Decision Support with AI: LLMs, Bayesian Networks, and Human Expertise
Large language models can analyze business problems and offer persuasive assessments. Their usefulness for many advanced tasks makes it tempting to turn to them for managerial decision support as well. For a manager considering a product launch, an investment, or a new policy, the central challenge is to anticipate the consequences of alternative courses of action. The value of an LLM’s advice depends on whether it offers a sound basis for judging those consequences. An LLM’s fluent, elaborate explanations can make it easy to overlook that we do not have direct access to the reasoning behind its recommendation.
This seminar develops a constructive use of LLMs for decision support: distilling their knowledge into explicit causal Bayesian networks. Human experts contribute knowledge, examine the represented relationships, and revise them. The Bayesian network model infers the outcomes of the actions under consideration and thereby supports the decision-making process.
Corporate Suites, Suite 501
880 Third Avenue, New York, NY 10022
November 5, 2026, from 2:00 p.m. to 5:00 p.m. (EST)
Seminar Overview
Part 1: What Decision Support Requires and How LLMs Can Contribute
Deciding Before We Know the Outcome
To find the optimal course of action at a decision point, we must be able to anticipate the consequences of actions we have not yet taken. In operations research, experimentation might be possible when dealing with recurring decisions of the same kind. For major one-off decisions, such as committing to develop and launch a new vehicle, there can be no experimentation at all. Hence, we need an oracle that reveals the outcomes before we commit to developing a particular type of vehicle or decide not to develop one at all.
Outside Greek mythology, however, oracles do not exist. In our world, mathematical or statistical models typically serve as substitutes. Today, LLMs are the most rapidly advancing type of model, and the question is whether they can serve as our oracle for decision support.
From a Persuasive Answer to an Explicit Model
An LLM gives us access to extensive knowledge through conversation. It can describe mechanisms, suggest relationships, and even supply numerical assessments. However, when asked to explain its reasoning, the LLM can provide another fluent explanation but may not be able to articulate the actual process that produced its original answer. We examine these limitations in detail and then propose an alternative approach: we transform the LLM’s knowledge into an explicit knowledge representation that we can use directly for decision support.
Distilling Knowledge into a Causal Bayesian Network
We refer to this LLM knowledge transformation as “distillation.” This process encodes the available LLM knowledge into named variables, associations, causal relationships, and probabilities in the form of a Bayesian network. Unlike an opaque LLM, a Bayesian network has a structure and inference mechanism that human experts can examine closely. Based on the knowledge obtained from the LLM, the Bayesian network model now represents how the domain of interest works. This allows us to query the Bayesian network to simulate all the actions under consideration and infer the outcomes and their corresponding expected utilities.
The Decision Maker’s Role
Performing inference with a Bayesian network does not mean that a decision is outsourced. Given the complete transparency of the Bayesian network, managers can audit the represented knowledge, check assumptions, examine how new evidence would change the anticipated outcomes, and thereby take responsibility for the recommended decision.
Part 2: A Worked Decision Example in BayesiaLab
We develop a causal Bayesian network live in BayesiaLab, drawing on knowledge elicited from an LLM and assessing it through expert judgment. The example makes the transition from verbal descriptions to explicit relationships and numerical assessments visible.
We use the model to simulate observations and interventions, then compare actions using the probabilities and utilities of their outcomes. By changing underlying assumptions in the model, we show how differing assessments can affect the expected utility and risk associated with each action under study. We also establish Value of Information as an important concept in the decision-making process. It can help us determine how gathering additional evidence can change the expected utilities of outcomes and thereby lead to a different decision.
The demonstration connects the seminar’s central ideas to calculations participants can follow: what the language model contributes, what the Bayesian network computes, and where human knowledge and preferences enter the analysis.
Who Should Attend
This seminar is for managers, analysts, consultants, researchers, and policymakers involved in any type of decision-making or decision support. Relevant settings include strategy, investment, marketing, finance, product development, public policy, and risk management, to name a few.
Concepts are introduced through a real-world decision problem and a worked example accompanied by a written tutorial. The emphasis is on understanding and examining the analysis, with technical terms explained as they become relevant.
Participation is free.
Venue
The seminar takes place at Corporate Suites, Conference Room 6D, 880 Third Avenue, New York, NY 10022, at the corner of Third Avenue and East 53rd Street in Midtown Manhattan.
About the Presenter
Stefan Conrady
Stefan Conrady has over 20 years of experience in decision analysis, analytics, market research, and product strategy, having worked with Mercedes-Benz, BMW Group, Rolls-Royce Motor Cars, and Nissan across North America, Europe, and Asia. As Managing Partner of Bayesia USA and Bayesia Singapore, he is widely recognized as a thought leader in applying Bayesian networks to research, analytics, and decision-making. Together with his business partner, Dr. Lionel Jouffe, he co-authored Bayesian Networks & BayesiaLab — A Practical Introduction for Researchers, an influential resource now widely cited in academic literature. With their deep expertise in Bayesian networks for Key Driver Analysis and Optimization, Stefan and Lionel are highly sought-after consultants, advising global leaders such as Procter & Gamble, Coca-Cola, UnitedHealth Group, L’Oréal, the World Bank, and many of the world’s largest market research firms.
Register Now (Free)