NEW: The Decision Factory — a novel about decisions under uncertainty. Get it on Amazon
Decision Science · · Adam DeJans Jr.

Why Your Simulation Is Not a Decision Policy

Many supply chain teams build sophisticated simulations but still struggle to make better decisions. Learn the difference between evaluating a decision and generating one.

decision sciencesimulationoptimizationsupply chainuncertaintypolicy design

Why Your Simulation Is Not a Decision Policy

One of the most common mistakes I see in supply chain analytics is the belief that a simulation model is a decision model.

A simulation can tell you what happens if you make a decision. It does not necessarily tell you what decision to make.

This sounds obvious, but entire projects fail because teams spend months building increasingly detailed simulations without ever defining the actual decision they are trying to improve.

The result is usually a beautiful digital twin, thousands of scenarios, impressive dashboards, and almost no measurable business impact.

Start With The Decision

Before discussing forecasts, machine learning, optimization, simulation, or artificial intelligence, ask a simple question:

What decision are we trying to make?

Examples include:

  • How much inventory should we buy?
  • Which customer orders should receive constrained supply?
  • Which products should be replenished this week?
  • Which shipments should be expedited?
  • How much capacity should be reserved for future demand?

If nobody can clearly state the decision, stop immediately.

The modeling effort is not ready.

A Useful Decision Framework

Most operational decisions can be framed using five components.

State

What do we know right now?

Examples:

  • Inventory positions
  • Open purchase orders
  • Supplier status
  • Current demand signals
  • Capacity availability
  • Financial constraints

The state describes the world at the moment a decision is made.

Decision

What can we actually control?

Examples:

  • Order quantities
  • Inventory targets
  • Allocation quantities
  • Production schedules
  • Transportation choices

If a variable cannot be changed by the business, it is not a decision variable.

Uncertainty

What happens after we make the decision that we do not know today?

Examples:

  • Demand realization
  • Supplier delays
  • Manufacturing disruptions
  • Transportation variability
  • Product returns

This is where forecasting belongs.

Forecasts are descriptions of uncertainty.

They are not decisions.

Transition Function

How does the system evolve after the decision is made?

This is often where simulation is valuable.

Inventory moves.

Demand arrives.

Lead times materialize.

Production completes.

Customers place orders.

The simulation describes how the system behaves.

Objective Function

How do we measure success?

Examples:

  • Profit
  • Cash flow
  • Service level
  • Fill rate
  • Revenue
  • Inventory turns
  • Total cost

Without a measurable objective, optimization becomes impossible.

Where Simulation Fits

Simulation is primarily an evaluation tool.

Suppose you have three replenishment policies:

  • Policy A
  • Policy B
  • Policy C

Simulation helps estimate the consequences of each policy under uncertainty.

It can answer questions such as:

  • Which policy generates higher expected profit?
  • Which policy has lower downside risk?
  • Which policy produces more stable service levels?

These are important questions.

However, simulation alone does not generate the policies.

It evaluates them.

The Practical Problem

Imagine a purchasing problem with:

  • 5,000 products
  • 52 weekly decisions
  • Supplier constraints
  • Container constraints
  • Budget limits
  • Service targets

The number of possible policies becomes astronomical.

Running simulation on random policies is not a strategy.

It is guessing.

At this point you need something capable of searching for better decisions.

Examples include:

  • Mathematical optimization
  • Dynamic programming
  • Approximate dynamic programming
  • Bayesian optimization
  • Reinforcement learning
  • Heuristic search

The exact technique matters less than understanding the role it plays.

Something must generate candidate decisions.

Something must evaluate them.

These are different jobs.

Questions To Ask Before Building Anything

When starting a project, I typically ask:

  1. What decision is changing?
  2. How often is the decision made?
  3. What information is available at decision time?
  4. What uncertainty arrives afterward?
  5. What constraints are real?
  6. What objective is being optimized?
  7. How will success be measured financially?
  8. What action will be taken differently if this project succeeds?

If these questions cannot be answered, the project is not ready for implementation.

Common Failure Modes

Forecast First

Teams spend months improving forecast accuracy.

Nobody defines the decision.

Nobody measures economic impact.

The forecast improves.

The business outcome does not.

Digital Twin First

Teams build a detailed simulation.

The model becomes increasingly realistic.

No decision policy is ever implemented.

The simulation becomes a reporting tool.

Constraint Blindness

A policy performs well in simulation because realistic operational constraints were ignored.

The solution cannot be executed.

Metric Confusion

The project optimizes service level.

Leadership cares about profit.

Nobody notices until deployment.

What To Do In Practice

For most real-world supply chain projects:

  1. Frame the decision first.
  2. Define decision variables explicitly.
  3. Document uncertainty separately.
  4. Build an evaluation environment.
  5. Develop policies that can generate decisions.
  6. Compare policies using simulation.
  7. Measure economic outcomes.
  8. Deploy the policy, not the simulation.

The simulation is the laboratory.

The policy is the product.

Confusing those two concepts is one of the fastest ways to spend a year building something sophisticated that nobody uses.