NextFin News - On March 2, 2026, a landmark study published by the Harvard Business Review has sent ripples through the global tech and consulting sectors, revealing that the increasingly "agentic" nature of generative AI is fundamentally altering the landscape of corporate decision-making. As major developers including OpenAI, Anthropic, and Google transition their Large Language Models (LLMs) from reactive tools into proactive conversational agents, they are now directing the flow of strategic dialogue by asking questions rather than merely answering them. However, the study, which benchmarked 13 leading LLMs against 1,600 human executives using the Leaders’ Question Mix (LQM) framework, found that AI models possess systemic biases in their inquiry styles that could lead to catastrophic strategic oversights if left unmanaged.
The shift toward AI-led inquiry is exemplified by the rapid deployment of tools like OpenAI’s DeepResearch and the Singapore-based Manus, which are designed to interrogate users to clarify ambiguous requests. While this "consultant-style" interaction aims to reduce errors, the data suggests a profound mismatch in priorities. According to the HBR report, while human executives distribute their focus relatively evenly across investigative, speculative, productive, interpretive, and subjective questions (ranging from 17% to 22% per category), LLMs exhibit extreme volatility. For instance, Google’s Gemini 2.5 showed a 23.1-point spread in its questioning mix, heavily favoring interpretive "So what?" questions while significantly underrepresenting productive "Now what?" inquiries. This trend is consistent across the board: all 13 models tested asked fewer productive questions—those concerning resources, timing, and execution—than their human counterparts.
This divergence arrives at a sensitive geopolitical and economic juncture. Under the administration of U.S. President Trump, who was inaugurated in January 2025, the federal government has aggressively championed the deregulation of the AI sector to maintain American dominance over global competitors. U.S. President Trump has frequently emphasized that "American intelligence must lead the world," a stance that has accelerated the integration of AI agents into the core infrastructure of Fortune 500 companies. However, the investigative findings suggest that this rush to automate leadership functions may be embedding a "productivity gap" into the very heart of American enterprise. If an AI agent directs a brainstorming session but fails to ask about feasibility or resource allocation, the resulting strategy may be visionary but entirely unimplementable.
The analytical core of this issue lies in the "semantic vs. strategic" gap. LLMs are trained on vast corpora of text where interpretive and speculative language is abundant, but the gritty, context-specific nuances of "productive" questioning—the kind that happens in a closed-door boardroom regarding budget constraints or logistical bottlenecks—are often underrepresented or proprietary. Consequently, when an LLM like ChatGPT 5 or Grok 4 directs a conversation, it tends to keep the dialogue in a conceptual loop. This creates a "speculative bubble" in decision-making where ideas are refined and interpreted ad nauseam, but the critical path to execution is never interrogated. For a CEO relying on these agents to draft a 2026 market entry plan, the AI might identify the "why" and the "what if," but its failure to ask "who is responsible?" or "do we have the capital?" creates a high-risk vacuum.
Furthermore, the study highlights a lack of consistency that complicates corporate governance. Even within the same families of models, such as xAI’s Grok 3 and Grok 4, the questioning styles differ statistically. This lack of a "standardized inquiry protocol" means that a company’s strategic direction could shift simply based on which version of a model a manager chooses to use on a given Tuesday. This volatility introduces a new form of "algorithmic noise" into the corporate hierarchy. In the framework of the LQM, the "subjective" category—questions about emotional buy-in and political resistance—is also frequently neglected by AI. In a high-stakes merger or a sensitive restructuring, an AI that ignores the "unsaid" emotional dynamics of a team can lead a leader into a minefield of internal resistance that the model was never programmed to perceive.
Looking forward, the role of the executive is likely to shift from "answer-seeker" to "inquiry-editor." As U.S. President Trump’s administration continues to incentivize AI adoption through tax credits for automated infrastructure, the burden of risk management will fall on the human ability to "disengage the autopilot." We predict the emergence of a new professional discipline: Algorithmic Prompt Auditing. Leaders will need to intentionally inject the question types that AI lacks—specifically productive and subjective queries—to balance the machine’s interpretive bias. The risk is not that AI will make decisions for us, but that by controlling the questions, it will invisibly narrow the horizon of what we consider possible, leaving the most difficult questions of execution and human impact unasked until it is too late to pivot.
Explore more exclusive insights at nextfin.ai.
