Is Experience Still Worth Paying For When AI Narrows the Skill Gap?

By Victor Hoang · Updated Oct 10, 2026

Here's the question that shows up in compensation reviews and workforce plans right now: if a newer hire with a good AI assistant can produce work close to what a ten-year veteran produces, why pay the veteran more? The research on this is better than the hot takes. AI does lift less experienced workers the most on routine tasks. But the premium on experience doesn't disappear. It moves, from doing the standard work faster to handling the exceptions, making the judgment calls, and knowing when the machine is wrong.

That shift changes what you should pay for, who you should hire, and how you should develop the people who'll be your senior staff in 2030.

Key Facts

• A study of 5,179 customer support agents found a generative AI assistant raised productivity by about 14% on average and about 34% for novice and low-skilled workers, with minimal impact on experienced and highly skilled workers (Brynjolfsson, Li and Raymond, QJE 2025).
• In an experiment with 444 college-educated professionals, ChatGPT cut task time by 0.8 standard deviations, raised quality by 0.4, and reduced inequality between workers because lower-ability workers benefited more (Noy and Zhang, 2023).
• In a 758-consultant study, accuracy on a task outside the AI's frontier fell from about 84.5% without AI to 60% and 70% in the two AI groups, and below-average performers gained about 43% on tasks inside the frontier.
• Employment for workers aged 22 to 25 in AI-exposed occupations is about 19% below the trend of less-exposed peers, with no comparable gap for experienced workers (Stanford Digital Economy Lab, descriptive evidence).

What the Research Actually Shows

The first half of the story is real, and leaders shouldn't wave it away. Three well-known studies point in the same direction.

The largest field study is on customer support. Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied 5,179 customer support agents who got a generative AI assistant in a staggered rollout. Productivity, measured as issues resolved per hour, rose about 14% on average. For novice and low-skilled agents the gain was about 34%, and the paper reports minimal impact on experienced and highly skilled workers. The paper was later published in the Quarterly Journal of Economics in 2025.

A controlled experiment on writing found the same pattern. Noy and Zhang gave 444 college-educated professionals occupation-specific writing tasks and exposed half of them to ChatGPT. Time taken fell by 0.8 standard deviations, output quality rose by 0.4, and the spread between workers narrowed because lower-ability workers benefited more.

And a study of 758 consultants at Boston Consulting Group, run with Harvard Business School researchers, found that below-average performers gained about 43% on tasks the AI handled well, while top performers gained less.

So yes: on well-defined, repeatable work, the AI narrows the gap between a newer and a more experienced worker. If your pay and promotion logic is built on "seniors do the standard work faster," that logic is eroding.

But look at who was studied and what they were doing. Customer support agents resolving chat tickets. Professionals on mid-level writing tasks. Consultants on realistic but self-contained exercises. These are populations and tasks where there's a recognisable good answer. They don't tell you what happens at the messy edge of a job.

The Edge of the Job Is Where Experience Pays

The same BCG study contains the finding that matters most for this question. Researchers designed one task to sit just outside what the AI could do well. Without AI, consultants got it right about 84.5% of the time. With AI, accuracy fell to 60% and 70% in the two AI groups. The AI's answer was confident, plausible and wrong, and many consultants accepted it.

The researchers called this the jagged frontier: AI is excellent at some tasks and unreliable at others that look similar, and the boundary isn't visible from the outside. Someone has to know where the edge is.

That's the job experience now does. A veteran support agent recognises the ticket that looks routine but isn't, the customer whose real problem is three layers below the stated one. A senior controller senses that a variance explanation reads well and still doesn't add up. A seasoned recruiter notices that the AI-ranked shortlist is quietly reproducing last year's bias. None of that is faster typing. It's pattern recognition built over thousands of cases, applied to the exact situations where a tool trained on typical cases fails.

Noy and Zhang's finding points the same way from another angle. In their experiment ChatGPT mostly substituted for worker effort rather than complementing worker skill, and it restructured tasks toward idea generation and editing and away from rough drafting. Editing is a judgment activity. The more of the first draft the machine writes, the more of the human's value lives in deciding whether it's right.

There's a second, quieter finding in the customer support paper. The authors offer suggestive evidence that the AI disseminates the best practices of more able workers and helps newer workers move down the experience curve. In other words, part of what looks like the AI "replacing" experience is the AI copying it. That's good news for novices. It also means the experience you can write down is the experience that becomes cheap, and the experience you can't write down gets scarcer relative to everything else.

Where the Premium Moves: A Four-Part Test

When a leader asks whether a senior role is worth its salary, I'd suggest replacing the question "is this person faster?" with a test of where their value actually sits. Call it the Exception Premium test. For each senior role, score how much of the job falls into each of four zones.

A role that's 80% routine execution is exposed. A role that's mostly exceptions, judgment and teaching has become more valuable per hour, not less, because each senior decision now governs a larger volume of AI-assisted output.

The test also tells you something uncomfortable: some of your senior people are paid for routine execution and have been coasting on tenure. AI makes that visible. Better to see it in your own review than to find it out when a competitor staffs the same function at a lower cost.

The Early-Career Evidence Is a Warning, Not a Contradiction

If AI helps novices most, you'd expect junior hiring to boom. The labour data suggests something more complicated.

A Stanford Digital Economy Lab study by Brynjolfsson, Chandar and Chen used payroll records from ADP covering millions of US workers. In its latest revision, employment among workers aged 22 to 25 in AI-exposed occupations is about 19% below where it would be had it kept pace with less-exposed peers. Experienced workers in the same occupations show no comparable gap. The decline is concentrated where AI mainly substitutes for human tasks, while employment is flat or rising where AI mainly complements workers. The gap comes largely from reduced hiring rather than layoffs. The authors are careful to call these early, descriptive indicators rather than causal estimates, and they find no evidence of economy-wide job displacement.

Read that alongside the productivity studies and a pattern emerges. AI makes a novice more productive on routine work, and that is exactly the work companies used to hire novices to do. Employers respond by hiring fewer novices. Meanwhile the experienced worker's edge in exceptions and judgment keeps paying.

This is the trap. If you stop hiring and developing juniors because AI covers their routine tasks, you're quietly cutting off the supply of future experts. The people who handle exceptions in 2032 are today's juniors, and they learn judgment by seeing thousands of ordinary cases first. We cover this dynamic in the broken bottom rung, and it's the main reason a pure cost-cutting reading of the research is dangerous.

Three Mistakes to Avoid

Mistake 1: Paying for tenure instead of exception skill. Years in the seat is a poor proxy. Two ten-year veterans can differ enormously in how much exception handling they actually do. Pay for demonstrated judgment, not birthdays.

Mistake 2: Assuming the novice plus AI equals the expert. The gains in the research are measured on tasks with clear answers. A novice with a good assistant can match an expert on a standard ticket and fail badly on a non-standard one, and may not know the difference. The BCG result is the cautionary tale: on the edge task, accuracy fell because people trusted confident but wrong answers. If you cut senior review layers because average output looks fine, you've removed the control exactly where it matters.

Mistake 3: Letting seniors stop teaching. If experience is where the premium now lives, then transferring it is the senior's most valuable output. Many organisations let that happen by accident through shadowing and hallway conversations, and AI shrinks the amount of shared routine work that used to create those moments. It has to become deliberate. The human premium and keeping human judgment sharp go deeper on why.

What to Do Next

For a CEO, COO or CHRO, the practical response is a short program, not a philosophy.

1. Map each senior role to the four zones. Ask managers to estimate, for the last month, what share of the senior's time went to routine execution, exceptions, review and teaching. A rough estimate is enough. You'll find a few roles that are mostly routine and a few that are almost all judgment.
2. Reprice by zone, not by tenure. Hold the premium for exception, review and teaching work. For roles that are mostly routine, stop adding seniority increments and look at the market rate for the task. Handle this transparently. The conversation is delicate, and compensation and fairness in the AI era covers how to keep it from turning into a trust problem.
3. Keep a junior pipeline on purpose. Even if AI covers the entry-level tasks, hire a smaller number of juniors and give them structured exposure to real cases, including the failures. This is the cheapest insurance against a future expert shortage. The thinking in upskill vs hire: the AI-native ROI case applies directly.
4. Build an "edge map" for each AI workflow. For every process where AI drafts or decides, write down the known failure cases and assign a senior reviewer to sample outputs. Treat the jagged frontier as a living document.
5. Change what you measure in performance reviews. Count exceptions resolved, errors caught before they reached a customer, and juniors who got better. Our piece on how performance reviews change with AI has a starting framework.
6. Tell people the logic. Experienced staff are anxious that AI makes them redundant and juniors are anxious that nobody will train them. A clear message that the company is paying for judgment and investing in developing it lowers both fears. See how leaders respond to AI job anxiety.

The Bottom Line

Experience is still worth paying for. But it's worth paying for different things than it was three years ago. The old bargain, where seniority bought speed and reliability on standard work, is being renegotiated by software that gives everyone a decent first draft. The new bargain is that seniority buys the judgment to handle what the software can't, the ability to spot when it's wrong, and the willingness to pass that on.

Companies that read the research as "experience is dead" will thin out their expertise just as the exceptions get costlier to miss. Companies that read it as "experience is unchanged" will keep overpaying for routine work that has gotten cheap. The leaders who come out ahead are the ones who can tell the two apart role by role.

Learn More

• The Broken Bottom Rung
• The Human Premium
• Keeping Human Judgment Sharp
• Compensation and Fairness in the AI Era
• Upskill vs Hire: The AI-Native ROI Case
• How Performance Reviews Change With AI
• Does AI Favor Generalists or Specialists?