Planning for Emergencies Without Good Historical Data: A Human-in-the-Loop LLM Approach

Preparedness planning has a strange problem at its core: the emergencies you most need to plan for are often the ones you know the least about.

Past incidents are sparse. The records that do exist are fragmented. And even when history is available, it can mislead. The next emergency may not resemble the last pandemic, the last storm, or the last supply disruption. Still, public agencies have to make concrete decisions today about what to buy, what to store, and what risks to prepare for.

That gap is what this work tries to address.

The hardest emergencies to plan for are usually the ones with the weakest data — and the least time to think about them.

Public agencies and private organizations alike are expected to prepare for high-impact disruptions that may be rare, unprecedented, or unlike anything in recent memory. At the same time, the planners responsible for that work are usually stretched thin. They are managing active risks, coordinating across stakeholders, and making time-sensitive decisions with limited staff and limited analytical capacity. Under those conditions, a complex research exercise to identify a broad portfolio of plausible but unfamiliar incidents is usually not feasible.

That is where our approach creates value. We developed and tested a human-in-the-loop planning method that uses a large language model to generate a structured first-pass portfolio of incident scenarios for a specific stockpiled item. Experts then review, challenge, and refine that draft before it enters the planning model. The result is a faster, more systematic way to build planning inputs when historical data is weak, expert time is scarce, and the next disruption may not look like the last one.

Why traditional planning breaks down

Emergency stockpile planning sounds straightforward until you ask what information the decision actually depends on.

To decide how much of an item to hold, planners need to understand which events could create demand, how severe those events might be, how long they could last, how much of the need would already be covered by facilities or suppliers, and what residual demand would remain. In many real-world settings, that information is not available in a usable historical form.

Past events are often poorly documented for planning purposes. Data may be fragmented across institutions, captured inconsistently, or simply never translated into item-level demand assumptions. Even when history exists, it may not provide much guidance. COVID-19 did not look like a standard historical pattern. A flash flood in a region that has not seen one in generations won't either. In these settings, history is often incomplete, and sometimes misleading.

The fallback is usually expert judgment. But that creates a second problem. Building a meaningful incident portfolio from scratch is demanding work. It takes time, concentration, and a wide-angle view of operational risk. And in both public and private sector organizations, the people best positioned to do that work are often the same people already carrying heavy operational loads. As a result, scenario development tends to become narrow, recent-event-driven, and inconsistent from one planning cycle to the next.

A more practical way to build planning inputs

We built this method to solve that exact bottleneck. Instead of asking planners to start with a blank page, the system generates a structured first draft of the incident portfolio for a specific item. That draft includes:

• event-driven surge scenarios
• supply chain disruption scenarios
• mixed scenarios where an incident and a supply constraint happen at the same time
• lower-relevance scenarios that keep the analysis from becoming a worst-case-only exercise
• residual demand assumptions
• probability rationales
• available stock assumptions
• explicit flags for values that depend on AI-estimated inputs

This is not positioned as a forecast and not treated as truth. It is a planning accelerator. The value is operational: experts no longer have to spend their scarce time generating every scenario from scratch. Instead, they can focus on the higher-value task — reviewing relevance, correcting assumptions, challenging probabilities, and deciding what belongs in the final planning model.

How the method works

The LLM is one component inside a broader planning framework.

At a high level, the process works in four steps:

1. Develop an incident portfolio for a specific item — A set of plausible scenarios is created, along with residual demand estimates and probabilities.
2. Simulate demand across the planning horizon — The scenario portfolio feeds a Monte Carlo simulation that produces an empirical distribution of residual demand.
3. Define inventory trade-offs — Decision makers characterize the consequences of too little inventory versus too much inventory, including financial and operational impacts.
4. Generate an inventory planning reference point — An inventory model translates those assumptions into an analytically grounded inventory quantity.

The LLM supports the first step: building the scenario portfolio that drives the rest of the analysis.

Why the human-in-the-loop model matters

In high-stakes planning, the right role for AI is not autonomous decision-making. It is structured support.

That is why the system was designed so uncertainty stays visible. AI-estimated stock values are flagged. Higher-probability scenarios require stronger justification. Zero-residual-demand scenarios are highlighted for review if they depend on uncertain assumptions. The output is meant to be inspected, challenged, and improved.

This is what makes the approach deployable in practice. It preserves expert control while dramatically reducing the time and mental effort required to get to a usable planning set.

Where this can be applied

Although this work was developed in the context of emergency stockpile planning, the underlying method is applicable to many other areas. It is relevant anywhere organizations face the same combination of conditions:

• weak or fragmented historical data
• rare or unprecedented disruptive events
• limited expert time
• a need to make planning assumptions explicit and reviewable

That includes public health preparedness, emergency response, supply chain resilience, continuity planning, strategic inventory decisions, and other operational risk contexts in both the public and private sectors.

In each of those settings, the challenge is similar: the organization still has to make decisions, even when the historical record is not good enough and the people doing the planning do not have the capacity to run a major research effort from scratch.

What this enables

The core benefit of this approach is not that it predicts the future better than humans. The benefit is that it improves the planning process itself.

It helps organizations:

• broaden the scenario space beyond recent experience and avoid biases
• reduce the effort required to create planning inputs
• make assumptions explicit
• direct expert attention to the places where judgment matters most
• build a more transparent and defensible basis for inventory decisions

That combination is valuable well beyond emergency stockpiles. Any organization planning under sparse data and constrained expert capacity can benefit from a method that turns scenario development from a blank-page exercise into a structured review process.

Looking ahead

The next opportunity is not just better prompts. It is better institutional use of this capability: reusable planning workflows, stronger review routines, and more systematic capture of expert corrections over time.

Done well, that does more than improve an AI tool. It gives organizations a more scalable way to plan for events they cannot afford to ignore — even when they do not have the historical data, time, or internal capacity to model those events from scratch.

If your team is making high-stakes planning decisions with incomplete data, we can help you build a practical human-in-the-loop planning system around that reality. We design and implement scenario-based decision-support workflows that combine expert knowledge, structured review, and LLM-assisted draft generation to reduce planning effort and improve consistency. The result is a method your organization can use repeatedly — for stockpiles, resilience planning, supply risk, continuity, and other complex operational decisions.