Generative AI is boosting productivity in routine, information-heavy business decisions—especially among less-experienced workers—but can harm outcomes when applied outside its capability frontier, meaning firms need careful workflow redesign and governance.

Main Finding

Generative AI (GenAI) is currently most valuable as an augmenting prediction-and-synthesis technology for business decision-making rather than a replacement for human judgment. Evidence shows consistent productivity gains in well‑specified, information‑heavy tasks (customer operations, marketing, software engineering, R&D), concentrated among lower-skilled or lower-performing workers. However, GenAI can degrade decision quality when applied outside a model’s true capability — a “jagged technological frontier” that creates risk of a false sense of confidence and automation bias. Mapping AI capability boundaries at the task level is the central open research and implementation priority.

Key Points

• Augmentation not automation: GenAI reduces the cost of prediction and information synthesis, increasing the returns to human judgment rather than eliminating it (prediction-machine framing).
• Productivity evidence (selected results):
  • ~14% increase in issues resolved per hour in customer support (Brynjolfsson, Li & Raymond).
  • +12.2% tasks completed and +40% quality for in‑frontier consulting tasks using GPT‑4 (Dell’Acqua et al.).
  • ~55.8% faster task completion in coding with GitHub Copilot (Peng et al.).
  • Writing and other tasks show larger gains among lower‑baseline performers (Noy & Zhang).
• Heterogeneous effects by skill: largest benefits accrue to newer or lower-skilled workers; GenAI tends to compress performance distributions (knowledge‑leveling).
• Capability frontier risk: when tasks exceed the model’s effective boundary, AI assistance can mislead users, producing worse outcomes than no-AI workflows (false sense of confidence / automation bias).
• Organizational integration matters: successful implementations redesign decision workflows to fit AI strengths and set governance, oversight, and verification processes.
• Macroeconomic potential: large aggregate value estimates (trillions) concentrated in language- and synthesis-heavy functions, but these are conditional on correct task matching, governance, and adoption.
• Ethical and governance constraints: explicability, accountability, misinformation/IP risks, and the need for oversight are critical for business decision contexts.

Data & Methods

• Study type: Qualitative, descriptive, comparative literature synthesis (peer‑reviewed field experiments, lab experiments, organizational theory, and macroeconomic/industry estimates).
• Comparison axes used by the authors:
  • Business function (customer ops, consulting/strategy, software engineering, cross‑functional).
  • Evidence type (large‑scale field experiment, controlled lab experiment, survey, macro estimation).
  • Reported productivity/quality effects and distribution across participant skill levels.
  • Whether studies explicitly consider capability‑frontier or task suitability risk.
• Representative sources summarized:
  • Field experiments: large customer‑support RCT (Brynjolfsson et al.), BCG consultant field trial (Dell’Acqua et al.).
  • Controlled experiments: coding (GitHub Copilot, Peng et al.), professional writing (Noy & Zhang).
  • Macroeconomic and exposure estimates: McKinsey (Chui et al., Bughin et al.), occupation exposure (Eloundou et al.).
  • Organizational and ethical frameworks: decision‑structure typologies, AI governance and explicability literature.
• Outputs: comparative tables (productivity by function; value potential vs. exposure; task‑suitability risk matrix) and synthesis identifying research gaps.
• Limitations noted by the authors: review is qualitative rather than a formal systematic meta‑analysis; task‑level boundaries remain under‑mapped; many macro estimates are conditional and heterogeneous.

Implications for AI Economics

• Task‑level mapping is essential: economic estimates and firm strategies must move from function‑level exposure to task‑level capability matching to avoid overestimating realized gains and to manage risk.
• Distributional labor effects: because lower‑skill/less‑experienced workers gain most, GenAI is likely to compress within‑occupation performance (short‑run leveling) while exposing high‑skill, information‑intensive occupations to larger task reallocation pressures.
• Complementarity reinforces human judgment value: cheaper prediction raises demand for judgment and accountability; firms and markets will value skills in interpreting, verifying, and integrating AI outputs.
• Policy and firm strategy: invest in worker training, oversight systems, and workflow redesign to capture productivity gains safely; regulatory focus on explicability and accountability will shape adoption and realized economic value.
• Measurement and research priorities for economists:
  • Quantify the jagged frontier: identify which tasks within high‑exposure functions are reliably automatable vs. augmentable.
  • Heterogeneous treatment effects: precise estimates by worker skill, task complexity, and organizational context.
  • Dynamic effects: long‑run impacts on wages, skill composition, and productivity growth conditional on different governance/implementation regimes.
  • Externalities and systemic risk: evaluate misinformation, legal/liability, and competition effects when AI‑generated analyses enter high‑stakes decisions.
• Bottom line for AI economics: large potential gains exist but are conditional. Realized macroeconomic impact depends critically on task suitability, governance, human‑AI complementarities, and how firms and policymakers manage the jagged capability frontier.