Generative AI boosts efficiency in routine, data-rich tasks but cannot reliably replace experience-based human judgment; organizations should adopt a five-part Human–AI Collaboration Framework to keep expertise and accountability central.

Main Finding

Current generative AI (notably large language models) cannot substitute for human judgment rooted in lived experience and contextual sense‑making. AI is most effective in routine, data‑rich tasks; it fails reliably in ambiguous, atypical, or context‑dependent situations. Organizations should therefore design workflows and governance to keep human expertise active and accountable while using AI to augment efficiency.

Key Points

• Distinct cognitive processes: Human experts use accumulated, embodied experience to interpret subtle cues, infer intent, and make sense of novelty; LLMs operate by pattern‑matching across statistical regularities and lack the same lived, situational grounding.
• Recurring failure modes when AI is applied to expert work:
  • Errors on unusual or out‑of‑distribution cases.
  • Missed contextual signals that are tacit or not captured in training data.
  • Over‑reliance on AI recommendations by workers and managers (automation bias).
  • Reduced visibility of real skill differences among employees when AI standardizes outputs, complicating assessment and learning.
• Cross‑domain evidence: Studies in hiring, performance management, healthcare, and knowledge work show similar patterns of benefit in routine tasks but harms or risks where contextual judgment is required.
• Practical response: the authors propose a five‑part Human–AI Collaboration Framework to preserve human judgment and accountability while capturing efficiency gains from AI.

Data & Methods

• Approach: interdisciplinary synthesis drawing on cognitive science, neuroscience, and organizational studies combined with empirical evidence from recent applied studies.
• Evidence sources: laboratory experiments, field studies, audits, and observational/case studies from domains including:
  • Hiring processes (automated résumé screening and interview summarization).
  • Performance management (algorithmic performance ratings and recommendations).
  • Healthcare (clinical decision support and diagnostic assistance).
  • Knowledge work (AI assistants used for drafting, summarization, and decision support).
• Analytical strategy: identify recurring patterns of AI failure and human–AI interaction problems; map these patterns to cognitive mechanisms (e.g., tacit knowledge, context sensitivity) and organizational processes (e.g., accountability, skill assessment); derive a prescriptive framework for collaboration.

Implications for AI Economics

• Task‑level complementarity: The paper reinforces a tasks‑based view of labor substitution — AI substitutes in routine, codifiable tasks but complements (and cannot replace) tasks requiring contextualized human expertise. Economic models should account for heterogeneous task susceptibility to automation.
• Value of firm‑specific human capital: Because experience and tacit knowledge matter, organizations’ returns to investing in employee experience and on‑the‑job learning remain high. AI that erodes opportunities for experience accumulation or masks skill differences can reduce incentives to invest in human capital.
• Labor demand and skill composition: Demand may shift away from purely procedural work toward roles that combine domain experience, sense‑making, and oversight of AI systems (e.g., curators, auditors, escalation experts). Wages and hiring signals will increasingly reflect this complementary skill bundle.
• Productivity measurement and incentives: Standardizing outputs through AI can inflate measured productivity while obscuring individual contributions and learning trajectories, complicating performance pay and promotion decisions. Firms should adjust measurement systems to preserve visibility of true skills and outcomes.
• Governance and policy: The need for auditability, provenance, uncertainty disclosure, and accountability in AI deployment has economic consequences (compliance costs, procurement standards). Regulators and firms should focus on ensuring human oversight for high‑stakes, context‑sensitive decisions.
• Research directions for AI economics:
  • Quantify task‑level elasticities of substitution between AI and different types of human capital.
  • Study long‑run effects of AI adoption on accumulation of firm‑specific experience and on career trajectories.
  • Evaluate the optimal design of incentives and performance metrics when AI amplifies or masks worker output.
  • Model organizational investments in training, interfaces, and governance that preserve complementarity and mitigate automation bias.

(If you want, I can expand the five‑part Human–AI Collaboration Framework into concrete components and implementation steps tailored to a specific domain—e.g., hiring or healthcare.)