Reasoning Machines: AI Agents and the Evaluation of M&As

Emil Mirzayev, Marco Testoni, and Bart Vanneste

Working paper

Mergers and acquisitions (M&As) can create value through synergies, but often destroy value due to imperfect managerial decision making. Advances in artificial intelligence (AI), and particularly in large language models (LLMs), offer new avenues to support the evaluation of M&A opportunities. This study investigates how well artificial evaluators, and especially increasingly autonomous AI agents, can predict whether M&As are value-creating. We decompose evaluation into analysis (examining an alternative’s strengths and weaknesses) and synthesis (combining these examinations into an assessment). Four evaluator conditions are examined, which differ in how they organize evaluation: synthesis only (base), analysis and synthesis by a single role (chain-of-thought), analysis shared between two roles with one of them also synthesizing (reflection), and analysis and synthesis by different roles (multi-agent). Using anonymized announcements of deals between U.S. public firms, only the multi-agent condition demonstrates substantive predictive power for how likely an M&A is to create value. This predictive power corresponds to substantial economic gains (i.e., the top half of deals deemed most promising gain 3.3% in abnormal returns, while the bottom half with the least promising deals lose 2.5%), also holds for deals that occurred after the LLMs were trained (i.e., in an out-of-sample test), and compares favorably to human expert evaluators (i.e., equity analysts updating their 12-month share price forecasts shortly after the M&A announcement). Furthermore, we find that the multi-agent evaluator’s predictive power is driven especially by its analysis. These results underscore the potential of AI agents in enhancing strategic decision making in M&As.

Link to full article