Are AI Agents Ready to Manage Your Money? A Critical Look at the Economic Rationality of Large Language Models
"New research benchmarks the decision-making skills of LLMs, revealing surprising gaps in their ability to handle complex financial scenarios."
Imagine a future where AI agents handle your personal finances, make investment decisions, and even negotiate on your behalf. This vision is fueled by the rapid advancement of Large Language Models (LLMs), which are increasingly being touted as capable decision-makers. However, before we entrust our economic well-being to these digital entities, a crucial question arises: are LLMs truly rational enough to handle the complexities of the financial world?
Recent research has explored leveraging LLMs to create decision-making engines, configuring them either to act directly as economic agents or to serve as key elements of broader systems. LLM-based agents are already showing strength in planning, solving complex problems, leveraging tools, and playing games. However, assessing their economic rationality is a different ballgame.
To address this concern, a team of researchers has developed a novel benchmark called STEER (Systematic and Tuneable Evaluation of Economic Rationality) to rigorously assess the economic rationality of LLMs. This benchmark draws upon established economic principles and cognitive psychology to evaluate LLMs across a wide range of decision-making scenarios.
Assessing LLM Economic Rationality
Researchers have developed benchmarks to evaluate whether large language models exhibit economically rational behavior. The STEER framework, introduced in early 2024, provides a methodology for assessing LLM agents' economic rationality across complex tasks requiring logic, probability, optimization, and reasoning about human behavior. The framework underscores the pivotal role of Reinforcement Learning from Human Feedback (RLHF) in augmenting LLMs' rationality and decision-making capabilities. These assessment tools are critical for determining whether LLM agents can be trusted with consequential decisions in economic and financial contexts.
Traditional Evaluation of AI Rationality
Assessing the rationality of artificial agents has historically relied on frameworks borrowed from classical economics, where rationality is defined as consistent utility maximization. However, applying these strict benchmarks to AI systems raises questions about their appropriateness, given that even human decision-makers routinely deviate from perfect rationality. Standard evaluation methods often test narrow dimensions of reasoning in isolation, potentially missing the complex interplay of cognitive skills that real-world economic decisions demand. As AI systems grow more capable, the field is grappling with the need for more nuanced evaluation paradigms that account for bounded rationality and contextual judgment.
The Evolution of Rationality in AI
The concept of rationality in AI has deep roots stretching back to the earliest days of the field, when researchers sought to formalize logical reasoning in machines. Over decades, understanding of rationality has been shaped by cross-pollination between economics, philosophy, psychology, and computer science. The rise of deep learning and, more recently, large language models has reinvigorated debates about what it means for an artificial agent to reason well. Foundational work on bounded rationality by Herbert Simon and later behavioral economics research by Kahneman and Tversky have profoundly influenced how AI researchers think about agent decision-making, though the application of these insights to modern LLMs remains an active area of investigation.
Introducing STEER: A Report Card for AI Rationality
STEER isn't just another AI benchmark; it's a comprehensive framework designed to evaluate LLMs against the gold standard of economic rationality. It moves beyond ad-hoc tasks by enumerating first principles describing how agents should make decisions, then evaluating an agent's degree of adherence. The normative question of how decision-makers should act has been the focus of more than a century of research in economics, cognitive psychology, computer science, and philosophy.
- Foundations: Tests core mathematical and logical reasoning abilities.
- Decisions in Single-Agent Environments: Explores preference formation and decision-making with single deterministic or probabilistic outcomes.
- Decisions in Multi-Agent Environments: Assesses strategic thinking and game theory concepts.
- Decisions on Behalf of Other Agents: Evaluates the ability to aggregate preferences and make socially responsible choices.
The Current State of Rationality Research in AI
A comprehensive 2025 survey published in the journal Artificial Intelligence Review highlights that despite rationality being central to AI, there remains no unified definition of what constitutes a rational agent. The survey examines how conceptions of rationality from economics, philosophy, and psychology have influenced AI, and identifies irrational behaviors that can paradoxically prove optimal in certain scenarios. Separately, researchers have conducted extensive experiments testing prominent LLM families for systematic behavioral biases, drawing on cognitive psychology and experimental economics methodologies. Their work documents consistent patterns in the biases LLMs exhibit, raising important questions about the reliability of these models for economic and financial decision-making tasks.
Documented Biases and Reasoning Failures
Research consistently demonstrates that LLMs exhibit systematic behavioral biases mirroring many of those found in human cognition. Experimental economics studies applied to prominent LLM families have revealed these biases persist across model versions and scales, suggesting they are not easily eliminated by simply increasing model size. Critics, including prominent AI researcher Gary Marcus, have argued that despite nearly a trillion dollars in investment, LLM reasoning remains deeply flawed, pointing to a 2025-2026 review that underscores the breadth of the problem. These findings challenge the assumption that current LLMs can serve as reliable economic agents without significant corrections or guardrails.
Benchmarking LLMs Against Human Rationality
Comparing the rationality of LLMs to that of humans is an inherently complex endeavor, as both exhibit systematic deviations from classical economic rationality. While LLMs may outperform humans on certain structured reasoning tasks, they appear to replicate many of the cognitive biases documented in behavioral economics. The challenge for researchers lies in determining whether these shared biases represent a fundamental limitation of language-based reasoning or an opportunity to develop more human-like and contextually appropriate decision-making systems. As the field matures, comparative frameworks that go beyond simple accuracy metrics will be essential for a fair assessment.
Beyond the Hype: Towards Truly Rational AI Agents
The STEER benchmark provides a valuable tool for evaluating and improving the economic rationality of LLMs. By identifying specific areas where models struggle, researchers and developers can focus their efforts on fine-tuning, curating new datasets, and developing specialized architectures. The journey toward truly rational AI agents is just beginning, but benchmarks like STEER are essential for guiding our progress and ensuring that these powerful tools are used responsibly and effectively in the economic sphere.
Integrating Rationality into AI Design
The growing body of research on LLM rationality suggests that neither uncritical adoption nor outright dismissal is warranted. Experts increasingly advocate for a middle path: acknowledging the limitations of current models while investing in methodologies to measure and improve their decision-making reliability. The absence of a unified definition of rationality in AI means that progress will require interdisciplinary collaboration spanning economics, cognitive science, and computer science. Ultimately, the question may not be whether LLMs are perfectly rational, but whether they are rational enough for specific, well-defined applications.
AI's Economic Trajectory and Investment Landscape
Nobel laureate Daron Acemoglu has predicted that AI will have a 'nontrivial, but modest' effect on GDP over the next decade, tempering some of the more exuberant projections about AI's economic transformation. Meanwhile, Morgan Stanley's 2026 analysis identifies AI as the central force influencing global growth, earnings, and investment strategy, suggesting significant market expectations despite tempered productivity forecasts. AI trend analysis from 2026 continues to track performance improvements, pricing shifts, and the competitive dynamics between US and Chinese AI development. These divergent perspectives highlight the uncertainty that surrounds AI's ultimate economic impact and the importance of grounding predictions in rigorous evidence.
Systemic Risks of AI in Economic Systems
Deploying AI agents in economic contexts introduces systemic risks that extend beyond individual model failures. Questions of accountability, regulatory frameworks, and the potential for correlated failures across AI-driven financial systems remain largely unresolved. The concentration of AI capabilities among a small number of technology firms raises concerns about market power and equitable access to AI-enabled economic tools. As governments and institutions race to develop AI governance frameworks, the tension between innovation and caution continues to define the policy landscape.
From Hype to Proven Business Value
The World Economic Forum, in partnership with Accenture, has documented 32 real-world AI deployments across more than 30 countries and over 20 industries that are operating at scale to deliver improved business outcomes. A separate GitHub repository catalogues over 500 case studies of generative AI and LLM system designs from more than 130 companies, illustrating how industry leaders deploy and optimize these models in production. Research from Anthropic models the economic consequences of transformative AI between 2026 and 2030, mapping paths for AI capabilities into implications for GDP, labor displacement, and wage dynamics. These real-world examples demonstrate that while AI's rationality limitations are real, practical deployment is already yielding measurable results in specific, well-scoped applications.