AI in finance is evolving from prediction engines to partners in decision-making, reshaping workflows, prices and credit allocation; its benefits hinge on how firms allocate authority, oversight and incentives across human–AI workflows rather than on raw model accuracy.

Main Finding

AI in finance is shifting from standalone predictive tools to components of human–AI hybrid financial decision systems. Outcomes depend less on model accuracy alone and more on the decision architecture that allocates authority, shapes reliance, provides actionable explanations, enables meaningful oversight, and manages reflexive feedback effects. Human–AI complementarity in finance is conditional, not automatic.

Key Points

• Unit of analysis shift
  • Move from evaluating isolated models to evaluating sociotechnical decision systems in which humans, algorithms, organizations, markets, and regulators jointly produce outcomes.
• Five connected constructs (integrated decision-system framework)
  • Delegation frontier: where decision authority shifts between humans and AI across information acquisition, recommendation, approval, execution, monitoring, learning.
  • Reliance wedge: the gap between actual reliance on AI and warranted reliance given model accuracy, uncertainty, incentives, and accountability.
  • Decision-useful XAI: explanations designed to improve actionability, auditability, contestability, and accountability (not just perceived transparency).
  • Meaningful oversight: human supervision that has the information, authority, time, independence, and responsibility to affect outcomes.
  • Reflexive AI loop: how AI-mediated decisions change behaviours, organizational routines, market conditions, and the data generating process for future decisions.
• Evolution of methods and roles
  • Traditional ML (scoring/ranking) → scalable case scoring, easier validation.
  • Deep learning → richer representation for sequences/text/multimodal data but raises interpretability and robustness needs.
  • Reinforcement learning → policy/action outputs that force delegation and control design (limits, kill-switches, monitoring).
  • FinLLMs / generative & agentic AI → language + tool use + agentic workflows; produce AI-generated decision artifacts that influence downstream human judgments and actions.
• Conditional complementarity
  • Hybrid performance hinges on task structure (routine vs. interpretive), presence of private information, feedback quality/frequency, incentives and accountability, explanation design, and governance arrangements.
• Reflexivity & systemic effects
  • AI-mediated choices alter prices, liquidity, credit allocation and produce strategic responses (e.g., gaming, selection effects), complicating evaluation and possibly creating externalities or stability risks.
• Practical tension
  • Strong standalone models can worsen decisions if decision rights, oversight, or explanations are mis-specified; conversely, modest models can improve outcomes if embedded in well-designed decision architectures.

Data & Methods

• Article type and scope
  • Integrative, conceptual review synthesizing literature across finance, management, human–computer interaction (HCI), explainable AI, governance, and FinLLMs. Not a systematic enumeration of every method.
• Empirical and methodological building blocks surveyed
  • Empirical domains: lending, trading, portfolio management, advisory work, compliance, insurance, financial supervision.
  • ML methods: tree ensembles, boosting, SVMs (tabular scoring); deep networks (LSTM, CNN) for sequences and order books; RL/DRL for sequential control; transformer-based FinLLMs and tool-using agents for language & workflow.
  • Evaluation approaches emphasized: backtesting and simulated environments (for RL), robustness/regime-sensitivity checks, challenger model frameworks, stress-testing, audit trails, and live monitoring.
  • Human–AI interaction evidence: lab and field studies on reliance, trust, and override behavior (e.g., calibration studies showing both under- and over-reliance), HCI research on explanation design, and governance/case studies on oversight failures.
• Recommended empirical strategies (implied)
  • Randomized delegation experiments, difference-in-differences, RCTs in operational settings, audit-and-monitor trails, structural dynamic models and multi-agent simulations to capture reflexivity and endogenous responses.

Implications for AI Economics

• Change in evaluation targets
  • Economic evaluation should move from pure predictive accuracy metrics to decision-outcome metrics: welfare, distributional impacts, error externalities, accountability costs, and system-level stability.
• Modeling reflexivity and endogeneity
  • Economists must model how AI-mediated policies change agent behaviour and the data generating process (selection, strategic adaptation, market impact). Static estimands are often invalid; dynamic/structural approaches and multi-agent models are needed.
• Incentives, delegation, and mechanism design
  • Allocation of decision rights (delegation frontier) and incentive structures determine whether AI improves welfare. Mechanism design and contract theory approaches can help design delegation rules, override penalties, and monitoring incentives.
• Measurement and identification challenges
  • Endogenous deployment of AI, feedback loops, and strategic gaming complicate causal identification. Use of randomized rollouts, instrumental variables tied to exogenous variation in deployment or interface, and carefully designed field experiments are critical.
• Policy and regulatory focus
  • Regulators need to mandate decision-useful XAI, audit trails, and meaningful human oversight rather than only transparency of model internals. Supervision should account for system-level risks (market liquidity, herding, procyclicality) from widespread agentic AI adoption.
• Distributional and welfare concerns
  • AI in credit, pricing, and compliance reallocates opportunities and scrutiny. Research should quantify who gains/loses, how bias and fairness interact with delegation, and costs of contestability and error remediation.
• Research agenda priorities
  • Empirically measure the delegation frontier and reliance wedge across settings.
  • Develop and test decision-useful XAI interfaces that improve auditability and contestability.
  • Design field experiments on delegation rules, override protocols, and oversight structures.
  • Build structural models capturing reflexive AI loops and estimate welfare consequences and systemic risk.
  • Study the macro/market implications of agentic FinLLMs and tool-using trading agents at scale.

Summary takeaway: AI’s economic impact in finance depends on decision-system design—who acts, who can override, how explanations and oversight function, and how AI-mediated actions reshape markets and data. For economists this means shifting methods and metrics toward interaction-aware, dynamic, and institutional analyses that capture incentives, reflexivity, and distributional effects.