Theory-Based Acquisition Strategies – an experimental analysis of AI impact on decision-making in M&As.

The experiment implements a three-arm randomized controlled trial (RCT) with two parallel AI interventions and one control group. All participants first receive a short online training video introducing the Theory-Based View (TBV) of strategy, emphasizing causal reasoning in strategic decision-making. After viewing the training, participants complete an M&A decision challenge and develop a brief written acquisition strategy.

Arm 1 – Control (TBV + Web Search):
Participants complete the TBV training and address the M&A challenge using only their own reasoning and publicly available information via Google Search. No AI assistance is provided.

Arm 2 – TBV + ChatGPT (General-Purpose LLM):
Participants complete the TBV training and use OpenAI’s ChatGPT (GPT 5.4, reasoning effort set to medium) as a general-purpose large language model to assist them in researching, formulating, and refining their strategic theory before submitting their final decision. Participants are allowed to use web search as well.

Arm 3 – TBV + Aristotle (Agentic AI):
Participants complete the TBV training and use "Aristotle", a specialized agentic AI system developed at Bocconi-IMSL that applies TBV reasoning principles. The agent autonomously supports causal reasoning and theory construction, providing targeted feedback and prompting to improve strategic coherence. Participants are allowed to use web search as well.

All other procedures, materials, and timing are identical across conditions. Total participation time is approximately 1 hour. Interventions are delivered online via the Qualtrics platform. Randomization is implemented automatically within the survey workflow using minimized allocation to maintain covariate balance across education, M&A experience, and baseline AI aversion.

2026-04-27

2026-05-31

Subjective Probability of Acquisition Success (0–100 scale): The participant's self-assessed likelihood that the counterpart will accept their expressed willingness-to-pay or a lower price, elicited in 10-percentage-point increments.

Subjective Probability of Positive Returns (0–100 scale): The participant's self-assessed likelihood of achieving a positive return on the acquisition, conditional on deal completion, elicited on the same scale.

Confidence in Probability Assessments and Strategic Reasoning (7-point Likert): Three sub-constructs are measured: (a) confidence in the stated probability of acquisition success, (b) confidence in the stated probability of positive returns, and (c) confidence in the soundness of the strategic theory applied.

Outcome Alignment (Brier Score): The calibration of each participant's probability assessments against realized binary outcomes, computed using the Brier Score. Lower Brier scores indicate better-calibrated forecasts. Two evaluation horizons are pre-registered, however only the long-term evaluation benchmark represents a primary outcome: this is based on seven board-identified strategic objectives evaluated over a 36-month horizon from acquisition close (November 14, 2025), not observable before Q4 2028.

All outcomes are collected post-intervention within the same Qualtrics session. Outcomes 1–3 capture subjective assessments of decision quality; Outcome 4 anchors those assessments to ground-truth realized performance. The binary acquisition decision (proceed or decline) is reported alongside the Brier score as a complementary measure of forecast accuracy.
The primary treatment effects are estimated through pairwise contrasts between:

(a) ChatGPT vs. Control,
(b) Aristotle vs. Control, and
(c) ChatGPT vs. Aristotle (ITT only).

All tests are two-sided with family-wise error rate controlled at α=0.05 using Bonferroni correction, with the Holm procedure applied at the analysis stage for improved power. ITT effects are estimated for all three contrasts; LATE effects via instrumental variables are estimated separately for each AI arm against the control group, and are not directly compared across the two AI treatments.

Subjective Probability of Acquisition Success and Subjective Probability of Positive Returns capture the participant's probabilistic assessments of two sequential events: whether the deal closes at or below their stated willingness-to-pay, and whether it generates a positive return conditional on closing. Together, these elicitations operationalize the precision and directionality of the participant's forecast about the acquisition opportunity.

Confidence in Probability Assessments and Strategic Reasoning measures the participant's meta-cognitive certainty across three sub-constructs: confidence in each of the two probability estimates and confidence in the underlying strategic theory. This outcome captures whether AI assistance affects not only the content of participants' judgments but also their epistemic relationship to those judgments. It is conceptually distinct from the probability elicitations themselves: a participant may assign a moderate probability with high confidence, or a high probability with low confidence, and these configurations have different implications for decision quality under uncertainty.

Outcome Alignment anchors the above subjective measures to ground truth, assessing whether AI-assisted participants produce probability assessments that are better calibrated to realized outcomes. This is the central criterion for evaluating forecast quality in the study: it rewards neither overconfidence nor excessive hedging, but the accuracy of probabilistic reasoning under ambiguity. The deferred long-run benchmark evaluates alignment against the actual strategic performance of the acquired business unit.

Together, these outcomes assess the study's core theoretical proposition: that AI assistance — and agentic, theory-driven AI assistance in particular — improves both the calibration of managers' probabilistic forecasts and the epistemic confidence with which they hold those forecasts, relative to unaided web search.

Outcome Alignment — Short-Run Market Benchmark (Brier Score, short-run): The calibration of each participant's probability assessments against Biogen's stock price abnormal returns in the announcement window around the acquisition date (September 18, 2025), treated as a market-based proxy for informed expectations about deal value creation. This benchmark is observable at the time of initial analysis and serves as the primary reference for classifying decision outcome alignment in the first publication.

Maximum Willingness-to-Pay (WTP, in $M): The highest price at which the participant would recommend proceeding with the acquisition. This outcome captures the participant's quantitative assessment of deal value and complements the binary acquisition decision and probability elicitations.

DAG Complexity and Size: Each participant's written strategic theory is processed through a semi-automated algorithm to extract a directed acyclic graph (DAG) representing its underlying causal structure. Three families of graph metrics are computed: (a) graph size (number of nodes and edges), (b) graph complexity(average-in-degree, maximum-in-degree, graph density), and (c)within-graph diversity, measured by the number of domain-relevant nodes spanning M&A due diligence categories identified in the literature (described in the pre-analysis plan Appendix A). These structural metrics operationalize the hypothesis that AI assistance expands the breadth and causal depth of participants' strategic theories.

Human–AI Interaction Quality (HIQ): For participants in the two AI treatment arms, all conversation threads are scored along four components using a structured rubric evaluated by a blinded LLM rater: (a)Domain-Specific Terminology, (b)Task-Relevant Specificity, (c)Input Delineation , and (d)Task Decomposition. Each component is entered individually as an ordinal variable in a correlational analysis linking interaction quality to primary outcomes. This analysis is explicitly non-causal: participants who prompt more skillfully may produce stronger theories independently of AI assistance, due to greater domain expertise or cognitive ability.

The secondary outcomes deepen the analysis of how AI assistance influences the mechanisms underlying strategic reasoning, forecast calibration, and decision quality.

Outcome Alignment — Short-Run Market Benchmark grounds the primary probability elicitations in an observable, theory-neutral reference point at the time of initial analysis. While the long-run board-dimension composite is the more theoretically meaningful criterion for evaluating acquisition quality, the market reaction in the announcement window provides an immediately available signal of whether participants' forecasts aligned with the expectations of informed investors at the moment the deal became public.

Maximum Willingness-to-Pay translates participants' qualitative strategic assessments into a quantitative valuation judgment, providing a continuous measure of perceived deal value that complements the binary acquisition decision. It captures whether AI assistance affects not just the direction of participants' recommendations but the financial precision with which they anchor those recommendations — a practically relevant dimension of M&A decision quality that probability elicitations alone cannot capture.

DAG Complexity and Size operationalize the structural hypothesis that AI assistance expands the breadth and causal depth of strategic theories, independently of their calibration against outcomes. Graph size measures how many distinct causal factors participants identify; graph complexity measures how densely those factors are interconnected; and within-graph diversity measures whether participants' theories span the full range of M&A due diligence domains rather than concentrating on a narrow subset. These metrics allow the study to distinguish between two qualitatively different mechanisms: AI tools may improve outcome alignment by helping participants reason more thoroughly across more domains, or they may improve calibration without appreciably altering theory structure. The DAG measures are designed to adjudicate between these possibilities.

Human–AI Interaction Quality (HIQ) characterizes the behavioral process through which participants engage with the AI tools, linking prompting behavior to primary outcomes in a correlational framework. Participants who decompose tasks, use domain-specific terminology, and structure their inputs clearly may extract more value from AI assistance — but because such participants may also be stronger reasoners independently, the HIQ analysis is framed as descriptive rather than causal. It nevertheless provides the empirical basis for understanding heterogeneity in treatment effects and for informing the design of future AI-assisted decision-support interventions.

Together, these outcomes form a multi-layered picture of AI's influence on strategic reasoning: from the financial precision of valuations, through the structural properties of causal theories, to the attitudinal and behavioral dynamics of human-AI interaction.

The study is a three-arm randomized controlled trial (RCT) designed to test how different forms of AI assistance influence strategic reasoning quality and confidence among managers making M&A decisions.

All participants first receive a standardized short training video introducing the *Theory-Based View (TBV)* of strategy, which teaches causal reasoning and theory formulation principles. Immediately after training, participants complete an online M&A case challenge requiring them to develop a brief acquisition strategy and justify their reasoning.

Participants are randomly assigned (via minimized randomization) to one of three experimental arms:

1. Control Group – TBV + Google Search: Participants complete the M&A task using only web search and their own reasoning.
2. "Intervention 1 – TBV + LLM:" Participants use a general-purpose large language model to assist with information gathering, idea refinement, and theory formulation, other than standard web search.
3. "Intervention 2 – TBV + Agentic AI:" Participants use a specialized agentic AI that provides structured guidance and feedback grounded in causal reasoning principles, web search.

The experiment is conducted fully online using the Qualtrics platform. Total participation time is approximately 1 hour.

After completing the task, participants answer post-intervention surveys measuring subjective outcomes and provide qualitative feedback. Written responses are later coded blind to condition by expert judges and by an LLM-as-evaluator for objectivity and robustness checks.

The study design allows direct comparison of (a) general AI vs. no AI, (b) agentic AI vs. no AI, and (c) agentic AI vs. general AI. We compare both the effect of being assigned to a treatment condition (ITT) and the actual effect on compliers (LATE) through an IV analysis.

Not available

An asymmetric randomization is conducted and implemented via the Qualtrics platform, using stratified randomization on key covariates to ensure balanced groups. The allocation procedure follows a 1:1.2:1.8 ratio to ensure enough statistical power even in case of partial non-compliance to the treatments.

Intervention assignment:
After the baseline survey, participants are allocated to interventions in the order of enrollment. The first ∼120 qualified respondents are assigned to Intervention 1 (TBV training vs. placebo), and the next ∼120 respondents are assigned to Intervention 2(AI assistance). Within the first group (Intervention 1), participants are randomized 1:1 to TBV training vs. placebo. We use stratification on important covariates – gender, field of education, years of experience, and baseline AI aversion – to achieve balance between the TBV and placebo groups and explore potentially meaningful heterogeneous treatment effects. Similarly, within the second group (Intervention 2), participants are randomized 1:1 into the two AI conditions (General AI vs. Agentic AI), again using stratified randomization on the same covariates to ensure balanced characteristics across these AI groups and enable heterogeneous treatment effects analysis. (No new participants are directly assigned to a “No AI” condition in Intervention 2, since the No AI comparison group consists of the TBV-trained participants from Intervention 1.) We will record the randomization procedure with software logs, including random seeds and assignment timestamps, to ensure transparency.

This stratified approach prevents detectable imbalances in meaningful observable characteristics and upholds group equivalence. All participants provide informed consent before randomization. We will verify ex-post that the groups are balanced on baseline covariates (e.g., demographics, experience, etc.), and if any notable imbalance arises by chance, we will control for those covariates in the analysis as a precaution.

Individual

No

A minimum of 400 senior managers

400 participants

Our final target sample size is 400 participants in total: approximately 100 in the control group, 120 in Intervention 1 and 180 in Intervention 2

The design can detect a small to moderate effect size (Cohen’s d = 0.45) with sufficient power. Assuming a two-tailed test with family-wise α = 0.05, desired power (1 − β) = 0.80 our sample size meets the requirement both for the ITT comparison, and the LATE comparison with expected compliance rates of 0.9 for Intervention 1 and 0.8 for Intervention 2.