Expert Agency And Apprenticeship Protection

Expert Agency & Apprenticeship Protection

Purpose

Addresses a risk that doesn't show up in any single project's success metrics but compounds quietly across years: if AI absorbs the junior-level tasks through which people normally build the pattern-recognition expertise of a senior professional, the organization loses its own future supply of people qualified to judge whether the AI is right. This skill does two things — draws a hard boundary around decisions that must stay with a human regardless of AI capability, and designs early-career roles so they still build real expertise even where AI now does most of the routine work.

Anchored in research

• The "apprenticeship risk" concept — junior professionals traditionally build expert judgment through repetitive, procedural work; if AI absorbs that layer without a deliberate redesign of how junior roles build expertise, the organization's future ability to supervise and validate AI erodes along with it. The source material attributes this concept to "Curt Strovink," but no identifiable person by that name could be confirmed in available search results — this attribution is dropped. The underlying concept is instead grounded in the independently well-documented 2026 discourse on this exact risk: American Recruiting & Consulting Group's analysis of AI "hollowing out the apprenticeship layer," SPARK6's "The Death of the Junior Analyst," and peer-reviewed work on novice-risk learning failures with emerging technologies (ScienceDirect, 2025) — see ../../references/hitl-partnership-heuristics-research.md for the full list.
• Skills-files-as-expertise-interface — the idea that expertise can be distilled into a machine-readable Markdown "skill file" that acts as a token-efficient interface, forcing an expert to make their process explicit rather than leaving it tacit. This is independently verifiable and current: Atlassian is a confirmed adopter of the open Agent Skills / SKILL.md standard (agentskills.io, opened December 2025), and has published its own design-system AI skills (an "ADS MCP server" plus detailed skill files) that its own team reports reduced AI token costs and improved output accuracy for their product builders. This is also, usefully, a direct validation of this repository's own methodology — this pack's skills are themselves an example of the pattern being described.

Method

1. Classify every AI-touched decision by stakes, not by task complexity. A task can be complex and still safely automatable if the cost of an error is low and reversible; a task can be simple and still require a human if the cost of an error is high or irreversible (see the same distinction already used in ../../../ai-strategy-and-governance/skills/task-level-decomposition-and-automation-fit/SKILL.md's error-tolerance criterion — apply it here specifically to the AGENCY question, not just the automation-fit question).
2. Name the decisions that must never be made autonomously by AI in this specific process — not as a generic policy statement, but as a concrete, named list: e.g. a final hiring decision, a clinical diagnosis, an irreversible financial commitment above a threshold, an action that affects someone's legal status. Write these into the routing design in ../hitl-maturity-and-confidence-routing/SKILL.md as hard escalation triggers regardless of confidence score.
3. Audit which junior-level tasks are being fully absorbed by AI in this organization, and for each one, ask explicitly: was this task also how juniors used to build the pattern recognition that lets a senior professional today catch an AI's mistake? If yes, removing the task entirely — rather than redesigning it — quietly consumes a training function the organization still depends on.
4. Redesign, don't just remove, the junior task where the answer to step 3 is yes. Options to consider, not a fixed recipe: keep the junior doing the task periodically alongside AI and comparing results (a structured, deliberate version of what ../hitl-override-metrics-and-feedback-audit/SKILL.md measures at the process level, but here used as a training exercise); rotate juniors through reviewing AI output specifically so pattern recognition builds from critique rather than from first-draft production; or preserve a smaller, deliberately-kept slice of manual practice even where full automation would be technically possible, treating it as a training investment rather than an inefficiency to eliminate.
5. Distill senior expertise into machine-readable skill files as a parallel mitigation — not a replacement for step 4, but a way to make tacit expert judgment explicit and inspectable, which has two effects: it gives the organization a token-efficient way to make expert reasoning available to AI tools (protecting experts' time from low-value manual repetition), and it forces the expert to articulate their own process clearly enough to write down, which is itself a form of documentation that survives their eventual departure. This is the exact pattern this skills pack itself demonstrates.
6. Revisit the classification and the audit periodically, not once. Both the stakes classification (step 1–2) and the junior-task audit (step 3) can go stale as AI capability and the organization's process both change — schedule a periodic review rather than treating this as a one-time setup exercise.
7. Produce a structured output: the named list of decisions requiring permanent human agency, the audit of at-risk junior tasks with a redesign recommendation for each, and (if applicable) which senior processes are candidates for skill-file distillation next.

What this skill does NOT do

• Doesn't make the hiring, workforce-sizing, or organizational-design decision for you — surfaces the risk and a structured way to think about it, the decision and its trade-offs stay with the organization.
• Doesn't replace ../../../ai-strategy-and-governance/skills/responsible-ai-and-governance-check/SKILL.md for regulatory human-oversight obligations — this skill's agency boundary is a design decision about organizational capability, not a compliance requirement (though the two will often overlap for high-stakes decisions).
• Doesn't guarantee a redesigned junior role will actually build the same expertise the old one did — this is a genuinely unsettled, actively debated area; treat any specific redesign as a hypothesis to validate with real outcomes, not a proven fix.
• Doesn't build the skill files themselves — see ../../../meta/skill_design_principles.md in this repo for what makes a good one, and treat this skill's own structure as a working example.

Refinement notes

Areas to keep deepening with real practice:

• your own examples of a junior-task redesign that measurably preserved expertise-building, vs. one that didn't
• a concrete stakes-classification template with worked examples across a few industries (into ../../references/)
• observations on how quickly pattern-recognition expertise actually degrades when a task is removed vs. redesigned — currently an assumption in the broader discourse, not something this pack has independently measured

Once this section is filled in and validated in practice, update the maturity field in skills_index.json to draft, validated, or canonical (see ../../../meta/maturity_levels.md). Don't add new fields to the frontmatter — name and description are the only ones allowed (see ../../../meta/frontmatter_schema.md).

Continue from here

• Individual-practitioner complement in this pack: ../self-efficacy-and-cognitive-rot-shielding/SKILL.md — this skill protects an organization's junior-expertise pipeline; that skill protects an individual practitioner's own judgment in their day-to-day work with AI. Use both — organizational design doesn't fix individual offloading, and vice versa.
• Related skill in this pack: ../hitl-maturity-and-confidence-routing/SKILL.md — where this skill's named "never autonomous" decisions become hard escalation triggers.
• Related skill in this pack: ../hitl-override-metrics-and-feedback-audit/SKILL.md — a healthy override rate depends on reviewers who still have the expertise this skill protects.
• Related skill in another pack: ../../../ai-strategy-and-governance/skills/task-level-decomposition-and-automation-fit/SKILL.md — the automation-fit classification this skill's stakes classification builds on.
• Related skill in another pack: ../../../ai-strategy-and-governance/skills/responsible-ai-and-governance-check/SKILL.md
• This pack's shared guardrails: ../../CLAUDE.md

References

• ../../references/hitl-partnership-heuristics-research.md — full sourcing and grounding-strength notes for this pack, including the dropped "Curt Strovink" attribution and its replacement sourcing
• ../../references/ — the pack's shared background material
• ../../CLAUDE.md — the pack's shared guardrails