Openroll recently presented "Reimagining Compensation in an AI-Driven World" at a Silicon Valley Compensation Association session, alongside Shane McCauley, VP of Total Rewards at Flock Safety. During the session, we polled the room on how comp leaders actually work and where AI fits into their day-to-day reality.
The argument at the center of it is simple: comp professionals are becoming orchestrators of work, not the ones executing every step of it. Not because AI is replacing judgment, but because it's finally possible to stop spending judgment on work that never required it. Everything below, the poll numbers and Shane's own story, is the evidence for that.
in the room
questions
Manual work is still the norm
We asked how long it takes to answer a typical stakeholder request. Most said hours. A significant portion said a full day or more.
I opened the session with a story that made the room go quiet. A friend of mine works in comp at a fast-growing Bay Area tech company. At 4pm on a Friday, her CFO emailed asking how many people had landed below range after the latest cycle, and which teams were hit hardest. He needed it by Monday. Getting there meant requesting data access, pulling reports out of Workday, downloading bands from a shared drive, building an Excel model from scratch, and turning it into a slide. She finished it Sunday night. Roughly 15 hours of work, all of it to answer a question about her own company's data, none of it requiring her actual expertise as a comp professional.
Her team has since rebuilt that workflow around AI connected directly to their systems, auditable at every step. The same kind of request now takes about 35 minutes. She wasn't replaced in that process, she directed it, checked the data pulled was correct, defined the methodology, and reviewed the output before it went to the CFO. The work didn't disappear. It moved from doing to directing.
When we asked how insights get shared once they're finally ready, the answers were consistent.
How do you share insights with stakeholders? (n = 60+)
59% of the room still delivers insights via spreadsheet.
These are ambitious, forward-thinking teams at companies that move fast and invest heavily in better ways of working. Their experience highlights a shared industry-wide opportunity: helping even the most capable teams move from manual workflows to more scalable ones.
AI's biggest obstacle isn't capability, it's trust
These aren't AI skeptics. Most are already experimenting with AI in some capacity.
But when we asked what concerns them most about AI-generated answers, the response was clear.
Speed stopped being the interesting variable a while ago. Auditability is the one that decides adoption now.
The question isn't whether AI can produce an answer fast, everyone knows it can. The question is whether that answer holds up in front of a CPO or a board. Speed is easy. Trust takes proof.
Input debt is the real problem
Shane McCauley shared the example that got to the core issue better than anything else in the session. AI doesn't fix a broken workflow. It amplifies whatever you feed it. Call it input debt: the gap between the data you have and the data your model actually needs, and every team we talked to is carrying some.
Shane's first attempt at using AI for job matching, mapping Flock's broad range of roles to market data, started by feeding full job descriptions into the model and asking it to semantically match them. It got through maybe 40% of the jobs, with low confidence even on the ones it matched. As he put it in the session, he'd already sunk about as much time into that approach as doing it by hand would have taken.
The fix wasn't a better model. It was better input. Flock already had a job hierarchy built for org charting, structured similarly to the market data. Instead of trying to get the model to semantically understand one job description versus another, Shane fed it that hierarchy directly and had it match branch by branch. Coverage jumped to roughly 80% on the first pass, and he finished matching the full job set in about four hours.
AI is only as good as the thinking that goes into what you feed it.
The teams getting real results aren't just automating existing processes. They're paying down input debt before AI touches it.
Judgment is the job
The consensus in the room was pragmatic. Use AI to eliminate manual work, but only if the output is auditable. Trust it for execution, not for judgment.
Here's what that split actually looks like in comp work. Pulling comparable pay data, flagging a role that's drifted out of its band, matching jobs to a benchmark database: that's execution, and it's exactly the kind of work Shane and the room want off their plate. Deciding whether a below-band offer is worth making anyway to land a critical hire, or how to message a compression issue to a manager before it becomes a retention problem: that's judgment, and nobody wanted AI making that call for them. The line isn't about how hard the task is. It's about whether the answer is defensible on its own, or whether it needs a person who understands the stakes behind the number.
Shane summed up the shift plainly:
"I haven't written a spreadsheet formula in months. I've moved into an orchestration layer, spending my time on the prompt and the output, and understanding where I actually need to apply my judgment."
Not because the analysis stopped happening or because his expertise matters less, but because the mechanical part of the job stopped requiring him personally. The time that used to go into building models now goes into the calls only he can make: whether a match is right, whether a range makes sense, what a number actually means when it lands in front of a leader who's going to ask why.
AI won't replace compensation professionals. It will replace the parts of compensation that never needed one.
That's the direction the room is moving. And it's the shift Openroll exists to support.
