Expertise is an activity. It persists through practice, through apprenticeship, through evaluation against internalized standards. Stop doing any one of those and the expertise doesn't vanish overnight. It thins. Quietly, over quarters, in ways that don't register on any dashboard anyone is watching.
Agentic AI disrupts all three at once. And it's doing so at a pace that outstrips most organizations' capacity to notice. The deployment confidence is high. The integration maturity lags far behind.
Routine execution gets automated first, because that's where the ROI case is simplest. Apprenticeship pipelines thin as entry-level roles contract. Postings for junior positions fell 35% between January 2023 and June 2025, with AI-exposed roles hit hardest. And evaluation degrades as senior reviewers spend more time assessing AI outputs than producing work themselves, gradually losing the internal baseline that made their judgment trustworthy.
Each of these alone would be manageable. Together they compound. But follow the compounding one more step and the problem folds in on itself: the ability to detect the compounding is subject to the same decay.
A multicenter study of over 23,000 colonoscopy procedures, documented in a 2026 scoping review, found that after routine AI-assisted practice, endoscopists' detection rates dropped from 28.4% to 22.4% when working without AI. They had been better before AI arrived. During normal operations the numbers looked fine, because during normal operations the AI was there. The degradation only surfaced when someone thought to remove the support and measure what remained. Nobody's quarterly review catches a problem that only exists when the tool is absent.
In most organizations, nobody is designing that test. The analyst whose judgment was built through years of manually cross-referencing sources has that judgment slowly replaced by a review workflow. The junior who would have developed similar judgment through operational mistakes, through recoverable errors made under supervision, never gets hired into the role. The senior leader evaluating the team's output has fewer and fewer reference points for what "good" looks like independent of the AI's contribution. Performance metrics were designed for a world where human expertise was the substrate, not a variable being quietly consumed.
The dashboards stay green. The system is working.
The METR developer productivity study found that experienced open-source developers using AI tools believed they were 20% faster. They were 19% slower. The error pointed in exactly the direction that would prevent correction.
Building the measurement infrastructure to catch expertise erosion requires the expertise that's eroding. Designing evaluation benchmarks, identifying what independent judgment looks like in a specific domain, knowing which edge cases reveal real understanding. All of it depends on practitioners who still carry the old baseline in their heads. That population is aging, retiring, being promoted out of operational work, or simply losing the sharpness that came from daily practice. NASA researchers studying pilots found the same pattern: motor skills held up surprisingly well under automation, but cognitive skills eroded first. The ability to stay alert, to notice when something had gone wrong. Exactly the skills you'd need to catch the erosion.
Every quarter that passes makes the next quarter's assessment harder to construct. The window to build these systems is the window during which the people who could build them still have the knowledge to do it. And that window is closing on the same curve as the problem it would address.
Ordinary organizational risk announces itself eventually. This one seals itself shut.
Things to follow up on...
- Who reviews the reviewers: A randomized experiment with 2,784 participants found that evaluators' prior disposition toward AI, not their training or expertise, determined how reliably they caught errors in AI outputs, raising questions about what happens as the reviewer population shifts generationally.
- The pipeline is contracting: US programmer employment fell 27.5% between 2023 and 2025 according to BLS data, while software developer roles declined only 0.3%, a divergence IEEE Spectrum explored as evidence that codifiable work is disappearing faster than design-oriented work.
- Nine percent have real autonomy: ServiceNow's 2026 Enterprise AI Maturity Index found that while 59% of organizations have moved beyond piloting agentic AI, only 9% have built meaningful autonomous workflows, suggesting the gap between deployment confidence and operational integration is wider than most leaders assume.
- Only 6% are redesigning: Deloitte's 2026 Global Human Capital Trends report found that just 6% of leaders say they're making real progress designing how humans and AI should work together, a finding the Microsoft Work Trend Index frames as the "Transformation Paradox" of organizations adopting tools without rethinking the structures around them.