AI can be a useful tool for gathering and organizing compensation data, particularly when public information is readily available. However, compensation strategy requires far more than data collection. Market benchmarks, salary structures, and pay decisions all require context, judgment, and a deep understanding of organizational priorities. While AI can help surface information, it cannot replace the human expertise needed to design compensation programs that are fair, defensible, and aligned with business objectives.
We’re all inundated with AI content: how AI is taking your job, how AI can help you do your job more efficiently, how you’re falling behind if you aren’t using it in every facet of your work, and how AI is apparently both saving and destroying the world. Ironically (and unsurprisingly), a significant amount of this content is written, edited, or transformed in some way by AI.
I’ll save my broader thoughts on AI for another article (including my aversion to adopting the nomenclature of ‘AI’ as broadly as we all have), but I want to talk about how this new reality is impacting my corner of the world specifically: Compensation and Total Rewards.
Before I do, a few disclaimers:
-
- When I speak of AI, I’m referring to the seemingly universal instances of large language models (LLMs) and their semi-agentic evolutions. ChatGPT, Claude, Copilot; the stuff most of us are interfacing with on a day-to-day basis.
- As a semi-reformed luddite on LLMs as a general concept, I am neither an expert, promoter, nor dissenter of AI tools.
- None of this was written by AI.
We’re fortunate to work with a fair number of not-for-profit organizations. When I put myself in the shoes of an Executive Director leading a 100-staff organization with a fixed revenue or funding model, I can absolutely see the attractiveness of asking an LLM to find market data, suggest changes to salary grids, support salary planning, and even draft communications for difficult compensation conversations.
In many organizations, especially lean ones, compensation administration is not a full-time function. It is layered onto finance, HR, operations, or the Executive Director themselves. In that context, the prospect of AI does not feel like a novelty, it feels like the obvious solution.
I don’t fault anyone for turning to their in-house LLM for any of this. In some ways, it almost feels irresponsible not to. But as a consultant in this space, we are starting to see some consequences with AI use, much like the consequences of failing to stay up-to-date with organizational compensation reviews.
Why Compensation Data Has Always Been Complicated
For most of my career in Compensation, there has been a fairly accepted reality that most public self-reported aggregator sources for compensation data (Indeed, Glassdoor, PayScale, etc.) are generally inferior. While they might provide some directional marker for general pay ranges (for example, how much more does a VP level employee get paid compared to a director), they are rarely considered reliable enough to support meaningful compensation decisions. The data is largely self-reported by employees, the data sample sizes are often too small to be statistically significant, and in my experience, this data is rarely illustrated using current-year equivalent dollars, with some sources using information that is years, or even decades old.
Working on the consulting/data provider side of the equation, my skepticism of these sources is admittedly obvious. But, even as a practitioner, my colleagues and I know to steer away from these sources as they just don’t provide sound information which we’d be comfortable basing compensation recommendations based on.
The AI Data Problem
I bring this up because when an LLM produces compensation “market data,” it isn’t always clear where that information originated. Is it drawing from formal survey sources, public disclosures, collective agreements, job postings, self-reported employee submissions, Reddit comments by disgruntled employees, or some messy combination of the above?
And yes, prompt-engineering your model can help narrow the sources being used. Public disclosure documents and other open data platforms are available to provide valuable information and AI is objectively good at finding, consolidating, and summarizing those sources. It is possible, with solid prompt engineering and fair amount of time, to yield usable, consolidated compensation benchmarking data, particularly for unionized roles where collective bargaining agreements are often readily available online.
That said, I would still encourage anyone using AI-generated compensation date to validate the underlying sources and any math you ask it to do. Because even when the data is technically accurate, compensation decisions remain far more complex than data collection.
Compensation Is a Grey Area
Total Rewards is, and will always be, a grey area. Data matters, and good data matters even more. But even the best data does not eliminate the need for judgment and nuance.
No organization is exactly akin to another, and when it comes to compensation benchmarking, we see this as a good thing. Being able to navigate the data and make inferences on why certain data points are high or low, what might be driving statistical values one way or another year-over-year, and why one organization may be doing something different than the next has always been where compensation professionals create value; be that in-house or external.
It’s this grey area where I remain skeptical on whether AI can objectively or intuitively navigate the source material. LLMs cannot fundamentally discern and account for outlier data points beyond the explanations provided at the source. Anyone familiar with how compensation data is reported knows that companies aren’t typically volunteering contextual information about individuals, programs, or internal processes. Even the major comp data providers cannot reasonably ascertain or report on context beyond quantifiable attributes like location, industry, number of employees, or revenue.
If Jane’s salary sits above the maximum of her salary range because she recently took a new role in a lower level of the organization, the data doesn’t know this. The data simply says that Jane’s company pays highly for her that role.
When Confidence Becomes the Problem
LLMs in all their confident and occasionally sycophantic glory, do not particularly like grey areas as a concept. They are designed to provide answers and will do so confidently with something that sounds thought out, deliberate, and complete, even when given very little to work with.
This bias rears itself when sourcing data through these AI models as well; the model wants to provide a good enough answer in as little time as possible, so when it finds one, two, or three data points (even if those are all sound bites of information), it concludes that is has enough information to get the job done.
A salary range produced in seconds can look just as polished as one built through a proper compensation review. It may have minimums, midpoints, maximums, percentages, job levels, and tidy explanatory notes. But polish is not the same as validity. A compensation model can be beautifully formatted and still be fundamentally inappropriate.
As an experiment, try this yourself. Using your LLM model of choice, tell it that your organization pays $75k for a Logistics Coordinator role, and then ask it to design an entire compensation model for the rest of the organization around that single data point. It will happily comply, but what it generates will be utterly problematic.
When a Bad Benchmark Becomes a Governance Problem
The bigger risk is not just that an organization lands on the wrong market number. It is that the wrong number gets embedded into a compensation structure. One questionable benchmark can cascade into job levels, salary ranges, pay bands, annual increases, internal equity decisions, and future employee expectations. At that point, the issue is no longer a bad AI output, it is a compensation governance problem.
After a couple of compensation cycles, this risk can transform from being a small misstep into a structural issue requiring a ground-up rebuild.
There is also the defensibility or credibility issue. When employees ask why a role sits in a certain range, why one job is valued differently than another, or why salary adjustments were made for some employees and not others, “the model suggested it” or “that’s what the data said” is not a compensation philosophy.
Organizations still need to be able to explain the rationale behind their decisions in plain language, articulate why leadership has made these decisions, and do so with accuracy and empathy.
Data Is Not Strategy
LLMs are trained to find patterns, and compensation strategy should not be about copying the most common pattern.
It’s not unheard of that an organization intentionally leads the market for one job family, lags for another, protects internal equity over external competitiveness, or makes a short-term affordability decision while building toward a longer-term target. Those choices are strategic and perfectly reflect the nuance required when administering compensation programs.
Ultimately, even if AI models could consistently provide sound, verifiable total rewards data, it would still be just data. Expert inference and reasoning beyond the numbers are still things we as humans hold a competitive advantage over LLMs in.
Why Human Judgment Still Matters
The justification for compensation professionals has not diminished because of the existence of LLMs.
Total Rewards experts know that, while informed by data, a sound compensation strategy needs to account for all sorts of variables when designing plans. Your plan also needs to balance the internal realities of budget constraints, employee equity, growth trajectory, attraction/retention concerns, organization philosophy, and practical implementation to name a few.
Most importantly, compensation decisions affect people. The human beings you employ within your organization still matter, and no AI model can fully understand their experiences, responsibilities, motivations, or concerns in the same way other human beings can.
AI can help organizations gather information faster than ever before. But compensation has never been about finding a number. It's about understanding what that number means, how it aligns with organizational priorities, and how it affects the people behind the job titles.
Your employees deserve a plan that takes nuance and judgement into consideration. There is rarely a single ‘correct’ way to administer compensation, but for every organization, there is usually a ‘best’ way.
AI is changing how organizations access and analyze information, and compensation planning is no exception. Used appropriately, these tools can improve efficiency and support decision-making. However, compensation strategy remains fundamentally human work.
As AI tools continue to evolve, organizations should view them as inputs into the compensation process—not replacements for the expertise, governance, and thoughtful decision-making that underpin a successful Total Rewards strategy.
Need a Second Opinion on Your Compensation Strategy? Book a conversation with our team to discuss your Total Rewards strategy.
About the Author:
