Even in the best of times, strategic planning is a challenge for business leaders. When global uncertainty and volatility rise, the task gets even more difficult—and more integral to business success. When the world is in flux, an organization’s optionality grows in importance, as does organizational resilience. In fact, as research has repeatedly shown, resilience is now a substantial driver of corporate outperformance.
But building resiliency and optionality into a strategic plan challenges humans’ cognitive (and financial) bandwidth. The seemingly endless array of future scenarios, coupled with our own human biases, conspires to anchor our understanding of the future in what we’ve seen in the past. Generative AI (GenAI) can help overcome this common organizational tendency for entrenched thinking, and mitigate the challenges of being human, while exploiting LLMs’ creativity as well as their ability to mirror human behavioral patterns.
One potent approach to incorporating GenAI into strategic planning is coupling it with agent-based modeling (ABM) that is already in use to simulate complex, dynamic scenarios. Instead of relying on ABM’s deterministically coded agents of yore—still constrained by the boundaries of human imagination—modern LLMs can make simulations much more flexible, human-like, and (fruitfully) unpredictable. What’s more, they can achieve this at a fraction of the time and cost of in-person planning workshops, providing a powerful tool to explore a wider range of futures and prepare for the unexpected with greater agility.
The promise of LLM-powered behavioral simulations
GenAI has the potential to supercharge the strategic arsenal that is (or at least should be) commonplace among large corporations, including war gaming and scenario planning. The key is using the technology to simulate interpersonal or inter-institutional dynamics, from boardroom discussions to international competition and engagement with regulators.
Software has long been deployed in similar ways in the social and natural sciences using agent-based modelling in which heterogenous, independent agents interact over time. ABM has been used to simulate the spread of infectious diseases or even the emergence of behavioral norms in human societies. The use of ABM typically involves the time-intensive design of each of the heterogenous agents, traditionally comprised of a series of hard-coded rules that determine how they respond to inputs and interact with other agents. Now, with the ability to use dynamic LLMs as agents, ABM is within reach for most companies. (This use of LLMs is not to be confused with 2025’s buzzword, agentic AI.)
Having this sort of flexible, cheap, scalable aid for strategy makes it much easier for businesses of all sizes to put in practice the OODA loop often used in military contexts: “Observe in order to adjust the reference scenario based on the cockpit indicators; orient by identifying the strategic options according to the company’s starting point; decide on the most effective option; and act quickly and accordingly.” The OODA loop makes organizations better, faster learners.
In fact, our argument reflects our own experience using a multi-agent LLM simulation platform built by the BCG Henderson Institute. We’ve used this platform to mirror actual war games and scenario planning sessions we’ve led with clients in the past. As we’ve seen firsthand, what makes an LLM multi-agent simulation so powerful is the possibility of exploiting two unique features of GenAI—its anthropomorphism, or ability to mimic human behavior, and its stochasticity, or creativity.
LLMs can role-play in remarkably human-like fashion: Research by Stanford and Google published earlier this year suggests that LLMs are able to simulate individual personalities closely enough to respond to certain types of surveys with 85% accuracy as the individuals themselves. Other studies show that, when appropriately prompted, LLMs can replicate human decision-making patterns in economic experiments, or can accurately reproduce linguistic patterns and political inclinations observed in actual social media user behavior.
In addition, LLMs are not deterministic models, meaning that the same inputs won’t always result in the same output. They are instead stochastic, endowed with an inherent degree of randomness in the way they generate outputs. Stochasticity is the root cause of GenAI’s so-called “hallucinations,” or the generation of false answers to factual queries, but it is also what gives them such creativity-enhancing potential. When it comes to creative simulations, and expanding one’s conception of possible futures, these “hallucinations” can be plus.
By using a modelling platform that incorporates multiple LLMs, it becomes possible to have a range of agents, such as regulators, customers, and competitors, each with their own set of idiosyncrasies and agendas, all of which more closely resemble real-life, dynamic interactions The use of multiple LLMs in a simulation also helps minimize the role of human biases as expressed in the prompting of a single model. Researchers at Columbia University found that using a multi-agent approach over a single-agent approach improves the accuracy of simulations of human behavior by roughly 75%.
New fixes for old problems
So, how can LLM-powered ABM help leaders develop and augment their strategic planning efforts?
Blind-spot detection
The very experience that makes high-level executives valuable—deep industry knowledge, pattern recognition from past crises, and relationships built over years—can also limit their ability to envision truly disruptive or just unconventional moves by others in the marketplace. This, in effect, creates blind spots for executives and their organizations, leaving them vulnerable to disruption and closed off to creative solutions. Amazon’s surprise entry into grocery retail with the 2017 acquisition of Whole Foods, for example, must’ve seemed implausible to incumbent retailers at the time and likely wasn’t considered as a scenario to anticipate. Such unlikely, but highly consequential possibilities are hardest to foresee due to human bias, but grappling with them can expand strategic imagination and foster more adaptive, option-rich planning.
These biases and constraints can manifest at all levels of business, not just in individuals. Similar constraints can also impact a group or team, where the innate collective drift towards groupthink limits idea diversity. At an organizational level, the prevailing culture can also entrench certain paths, sidelining new ideas, and leading to institutional inertia. The use of AI can help overcome these constraints, in part, by allowing organizations to better identify unknown unknowns. Once transformed into known unknowns, it becomes easier to get ahead of such possibilities by building flexibility into plans.
Extending the reach of strategic planning tactics
Extending the reach of strategic planning tactics
The cost associated with strategic planning can also make it infeasible to gather all of the stakeholders needed to game out dozens of possible scenarios. Such costs can affect the frequency with which companies assemble their brain trust, or make it less likely that these exercises are conducted organization-wide. Yet the more frequently organizations can engage in scenario planning and war gaming, the better prepared they can be for navigating unforeseen, exogenous shocks to their business.
As complements to live strategy sessions, LLM-powered ABMs are easy to use, cheap to deploy, and scalable. As a result, they can facilitate the spread of the successful strategic planning tactics to a wider range of teams within a business (and beyond to new organizations) as well as more frequent reassessments of strategic plans and decisions.
Facilitating convergence
LLM-powered ABM can’t and shouldn’t replace classic, in-person strategy sessions because these face-to-face interactions help leaders unite around a shared strategic direction. But these new tools can help make that happen by building confidence in an organization’s strategic decisions. Simulations can be run repeatedly, eliciting patterns where a sort of Venn diagram of commonalities across scenarios emerges. While frequency doesn’t necessarily correspond to probability, it can nudge attention toward previously overlooked paths and help alignment on new strategies.
When we compared the output of our LLM-powered, multi-agent ABM with strategy workshops held with a top life insurance company, for instance, our simulations arrived at three of the same strategic recommendations that the human-led workshops produced, allowing the company’s leaders to go forward with additional confidence in those decisions. The LLM-powered ABM simulations also pointed to two new options that hadn’t come up during in-person workshops, including the strengthening of workforce training in emotional intelligence and AI literacy.
How to start, and why do it now
For organizations that want to get started, the first step is to augment strategic decision-making processes with multi-agent GenAI platforms. This doesn’t require starting from scratch; leaders should prioritize using existing frameworks to help establish inter-agent goals, context, and dynamics. It’s important to use the platform before and alongside existing strategic planning sessions, and to run it at scale in order to find the Venn diagram of most resilient strategic steps. Those results can then be used to build consensus on strategic decisions, to harness and maximize the potential of the future rather than fearing the uncertainty that comes with it.
Starting now can help make AI a routine input, helping organizations acclimate to an environment of higher-frequency change and adaptation informed by real-time learning. Evidence shows that resilience and optionality are now more important than they have been in decades. The sooner that a company can upgrade its strategic planning and foresight capabilities, the more likely it’ll be to thrive in a world of heightened uncertainty.
