AI raises productivity only when it functions as a calibrated teammate and humans are trained to collaborate; without shared understanding, flexible communication and calibrated trust, firms risk errors and wasted investment. Cross‑training and co‑learning systems appear most promising, but heterogeneous, mostly small‑scale evidence limits precise, generalizable conclusions.

Main Finding

Human-Agent Teaming (HAT) effectiveness depends on mutual readiness: AI needs adaptive, metacognitive, and communicative capabilities and humans need regulatory capacities (trust calibration, metacognitive awareness) and operational team competencies (communication, coordination, adaptability). Persistent barriers—communication inflexibility, limited shared mental models, and trust miscalibration—limit reliable collaboration, while cross-training and co-learning approaches show the most promise for building calibrated, productive HAT. Methodological heterogeneity in the literature constrains precise, generalizable conclusions.

Key Points

• Two complementary readiness levels matter:
  • Operational team competencies: communication, coordination, adaptability.
  • Regulatory capacities: trust calibration and metacognitive awareness.
• Recurring barriers to effective HAT:
  • Communication inflexibility in AI (limited ability to adapt explanations/behavior to human partners).
  • Limited shared understanding / misaligned mental models between humans and AI.
  • Trust miscalibration (under- or over-trust) leading to misuse or disuse of AI support.
• Regulatory capacities are especially critical but under-specified and under-measured in empirical work.
• Promising interventions:
  • Cross-training (humans learn aspects of the AI’s role; AI is trained on human behaviors) to build shared representations and expectations.
  • Co-learning systems that adapt dynamically to individuals and teams.
  • Design principles: calibrated transparency/explainability, adjustable autonomy, real-time feedback, and flexible communication modalities.
• Human readiness strategies:
  • Metacognitive training to improve situation awareness and error detection.
  • Trust-calibration exercises (simulations with varied reliability) and interface cues that reflect system uncertainty.
  • Team-training focusing on coordination with AI teammates (roles, handoffs, exception handling).
• Research gaps and limitations:
  • Heterogeneous methods and measures across studies make synthesis difficult.
  • Lack of standardized operationalization of trust calibration and metacognitive readiness.
  • Few large-scale longitudinal field studies on HAT outcomes.

Data & Methods

• Review type: Structured narrative review (not a meta-analysis).
• Time frame: 2010 through January 2026.
• Databases searched: Google Scholar, Scopus, PsycINFO, IEEE Xplore, ACM Digital Library, Semantic Scholar; supplemented with forward citation tracking.
• Screening: 572 records screened; 192 articles included in the final synthesis.
• Inclusion focus: empirical and conceptual work on human-agent teaming, team competencies, trust calibration, metacognition, communication, coordination, adaptability, cross-training/co-learning interventions.
• Limitations of the review methodology:
  • Narrative synthesis due to heterogeneity of study designs, tasks, metrics.
  • Possible publication and selection biases inherent to database and citation-search approaches.
  • Limited capacity to quantify effect sizes or produce pooled estimates of intervention impacts.

Implications for AI Economics

• Productivity and value creation:
  • Gains from AI depend on effective HAT readiness; technological capability alone will not fully translate into productivity unless communication, shared understanding, and trust are addressed.
  • Investments in co-learning-capable systems and workforce training are likely to have higher returns than investments in one-sided automation.
• Labor demand and task allocation:
  • AI systems that function as calibrated teammates augment human roles when humans are trained to supervise, interpret, and collaborate; poorly prepared teams may exacerbate task inefficiencies or de-skill workers.
  • Cross-training can shift the frontier of comparative advantage—workers may take on higher-value coordination and supervisory tasks while AI handles routine execution.
• Adoption, diffusion, and firm strategy:
  • Firms face adoption risk: miscalibrated trust or communications failures can produce costly errors and reputational losses, raising adoption thresholds.
  • Procurement decisions should value systems’ ability to support co-learning, transparent uncertainty communication, and adjustable autonomy—not just raw performance metrics.
• Cost and investment considerations:
  • Upfront costs for metacognitive/trust training, interface design, and co-learning infrastructure are investments that reduce downstream error costs and increase reliability.
  • Economic evaluations should include non-production benefits such as resilience, reduced error-related losses, and reduced oversight burden.
• Measurement & regulation:
  • Need for standardized metrics (e.g., calibrated trust indices, measures of shared mental model alignment, HAT reliability) to evaluate ROI and compare systems.
  • Policy/regulatory frameworks could incentivize explainability, uncertainty reporting, and human-centric teaming features to mitigate systemic risks.
• Externalities and market effects:
  • Systemic under-preparedness could create negative externalities (safety incidents, liability exposure) that spill across firms and sectors; coordinated standards and training programs can internalize these risks.
• Research and data priorities for economists:
  • Longitudinal, field-level studies that measure productivity, error rates, labor reallocation, and training returns under different HAT readiness interventions.
  • Cost–benefit analyses comparing investment in human readiness (training, team protocols) versus incremental AI capability improvements.
  • Modeling complementarity vs. substitution dynamics conditional on levels of human HAT readiness and regulatory capacities.

Overall, economic benefits from AI-as-teammate depend critically on complementary investments in human readiness and system design that enable calibrated, adaptive collaboration.