Decision Architecture: The Real AI Edge

Artificial intelligence is generating tremendous excitement across the investment industry. However, a critical yet overlooked factor is preventing investment firms from unlocking AI’s full potential.

A growing body of recent research suggests that the most consequential impact of AI will not come from automating isolated tasks, but from reconfiguring entire systems of work. For investment firms, this requires reshaping human-machine interaction, just as the shift from steam engines to electricity demanded entirely new work processes.

The need to reshape the broader ecosystem in which AI and humans interact also challenges two dominant views in the investment industry.

First, that AI is primarily a task-automation tool. Second, that performance gains come mainly from scaling models. Evidence increasingly suggests that true performance, reliability, and economic value stem not from automation or model size alone, but from how effectively AI is embedded into decision-making environments and how well it complements human cognition,

This post is the fifth installment of a quarterly reflection on the latest developments in AI for investment management professionals. Drawing on insights from investment specialists, academics, and regulators contributing to the bi-monthly newsletter The Augmented Intelligent Investor, it builds on earlier articles that explored AI’s promise and pitfalls, where AI ends and judgement begins, and practical risk management techniques in times of AI. This installment explores a critical yet overlooked area that investment firms must address to fully unlock AI’s potential.

In the industrial age, steam-powered factories were built entirely around a central engine. Machines and workers clustered tightly around the power source because energy could not travel efficiently over distance. The entire factory layout was dictated by the limitations of steam. When electricity arrived, it promised far greater flexibility: energy could be delivered anywhere, on demand.

However, for years, productivity gains remained modest. Factories simply swapped steam engines for electric motors while keeping the old, centralized layouts. The real leap in output only occurred when factories were completely redesigned to exploit electricity’s decentralized nature.

The same principle applies today to investment firms embracing AI. The machine is already unlocking significant improvements in investment analysis and across many areas of the investment process (Wierckx, Zilic, Kuhn, Schuller, 2025). However, these gains are largely incremental because most firms are still operating with decision environments built for the pre-AI era.

This emphasis on redesigning decision-architecture directly ties into the central insight from the work of Gans and Goldfarb (2026). They point out that production is not a collection of independent tasks, but a network of interdependent activities. When AI is introduced into such systems, it does not simply replace labor at the task level; it reshapes the allocation of effort across the entire production process. This could lead to nonlinear outcomes: partial automation can have muted effects, while full system redesign can generate discontinuous productivity gains.

For investment firms, it implies that those treating AI as a “plug-in automation” tool, are systematically underestimating both its upside and its complexity. The real gains come when investment firms redesign workflows, redefine decision making processes, and restructure feedback loops between human and machine agents.

The paper Agentic Reasoning for Large Language Models (2026) formalizes this issue by arguing that current LLMs are fundamentally reactive systems. They generate outputs sequentially without strategic planning or explicit self-monitoring. The authors propose a shift toward agentic control loops: observe → plan → act → reflect → update state → repeat.

The significance of this formulation extends beyond AI-model design. It also provides investment firms with a blueprint for how their investment processes must be reorganized. In investment process that embraces AI, reasoning is no longer a single cognitive act but a distributed process with checkpoints, memory, and role separation.

This has direct implications for the way investment firms should be organized. If AI systems are to function reliably in high-stakes environments, including investment decision-making and trade execution, they must be embedded in workflows that enforce iteration, critique, and state tracking. Without these structural constraints, even highly capable models degrade through what the literature describes as “silent error compounding.”

The next question is how humans interact with those new AI augmented environments. Here, the latest literature converges on a critical but often underappreciated constraint: human cognition itself.

A recent meta-analysis(Gullich, 2025) shows that while early specialization accelerates initial progress, cross-domain experience predicts higher long-term performance. The implication is that high-level judgment depends on integrative cognition rather than narrow expertise.

This matters in AI-augmented environments because system-level reasoning requires exactly this kind of integrative capacity. As AI systems become more capable of generating options, simulating scenarios, and producing analyses, the human role shifts toward evaluation, synthesis, and strategic arbitration across domains.

Empirical labor studies, including No AI Jobpocalypse (Parikh, 2026). show that indeed roles are being redefined: routine components are automated, while cognitive and integrative tasks become more valuable. This confirms Gans and Goldfarb’s prediction of nonlinear adjustment dynamics.

However, cognitive science research also highlights fragility in these processes. Studies such as The Effects of Smartphone Addiction on Learning (Sunday, 2021) show that constant interruptions degrade attention, working memory, and critical thinking. In AI-rich environments, this is crucial: the more capable the system becomes, the more fragile human oversight may become if cognitive conditions deteriorate.

Another reason why AI-augmented decision environments must be optimized for human cognition, is that AI accelerates output faster than it improves epistemic quality (Gullich, 2025). The result is an inflation of plausible but weakly validated knowledge. This creates what might be called a “complexity illusion”, or a situation where volume and sophistication of expression are mistaken for rigor.

Machines are currently overburdened by the task of generating evidence-based epistemological insight at the frontier of knowledge, much as humans themselves can be overwhelmed in this domain.

Rather than alleviating human cognitive strain, machines primarily contribute to the personalized and scalable amplification of human cognitive dissonances. While this may enhance productivity in the provision of goods and services, it does not necessarily deepen our understanding of the world (Schuller, 2026).

We should remember that AI systems are not neutral tools but at best cognitive enablers that shape reasoning behavior. This means that poorly designed interaction loops can amplify existing confirmation biases. In contrast, well-designed systems can function as adversarial collaborators: forcing counterfactual reasoning, surfacing hidden assumptions, and stress-testing hypotheses.

The research shows that AI’s greatest impact will not come from automating individual tasks, but from redesigning the systems within which decisions are made. For investment firms, the key challenge is therefore not simply adopting more powerful models, but creating decision environments that effectively combine machine capabilities with human judgment.

Competitive advantage will increasingly belong to firms that redesign workflows, information flows, and decision processes to leverage the complementary strengths of humans and AI. The future of AI is ultimately less about technology itself and more about the architecture of decision-making.

Yet even in increasingly AI-augmented environments, human judgment remains indispensable. Machines can generate insights, challenge assumptions, and expand analytical capacity, but the responsibility for interpreting evidence, exercising judgment, and determining what is true remains fundamentally human.

References

Gans, J. and Goldfarb, A. (01/2026) “O-Ring Automation”, NBER Working Paper Series, January 2026

Gullich, A., et al. (12/2025) “Recent discoveries on the acquisition of the highest levels of human performance”, Science, December 2025

Kusumegi, K., (12/2025) “Scientific production in the era of large language models”, Science, December 2025

Parikh, T. (02/2026) “Don’t fear the AI ‘jobpocalypse’” Financial Times 

Schuller, M. (03/2026) "The Timeless Pursuit of Evidence" – Panthera Group

Sunday, O.J, et al. (2021) “The effects of smartphone addiction on learning: A meta-analysis”, Computers in Human Behavior Reports, Volume 4, August–December 2021

Wei, T. et al, (01/2026) “Agentic Reasoning for Large Language Models”, Yale University, Google Deepmind, UC San Diego, et al., 

Wierckx, P. et al (10/2025) “A Multi-Dimensional Classification System For AI Agents In The Investment Industry” DePaul University, Panthera Group, SSRN.