This sixth article in the AI Contradictions series, moves the focus from how AI changes engineering work to how it changes the formation of engineering capability. It asks what happens when automation removes the task before management has decided whether the capability built through that task is still needed.
| Audience: | CTO đźž„ CIO đźž„ Director of IT Strategy |
| Primary Sectors: | All Industry Sectors |
| Decision Horizon: | Before AI-driven role redesign, staffing assumptions, or progression changes remove recurring engineering work. |
Executive Summary
AI can improve developer output without proving that the capability previously developed through that work is still being formed. The contradiction is that AI gives management better evidence that work can be removed before it gives management equivalent evidence that the capability produced through that work can be removed too.
Decision Posture: When AI-enabled productivity is being used to justify a material change in headcount, role scope, progression paths, or the engineering competency model, require the engineering owner to classify the displaced capability as obsolete, replaced, or still required. Remove the old work when the capability is obsolete or another mechanism credibly develops it. If the capability remains necessary and its formation path disappears, do not treat the workforce change as uncomplicated productivity gain.
The reason for the default rule is decision asymmetry. Task-level productivity becomes visible quickly; the consequences of removing a capability-building pathway may take years to surface. Workforce decisions can be made on the first signal before the second exists.
Our Analysis
Current research does not prove that AI is hollowing out engineering apprenticeship. It establishes the more useful management problem: AI-assisted performance and capability formation are different measures, while management decisions can easily treat them as the same one.
The Narrative vs The Reality
The productivity narrative is straightforward: coding assistants automate lower-level work, less-experienced engineers get faster, and senior engineers shift toward design and supervision. If AI can handle yesterday’s apprenticeship tasks, preserving those tasks can look like protecting inefficiency.
Three realities complicate that logic.
- The productivity benefit is real, and can be larger for less-experienced developers. Three randomized field experiments covering 4,867 professional developers found AI access increased completed tasks by 26.08% overall, with greater adoption and productivity gains among less-experienced developers.1
- Better assisted performance does not reliably demonstrate more underlying capability. In a randomized programming study of 275 students, both scaffolded AI and unrestricted ChatGPT substantially improved exercise performance, but neither produced greater knowledge gains or code-comprehension gains than the control.2
- The problem is not AI use itself. A 2026 systematic review identified 85 eligible STEM studies and found substantial heterogeneity; after publication-bias correction, the overall positive learning effect became uncertain, while differences associated with how AI altered the learner’s cognitive activity remained. The authors conclude that AI appears more promising when it augments activity than when it substitutes for it.3 A separate meta-analysis of 32 programming-education studies found positive effects on both post-test and practice performance.4
The Signal in the Noise
The management risk is not that “AI makes juniors worse,” it is using evidence that a task can be delegated as evidence that the capability behind it can safely disappear.
What Changes the Decision
Determine whenever AI automates a task whether the organization still requires the capability that task helped produce.
Tactive’s recommended control is to classify the displaced capability as either obsolete, replaced, or still required before AI productivity assumptions become material workforce changes.
- Obsolete: Remove the work and the associated learning requirement.
- Replaced: Remove the old work once another credible formation mechanism exists.
- Still required: Do not remove the capability from the role or progression model until its formation path has been replaced.
This gate should not apply to ordinary task automation. Trigger it when AI changes headcount assumptions, role scope, progression paths, or the competency model.
Why This Matters Now
Organizations do not need to wait for evidence of an apprenticeship crisis before confronting this decision. A developer-specific randomized 2026 article preprint found that developers using AI while learning an unfamiliar library performed worse on subsequent conceptual-understanding, code-reading, and debugging measures on average. Meanwhile cognitively engaged AI-use patterns preserved learning better.5 It is supporting evidence of the mechanism, not proof of long-run enterprise effects.
Nor should current hiring weakness be treated as proof of AI-driven apprenticeship erosion. LinkedIn’s August 2026 analysis says the broader entry-level slowdown across five markets is more consistent with economic uncertainty than AI. It nevertheless finds junior hiring running 3-10 percentage points below overall hiring within its broad category of AI-augmented occupations, which includes software engineering.6 That is workforce context, not capability evidence. The urgency therefore comes from timing. Organizations can redesign roles and progression paths now; reliable evidence about the multi-year capability consequences will arrive later.
What to Watch for Next
Progression models change faster than competency models, and stronger evidence that AI-native tutoring, simulation, or assessment can replace older apprenticeship loops. Where replacement works, retire the old work.
Recommended Actions
Do This
- Install the gate at workforce-model decisions, not at every automated task. When AI productivity is used to justify a material change in headcount, role scope, progression paths, or the competency model, require the VP/Head of Engineering or service owner to attach a lightweight capability disposition: obsolete / replaced / still required. A “still required” capability without a replacement formation mechanism blocks permanent removal from the role or progression model, not use of AI itself.
- Test for residual capability, not the obsolete task. For capabilities marked “still required,” define a proof point around the condition in which the organization still expects human judgment: diagnosing an AI-generated defect, recovering a failed service, challenging a faulty recommendation, or explaining a design tradeoff. Do not recreate the manual work AI has replaced merely to test whether someone can still do it.
- Put capability replacement into the AI business case. When projected AI productivity supports headcount assumptions, narrower entry-level scope, or removal of progression work, require the CIO and Finance to identify the cost and owner of any replacement formation mechanism before recognizing the change as sustainable labour productivity. If no replacement is needed, management should be able to name the capability that has become obsolete.
Avoid This
- Preserving manual work because it was once educational. If the capability is obsolete, eliminate both the task and its apprenticeship burden.
- Treating AI-assisted throughput as evidence of competence. Performance measures tell management whether work was completed; they do not necessarily establish that the underlying judgment was acquired.2,3
- Imposing a “juniors use less AI” rule. Less-experienced developers can benefit substantially from AI, and AI itself can support learning when the interaction preserves the cognitive work that matters.1,3,4
Bottom Line
AI can prove that a task is automatable before management knows whether its underlying capability is dispensable. Retire the work only when the capability is gone or replaced, not simply because AI can now perform it.
Evidence and Sources
- Cui, Kevin Zheyuan, Mert Demirer, Sonia Jaffe, Leon Musolff, Sida Peng, and Tobias Salz. 2026. The Effects of Generative AI on High-Skilled Work: Evidence from Three Field Experiments with Software Developers. Management Science. The Effects of Generative AI on High-Skilled Work
- Bassner, Patrick, Ben Lenk-Ostendorf, Ramona Beinstingel, Tobias Wasner, and Stephan Krusche. 2026. Less Stress, Better Scores, Same Learning: The Dissociation of Performance and Learning in AI-Supported Programming Education. Computers & Education: Artificial Intelligence 10: 100537. Less Stress, Better Scores, Same Learning
- Boolzen, Chiara, Jochen Kuhn, Salome Flegr, Eva-Maria Rott, Niklas Stausberg, and others. 2026. Evidence of Impact and Interpretational Limits of Generative AI in STEM Education: A Systematic Review and Meta-Analysis on Cognitive Learning Outcomes. Artificial Intelligence Review. Evidence of Impact and Interpretational Limits of Generative AI in STEM Education
- Deng, Hongji, Hui Chen, and Yan Dong. 2026. Do AI Chatbots Improve Students’ Learning Performance in Programming Education? Evidence from a Meta-Analysis. Journal of Educational Computing Research 64 (5). Do AI Chatbots Improve Students’ Learning Performance in Programming Education?
- Shen, Judy Hanwen, and Alex Tamkin. 2026. How AI Impacts Skill Formation. arXiv preprint. Used as supporting developer-specific evidence rather than the primary foundation for the recommendation. How AI Impacts Skill Formation
- Hood, Rosie. 2026. AI Labor Market Update — August 2026. LinkedIn Economic Graph Research Institute. Covers April–June 2026 across the US, UK, France, Germany, and India. AI Labor Market Update — August 2026
Learn More @ Tactive
- Before Coding Agents Get More Authority, Find or Enforce the Constraints They Cannot Miss
- Mandating AI Changes what the Adoption Metric Really Means
- The First Release Is Too Early to Declare AI Code Good
- AI Coding Has Made Qualified Review the Scarce Resource
- AI Has Made Coding Cheaper, but Software Ownership More Expensive
- AI Coding Is Not a Productivity Story Yet