Less-experienced workers often gain more from AI when the system provides reliable guidance for a structured task. Experienced workers can gain less, see no change, or slow down when they must verify context-sensitive output. The pattern is not universal: expertise still matters for judging errors, handling exceptions, framing problems, and using AI on work outside a model’s reliable range.
AI
Less-experienced workers often gain more from AI when the system provides reliable guidance for a structured task. Experienced workers can gain less, see no change, or slow down when they must verify context-sensitive output. The pattern is not universal: expertise still matters for judging errors, handling exceptions, framing problems, and using AI on work outside a model’s reliable range.
Key findings
• A customer-support deployment increased average productivity by about 15%, with roughly 30% gains among novice and lower-skilled agents.
• Highly skilled support agents did not improve in the same study, and their response quality fell slightly.
• In a P&G product-innovation experiment, less-experienced participants using AI reached performance comparable with experienced colleagues, while AI-assisted individuals matched the solution quality of two-person teams without AI.
• In a randomized study of experienced open-source developers, allowing early-2025 AI tools increased completion time by 19%.
• Workers and experts can misjudge the effect. Developers expected AI to make them 24% faster and still believed it had made them 20% faster after the measured slowdown.
• The practical conclusion is not “AI helps juniors and hurts experts.” It is that task structure, context, verification cost, and baseline expertise interact.
Methodology and definitions
This review compares field and randomized evidence from customer support, product development, and software maintenance. The outcome measures differ and should not be averaged.
Experience level can mean tenure, prior performance, task familiarity, domain knowledge, or technical skill. Studies operationalize it differently.
Productivity can mean issues resolved per hour, time to complete, output quality, or a combined score. A speed gain with lower quality is not equivalent to a quality-adjusted productivity gain.
Augmentation means AI helps a worker perform a task while the worker retains judgment and responsibility. Automation means more of the task is delegated. The boundary can vary within one workflow.
The sources include Stanford’s 2026 labor evidence synthesis, the Harvard Business School P&G field experiment, METR’s randomized developer trial, OECD workforce evidence, and PwC’s job-ad and company-financial analysis.
Customer support: the largest gains went to novices
Stanford’s 2026 evidence review on AI and jobs summarizes research on a generative AI assistant deployed to customer-support agents at a large call center.
Overall productivity increased by about 15%. The effect was concentrated among novice and lower-skilled workers, who improved the number of issues resolved per hour by roughly 30%. Highly skilled agents showed no productivity improvement, and their response quality fell slightly.
One interpretation is knowledge transfer. The assistant was trained on historical interactions and could surface patterns resembling the practices of high performers. New agents gained faster access to language, diagnostic steps, and resolutions that experienced agents had developed over time.
That interpretation should not be converted into a claim that AI replaces expertise. Experienced agents may already know the recommended answer, leaving little headroom. They may also handle harder cases where generic guidance is less useful. The small quality decline among highly skilled agents suggests that following AI recommendations can sometimes pull experts away from better judgment.
The study’s structure is favorable to measurement: support work has repeated cases, observable resolution rates, and a large historical knowledge base. Results may be weaker in roles where quality is delayed or subjective.
Product innovation: AI reduced expertise boundaries
The Cybernetic Teammate field experiment, published as a Harvard Business School working paper in 2025, studied 791 Procter & Gamble professionals. Participants worked individually or in cross-functional teams on real product-development challenges, with some groups using an internal GPT-4-powered tool.
AI reduced time by 16% for individuals and 13% for teams. Individuals with AI produced solutions comparable in quality with two-person teams without AI. Less-experienced participants achieved performance comparable with experienced colleagues.
AI also changed the shape of expertise. Research and development participants without AI tended toward technical ideas, while commercial staff tended toward market ideas. AI-assisted participants produced more balanced solutions that crossed those silos.
The highest-quality work still came from human teams using AI. Ideas ranking in the top 10% were three times more likely to come from AI-assisted teams than from individuals without AI.
This suggests two distinct operating models. An AI-assisted individual can provide efficient coverage for routine innovation work. A cross-functional team with AI may be preferable when the goal is a breakthrough solution and multiple forms of judgment matter.
The study occurred at one large company and focused on product innovation. It does not show that every junior employee can match a senior employee in negotiations, safety decisions, leadership, or relationship-dependent work.
Software maintenance: experienced workers became slower
METR’s randomized developer productivity study included 16 experienced open-source developers working on 246 real issues in mature repositories they regularly contributed to. Each task was randomly assigned to allow or disallow AI.
When AI was allowed, tasks took 19% longer. The confidence interval reported in METR’s later summary ranged from a 2% to 39% slowdown.
The setting differs sharply from the customer-support study. These developers had extensive context about large codebases. Their tasks required navigating tacit design decisions, dependencies, and repository conventions. AI suggestions could look plausible while creating verification and correction work.
The perception gap is equally important. Before randomization, developers forecast a 24% speedup. After participation, they still estimated a 20% speedup despite the measured slowdown. Self-assessment alone would have produced the opposite conclusion.
METR explicitly states that the result is a snapshot of early-2025 tools and does not represent most software development. In its February 2026 update, raw late-2025 estimates were consistent with small speedups, but selection effects, pay changes, multi-agent use, and time-measurement problems made the evidence unreliable.
A better model of skill and AI
The evidence supports a task-expertise matrix:
Task state: Structured task, novice worker; Likely AI effect: Larger potential gain; Main control: Teach verification and escalation
Task state: Structured task, expert worker; Likely AI effect: Smaller gain or quality risk; Main control: Let experts override and improve guidance
Task state: Ambiguous task, novice worker; Likely AI effect: Risk of accepting generic output; Main control: Require templates, sources, and review
Task state: Ambiguous task, expert worker; Likely AI effect: Potential ideation gain plus verification cost; Main control: Measure on real work, not perception
AI can compress the learning curve for known patterns. It does not remove the need to recognize when the pattern does not apply.
OECD’s 2025 SME workforce report found that twice as many GenAI-using SMEs said the technology increased the need for highly skilled workers as decreased it, 20% versus 9%. Data analysis and interpretation, creativity, and innovation were among the skills perceived to rise in importance.
PwC’s 2025 AI Jobs Barometer found skills requested in job advertisements changed 66% faster in highly AI-exposed occupations. Advertised AI skills carried an average 56% wage premium in 2024. Those are labor-market associations, not proof that a training course causes a 56% raise.
Practical SMB implications
Before deployment, complete an AI readiness audit for the workflow and the people performing it. Measure current performance by experience level where sample sizes allow.
Do not train only on prompts. Teach:
• What source material the AI can use.
• How to recognize unsupported claims.
• When to accept, revise, or reject output.
• Which cases require escalation.
• How to record a failure so the workflow can improve.
• Which actions are prohibited without approval.
Run evaluation by worker cohort. If juniors gain speed but error severity rises, tighten review. If experts slow down, test whether the tool should handle preparation, retrieval, or documentation rather than the core judgment they already perform efficiently.
Use experts as workflow designers and reviewers, not merely end users. Their corrections can define examples, rules, and exceptions. Protect independent skill by requiring workers to explain consequential decisions rather than accept output silently.
Measure quality-adjusted throughput, not activity. A support team sending more responses with more rework has not improved.
Limitations
The cited studies use different definitions of experience and productivity. Customer support, product innovation, and software maintenance have different feedback loops and error costs.
Tools changed between the studies. Small samples, especially METR’s 16 developers, limit generalization. PwC job-ad evidence is observational. The P&G study was conducted at one company.
What would change our view
We would revise the novice-advantage pattern if replicated studies with current tools found equal gains across skill levels after quality adjustment. We would become more concerned about deskilling if longitudinal research showed that AI-assisted juniors failed to build independent judgment or that experts’ performance deteriorated with sustained use.
FAQs
Do junior employees benefit more from AI?
Often in structured tasks, but not always. Customer-support research found larger gains among novice workers. Ambiguous tasks can expose juniors to greater risk if they cannot identify bad output.
Does AI make experts less productive?
It can in some settings. Experienced open-source developers took 19% longer with early-2025 tools in one randomized study. Other expert tasks may benefit, so measurement must be workflow-specific.
Why do experts gain less from AI?
Experts may already know the common solution, handle harder cases, and spend more time checking context-sensitive suggestions. Their baseline performance also leaves less room for improvement.
Can AI replace employee training?
No. AI can surface patterns and accelerate some learning, but employees still need domain knowledge, verification skill, escalation judgment, and practice without overreliance.
How should a company measure AI productivity by skill level?
Compare quality-adjusted completion time, rework, exceptions, and adoption across cohorts. Use real cases and avoid relying only on worker perceptions.
Get a 20-Minute AI Workflow Audit
AI Operator can identify which worker cohort and task type is likely to benefit, then design a quality-adjusted pilot with the right training and review boundary.