Getting AI to respond to recruiters like a compensation expert

At a high level, the goal sounds simple.

‍Can you make AI respond to recruiters the way a strong comp partner would?

But when I tried breaking down how to “clone” a comp partner, it became clear pretty quickly that this isn’t one problem.

It’s a stack of problems.

You’d need deep domain knowledge. Not just definitions, but conventions (how offers are structured, how percentiles are used, what’s considered a reasonable recommendation versus an aggressive one, etc).

These are things comp people internalize over time.

You’d also need company-specific context. Every company has its own rules, even if they sound similar on paper. Two companies can both say they target the 90th percentile and still behave completely differently when it comes to real offers.

And then there’s the need to understand the audience. The way you explain a recommendation to a recruiter is different from how you’d explain it to a hiring manager or an executive.

That’s what a real comp partner is doing all day, often without thinking about it.

Now try to get AI to do that.

AI doesn’t actually reason through any of this the way a human does. It’s not sitting there thinking about your comp philosophy or your policies. It’s just taking the text you give it and predicting what the next word should be.

It’s a prediction machine.

It might repeat what you’ve told it. It might sound right. But getting it to consistently apply that information the way a comp person would…. that’s where things get messy.

‍Consistency isn’t just an AI problem. It’s a human problem, too.

Comp is a policy-driven function. There is a “right” way to do things. That’s the whole point—managing spend, ensuring fairness, maintaining trust.

But in practice, it’s executed by people and people are inconsistent.

You might have a handful of comp partners supporting dozens of recruiters. Those recruiters are asking questions constantly. Some of them are experienced, some of them are brand new, and most of them don’t have deep fluency in your comp programs.

Ten comp partners might execute the same comp philosophy in ten slightly different ways.

One person is stricter. Another is more flexible. Someone’s having a bad day and just wants to get an offer out the door, so they bend the rules a little.

Individually, those decisions feel small. But when you zoom out, they show up in your P&L.

A lot of the work isn’t strategic. It’s repetitive. It’s answering the same types of questions over and over—ranges, policies, edge cases, “can I do this?” scenarios.

And the reality is, most HR and comp teams spend a meaningful portion of their time on exactly this kind of work. Studies put it somewhere in the 30–40% range, which lines up with what I’ve seen.

So you end up with this tension where your most valuable comp talent is spending time on low-leverage tasks.

Execution across the organization is uneven and the system relies on constant human intervention to stay on track. Then you introduce AI into this.

The expectation is that it will fix the inconsistency. But if you’re not careful, it might make things worse.

One of the early failure modes we ran into was pretty simple. We gave the model a lot of documentation—multiple years of comp policies, different program details—and assumed more context would lead to better answers.

What actually happened is the AI just blended everything together.

It doesn’t know what’s current versus outdated. It doesn’t know which details matter for a specific question. It just synthesizes across everything it sees, which means you can get answers that sound grounded but are completely wrong for the situation.

Not because the model is “making things up” randomly, but because it’s trying to be helpful without a real sense of what matters.

Now you have two imperfect systems: inconsistent humans and probabilistic AI.

And neither, by default, is reliably executing your comp strategy the way you want.

‍Stop making AI smarter and start making it behave correctly

Instead of asking, “How do we get AI to think like a comp expert?” I started thinking about, “How do we constrain it so it operates like one?”

A few things started to matter a lot.

You have to take the way your team actually works—how you structure offers, what data you use, how you think about pay mix and exceptions—and distill that into something usable.

Find the most relevant, most concise version of your strategy that still produces the right behavior.

I think about it like training a new comp partner. If I had an hour with them, I wouldn’t give them four years of documentation. I’d give them the essentials so they can operate effectively right away.

Next we started designing for who the AI is talking to, not just what it’s saying. A recruiter needs something different than a comp analyst.

A correct answer delivered the wrong way is still a bad answer.

Under the hood, that means defining how the system should communicate in each context, based on how comp people actually work with those stakeholders.

The third piece is consistency.

One of the more practical things we’ve done is encoding compensation domain knowledge in our system prompts — the “invisible hands” that guide AI systems.  We configured our agents to reflect how a strong comp partner would approach common situations.

1. How do you respond to an exception request?
2. How do you structure a recommendation to a recruiter? Or an exec?
3. How do you explain ranges clearly without overcomplicating it?

When you standardize that, you’re improving AI output while raising the baseline for your team. 

A more junior person can operate closer to your most experienced comp partner because the system is guiding how they work.

Recruiters get answers to common questions instantly, without pulling a comp partner into every interaction. They access ranges and policy-aligned guidance in one place instead of jumping between tools. And when it comes to offers, you can structure the way inputs are gathered and outputs are generated so it actually reflects how comp would approach it.

The goal isn’t to replace comp.

It’s to remove the work that doesn’t require your judgment so you can focus on the work that does.

AI can already improve comp alignment at scale

Turns out comp’s “old” system was also probabilistic.

You had policies, but execution depended on people. And people vary.

AI doesn’t magically fix that.

But if you design it correctly, it gives you something you don’t get with humans.

It will follow the same rules, the same way, every time.

No shortcuts. No inconsistency. No “just get it out the door” decisions.

That’s the part I think people are underestimating. The value isn’t just speed. It’s alignment.

Not “does this sound like a smart answer?”

But “does this behave the way our best comp partner would want it to behave?”

We’re not fully there yet.

But we’re starting to see what it looks like when you get closer.