LLM-First Policy Clarification and Institutional Knowledge Atrophy

LLM-First Policy Clarification and Institutional Knowledge Atrophy: Loss of Peer Consultation, Mentoring, and Long-Term Policy Expertise

2026-05-17 · llm-reasoning governance-policy workforce-skills knowledge-management organisational-learning agentic-ai organisational-design tools-infrastructure · medium · source → · wiki →

key claims

The evidence suggests that generative AI can displace some routine peer consultation by delivering codified expert practice directly to less experienced workers at the point of needBrynjolfsson et al. (2025)Forum (2025)
The same shift can weaken long-run skill formation when workers rely heavily on AI for unfamiliar tasks, because direct software evidence shows lower conceptual understanding, debugging, and independent error-correction after delegation-heavy useTamkin (2026)Macnamara et al. (2024)
Apprenticeship evidence suggests that peer clarification pathways carry institutional memory partly through modeling, coaching, scaffolding, and exposure to multiple experts, which AI answer delivery does not inherently preserveHolum (1991)
AI-first policy clarification is therefore more likely to centralize codified knowledge inside the tool while reducing the social interactions through which contextual judgment and exception handling are transmittedBrynjolfsson et al. (2025)Holum (1991)Tamkin (2026)
The long-run organisational risk combines weaker individual expertise with a feedback loop in which lower peer consultation produces less mentoring and leaves fewer people capable of challenging future AI interpretationsTamkin (2026)Mitchell (2026)Mitchell (2026)
Barger et al. suggest that a simulated AI coach can support some bounded coaching interactions in a single session, so the supported conclusion is conditional displacement of mentoring rather than universal replacement failureBarger et al. (2024)
The evidence supports a policy-assistant operating model where AI handles low-stakes retrieval and drafting but ambiguous, exceptional, or high-consequence interpretations still trigger human explanation, challenge, and escalationMedicine (2025)Forum (2025)Mitchell (2026)

Research Question

How does shifting from peer policy clarification to Large Language Model (LLM)-first interaction affect institutional memory transfer, mentoring, and long-term policy expertise?

Findings

Executive Summary

Shifting routine policy clarification from colleagues to an AI-first workflow is likely to reduce incidental mentoring and weaken long-run policy expertise unless organisations deliberately preserve human escalation and apprenticeship pathways. The strongest direct evidence shows a mixed pattern: generative AI can quickly diffuse codified best practices to newer workers, but heavier reliance during unfamiliar tasks reduces conceptual understanding and debugging ability rather than building durable expertise automatically. Apprenticeship evidence suggests that peer consultation matters because it transfers visible reasoning, contextual judgment, and multiple models of expert practice, not only final answers. For policy interpretation, the evidence supports augmentation rather than substitution. Use AI to accelerate low-stakes retrieval and drafting, but preserve colleague explanation, challenge, and escalation for ambiguous or high-consequence cases.

Key Findings

The evidence suggests that generative AI can displace some routine peer consultation by delivering codified expert practice directly to less experienced workers at the point of need.
The same shift can weaken long-run skill formation when workers rely heavily on AI for unfamiliar tasks, because direct software evidence shows lower conceptual understanding, debugging, and independent error-correction after delegation-heavy use.
Apprenticeship evidence suggests that peer clarification pathways carry institutional memory partly through modeling, coaching, scaffolding, and exposure to multiple experts, which AI answer delivery does not inherently preserve.
AI-first policy clarification is therefore more likely to centralize codified knowledge inside the tool while reducing the social interactions through which contextual judgment and exception handling are transmitted.
The long-run organisational risk combines weaker individual expertise with a feedback loop in which lower peer consultation produces less mentoring and leaves fewer people capable of challenging future AI interpretations.
Barger et al. suggest that a simulated AI coach can support some bounded coaching interactions in a single session, so the supported conclusion is conditional displacement of mentoring rather than universal replacement failure.
The evidence supports a policy-assistant operating model where AI handles low-stakes retrieval and drafting but ambiguous, exceptional, or high-consequence interpretations still trigger human explanation, challenge, and escalation.

Assumptions

Workplace advice-seeking evidence from customer support and organisational decision settings transfers directionally to policy-clarification work because both involve repeated questions, tool-mediated guidance, and judgment under uncertainty.
Institutional memory transfer in policy work depends partly on apprenticeship-style observation and explanation, even though the consulted literature does not measure policy teams directly.

Analysis

The evidence does not support a simple pro-AI or anti-AI story. Field evidence shows that generative AI can spread codified best practices and improve novice performance, which gives organisations a real incentive to keep more routine questions inside the tool. At the same time, the strongest direct skill-formation study shows that delegation during unfamiliar work lowers conceptual understanding and debugging ability, which means fewer peer interactions are not a free substitute for learning. The decisive mechanism is that peer clarification carries explanation, context, and staged responsibility, while AI-first clarification mainly carries fast codified output. A plausible rival explanation is that organisations could replace lost hallway mentoring with formal training or hybrid coaching, and the bounded coaching evidence suggests some of that substitution is possible. That rival does not remove the risk, because it requires deliberate organisational design; absent that design, AI-first convenience will tend to remove the social friction that previously triggered explanation, challenge, and escalation.

Risks, Gaps, and Uncertainties

No consulted study directly measures peer-policy-consultation frequency before and after AI-first policy-assistant adoption.
The coaching counter-example is based on a simulated future AI coach and one session, so it does not establish equivalence for long-term mentoring or institutional memory retention.
The institutional-memory conclusion is directionally well supported but still partly inferential because the consulted mentoring literature is broader than the specific policy-clarification use case.

Open Questions

How much does peer-consultation frequency actually change after policy-assistant rollout inside regulated organisations?
Which workflow designs preserve apprenticeship benefits while still capturing AI speed gains for routine interpretation?
What early warning signals best reveal that contextual policy expertise is thinning out before a control failure occurs?