsheet she'd spent fifteen years evolving, macros and all. Her company just sat in the gap where the manual process was costing five senior people a week a quarter and an enterprise platform still didn't make sense.

Then she remembered what her CEO had said at the All Hands. Experimenting with AI wasn't optional, and everyone should build something that week.

She opened Claude and asked whether she'd lost her mind or whether this was something she could build. It told her to buy a tool. She asked again, this time mentioning she had access to Claude Code. The answer came back that this changes everything.

She'd never opened Terminal on her Mac before.

What she shipped is a compensation management system with Google sign in enforced at the database query level, so an employee can only ever see their own pay and their own history. Business operations upload attainments from a CSV template. Payroll runs the calculation, exports it, and can compare any two runs to see exactly what changed between them. It generates payout letters explaining how each number was reached, which the company never had before.

Q1 actuals passed for every employee. It runs 110 tests on every change she makes. The repository sits with IT now, moving to a hosted environment, and payroll will keep running the old manual process alongside it for a cycle or two as a safety net.

She treats the thing she built as an entry level headcount she finally got approved. She teaches it the process, she checks its work, and when it gets something wrong she doesn't start again, she tells it that going forward it should do one step before the other. On trust she's blunt. She isn't ready to trust it. She watches it make mistakes. She checks everything. Checking still costs far less than doing, and it leaves an audit trail she can walk back through.

That only works because she knows all eleven of the company's variable comp plans, what gets calculated first and what feeds into it. Without that she'd have no way to look at an output and say no, that's wrong.

It also cost her time before it saved any. She built an equity refresh calculator in Python while running the actual refresh by hand in parallel, stopping at every step to feed inputs in. It took considerably longer than just doing the work. Then she ran the finished tool and it reproduced a week of calculation in under a minute. The saving arrives next cycle.

She's clear about why this matters to her specifically. At the startups she works with, nobody is hiring a team underneath her. There was never going to be one. So she trained up the help herself.

The hours went in on her own time. Most evenings, Netflix running on her phone to the left, Claude Code on the laptop to the right. Nobody funded that, and nobody asked her to account for it. Who pays for the time you spend learning this?

She almost didn't show any of it. She had flu the week we were first meant to record and admits she was relieved to reschedule, embarrassed to show me something that wasn't a finished product. A post about the pressure on women in tech to make everything perfect before showing it changed her mind. She shipped V1 to IT that week and turned up to the recording.

Her advice to her own past self was to stop thinking about starting.

She walks through the tool on screen in the episode. Watch on YouTube or listen on Spotify.

RECENT EDITIONS

ON THE RADAR

FROM RANGE

Marko Uskokovic went public this week about joining Range. Worth a follow if you're building.

Theresa's CEO told the company to build something that week and put the tools on her laptop. Plenty of people I speak to get the opposite, and some get told to adopt AI and blocked from installing anything in the same month. Which one are you?

Nobody has ever measured that, so Range built The Permission Gap. Twelve questions on what your employer actually allows you to do with AI, about five minutes, and your own result comes back the moment you finish. It asks nothing about salaries, budgets or market data.

Contributors get the full benchmark before it goes public.