At the CHRO events we hosted during Gartner’s HR conference in London and Orlando this month, one tension kept coming back: work is changing faster than pay can follow.
One leader’s organization is still paying premiums for IT skills that stopped being scarce years ago. The premium was right when it was set. The market changed, and the pay stayed where it was.
Others at the table worried about the current version of the same problem. Organizations are paying large premiums for AI skills right now, and they expect to be carrying those premiums long after the market catches up.
Pay isn’t built for shifting jobs
Most of the pay system is designed for stability, and for good reasons. Pay is more stable than the work it values. And employees plan their lives around a salary. Job architecture takes years to build and months to change. Merit budgets are set once a year. And base pay ratchets: each raise becomes the floor for every raise after it, and cuts to base are rare.
Performance ratings are the mechanism meant to keep pay connected to current contribution, but they work on raises. A rating tells you how well someone did the job this year. It doesn’t address whether the premium already in their base salary still reflects what the market pays for the skill. And because a merit increase is a percentage of base, it builds on that premium without ever testing it.
The work underneath pay is changing much faster. Gartner estimates that 32 million jobs will be redesigned every year through 2031, a figure its analysts repeated in Orlando, and only 26% of business leaders report successful organizational redesigns. Gartner also reports that AI-related roles are paid up to three to four times the average worker, while AI skill life cycles are shortening to as little as two to five years.
A premium is set in a hot market, for a skill that is scarce that year. The work changes. The skill becomes common, or AI absorbs part of it. The premium stays, because nothing in the pay system is designed to notice that the reason for it has expired.
Premiums are set on the part of work that changes fastest
A job is a bundle of tasks, packaged so the work can be assigned, managed and priced. Pay is attached to the bundle: the grade, the range, the job family. AI is pulling that bundle apart at the task level first. OpenAI’s analysis of more than 800,000 ChatGPT conversations found that 43.5% of occupation-specific AI use involved tasks that historically belonged to a different occupation: people doing work they used to hand off to a specialist. Anthropic’s labor market research, meanwhile, finds the distribution of jobs has barely changed. The tasks are shifting quickly. The job titles, and the pay ranges attached to them, are staying put.
A skill premium sits right on that mismatch. It pays for one task-level skill, the fastest-changing part of the work, through base salary, the slowest-changing part of pay. When the skill becomes common, base salary doesn’t adjust, so the premium stays.
“AI skills” is also a blunt thing to pay for. A task AI completes on its own and a task where AI assists a person who stays responsible for the result are different kinds of work. They probably shouldn’t be trained for, governed or paid the same way, and the job evaluation and skills-based pay systems most organizations use today can’t see the difference. I wrote about this distinction at more length in Before We Ask What AI Does to Work.
An AI premium can also reward the wrong effort. Economists Daron Acemoglu, Dingwen Kong, and Asuman Ozdaglar at MIT modeled what happens when AI agents give people accurate recommendations for their specific situation. Each decision improves. But people put less effort into learning, and that learning is also what builds the shared knowledge everyone draws on, including the AI. Under some conditions in their model, that shared knowledge runs out. Early evidence points the same way: activity on Stack Overflow fell as programmers turned to ChatGPT for the same help. Inside a company, the shared knowledge is the senior engineer explaining why the system was built this way, and the analyst who documents the edge case. That work is rarely what a skill premium is priced on. A premium for using AI, on its own, pays for the replacement and leaves the learning unpaid.
A premium compounds
Finance tends to treat a premium as a one-time number. Base pay compounds, so the premium does too.
Syndio’s research on the hidden cost of ungoverned pay decisions models this for new hires, in parallel versions for EMEA and the U.S. In both, about 30% of new-hire offers land above the organization’s internal equity range, by roughly 8%. In the EMEA model, a €70,000 salary carries a €5,600 premium the day the offer goes out, small enough that finance books it as normal variance. With every later raise calculated on top of it, it grows to nearly €30,000 over a five-year tenure. In the U.S. model, an $8,000 premium on a $100,000 salary grows to more than $42,000 over the same five years. Each comes from one offer decision that took about fifteen minutes.
A skill premium follows the same arithmetic, with one difference. A new-hire overpayment is often an accident. A skill premium is a deliberate choice, made for a reason that was true at the time. That makes it harder to question later, because it looked like good judgment when it was made.
Cutting 10% of headcount doesn’t cut workforce costs by 10%, a point one session made directly. Premiums that have outlived their reason are part of the gap. A layoff removes people. The people who stay keep their premiums.
The managers holding the decisions
The CHROs at our events kept returning to managers. Several want to push more pay decisions out to managers and recruiters, because the comp team can’t make them all. At least one company in the room has never used performance ratings and is tying pay to skills instead. Others described job architecture that can’t keep pace with how work is being redesigned.
Those managers are stretched. A Gartner session on CHRO priorities for 2027 described managers who aren’t equipped to evaluate work split between people and AI agents, and who are carrying heavier loads with less capacity. They’re increasingly the ones now setting the premiums.
A manager deciding on a premium for an AI engineer is usually doing it against a deadline, with a competing offer on the table and no view of what comparable people were paid last quarter. Asking that manager to also predict when the premium should end is unrealistic. The manager isn’t the problem. The decision has no record attached to it.
The fix: write down why, the day you decide
Pay problems are usually found in an audit, months or years after the fact. As we said when we closed the event, problems start far earlier, when someone makes a pay decision and doesn’t record why, which presumes they had a clear rationale to begin with. If the reason is explainable and documented at the moment of decision, the same record answers a regulator months later and the employee who asks their manager how they’re paid compared to peers. Rebuilt later, it usually answers only the regulator and misses the opportunity to build trust with the employee.
That record also fixes the premium problem. A premium recorded with its reason (“15% above range because this skill is scarce in our market as of Q3 2026”) is capable of being evaluated later on. Next cycle, someone can ask whether the reason still holds. A premium without a recorded reason becomes part of base pay by default, because nobody can say what it was for. Over five years, that default is how a skill premium outlasts the scarcity it was paying for.
None of this requires pay to move as fast as work. Base pay can stay stable while each exception keeps its reason attached, so the organization can tell which premiums still serve the strategy and which are left over from a market that no longer exists.
Regulation is pushing the same way. If you employ even one employee in any of the 27 EU countries, the Pay Transparency Directive requires you to explain when employees exercise their right to ask how their pay compares to peers, which requires pay structures built on objective, gender-neutral criteria. On top of that, the first gap reports are due in June 2027 for employers with 150 or more workers. A premium no one can explain is a cost and a compliance exposure at the same time.
A test you can run this month
At the event, we gave the CHROs in attendance a small test, which can be applied to premiums too. Pick one decision your team made recently, ideally an offer with a premium attached. Ask the person who made it to write down why, in two sentences. Then ask two questions. Could the employee’s manager say those two sentences to the employee? And will the reason still be true a year from now?
If the answer to the second question is no, you’ve found a premium with an end date that nobody wrote down.
Which premiums in your payroll could you still explain today?
Go deeper on governing pay decisions
Read the research: The Hidden Cost of Ungoverned Pay Decisions (U.S.) and the EMEA 2026 edition model how a single offer decision compounds over time.
For the regulatory picture: the EU Pay Transparency Directive Tracker shows where each country stands, and the U.S. Pay Scale Transparency Cheat Sheet covers U.S. requirements. Gianni Mui’s article, AI is rewriting job descriptions. Pay transparency has to keep up, looks at the same problem from a practitioner angle.
