Who Speaks for the Organization? Simulating Strategic Organizational Reasoning in Regulatory Comment Processes

Anita Srinivasan

Keywords: LLM agents, policy simulation, regulatory comments, organizational reasoning, multi-agent deliberation

TL;DR: We compare three LLM agent architectures for simulating organizational regulatory comments, finding monolithic agents predict positions perfectly while deliberative multi-department agents produce richer arguments but introduce emergent errors.

Abstract: Large language model (LLM) policy simulations typically model individuals expressing authentic preferences. Yet the actors who most shape regulatory outcomes are organizations producing strategic documents designed to build legal records, not express beliefs. We test whether modeling internal organizational deliberation—decomposing an organization into legal, business, and government affairs departments before synthesizing a public comment—produces more realistic simulated regulatory comments than either individual agents or monolithic organizational agents. Using the SEC's proposed amendments to Rule 3b-16 (expanding the definition of "exchange" to cover DeFi protocols) as a case study, we simulate comments from eight real organizations across three conditions and validate against the actual filed comment record. We find that monolithic organizational agents achieve near-perfect position prediction (100% accuracy) and high argument alignment, while deliberative agents produce marginally broader arguments but introduce a position error in one case—a community banking association whose internal departments recommended cautious support but whose synthesis agent flipped to opposition. We also apply the same architectures to an anticipated SEC rulemaking with no existing comment record, demonstrating the tool's forward-looking predictive use case.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 13