A logistics planner can save $120 by sending one truck on a longer delivery round that arrives late at a customer. Is it a good trade?
There is no mathematical answer until the business says what a late arrival costs. The consequences might include an explicit contractual charge, overtime at the receiving site, a disrupted production line or a relationship that is difficult to price. An optimization model can take those consequences seriously, but it cannot invent their value.
This is where the familiar business language of priorities and promises becomes a mathematical penalty.
Prices for outcomes
Consider an illustrative decision between two delivery plans. Plan A costs $800 in transport and arrives 30 minutes late. Plan B costs $900 and arrives on time. If the business assigns a lateness cost of $2 per minute, the scored cost of A is $860, so A wins. At $5 per minute, its scored cost becomes $950, so B wins.
Nothing about the routes changed. Only the business's valuation changed. The objective makes that choice explicit: transport cost plus the modeled cost of lateness.
This is why a penalty weight should not be selected simply because it looks mathematically convenient. It influences the decisions the optimizer will recommend. If the penalty represents a real contract, accounting may provide a figure. If it represents customer experience, the team may need to estimate its consequences, test different assumptions or enforce a service threshold instead.
Service is not always a straight line
A company may value the first few late minutes differently from a two-hour delay. For example, a delivery within a five-minute grace period might be acceptable; after that, each minute adds cost; after an hour, a missed production connection could trigger a much larger loss.
Mathematically, this is a piecewise relationship: the cost changes its rate or structure at particular thresholds. The model can represent a fixed late-delivery charge, a cost per minute, bands of severity or a hard limit beyond which a plan is not permitted.
Another approach is to focus on service promises directly. Instead of estimating an uncertain dollar value for disappointment, the business might require at least 98% of orders to arrive on time and then minimize operating cost among the plans that meet that standard. The percentage is an illustrative policy, not an industry benchmark.
Customer satisfaction is not a single meter
Customer satisfaction is especially difficult because it is partly observed and partly inferred. Survey scores, renewal rates, complaints, missed appointments and delivery accuracy capture different aspects of an experience. Increasing one does not guarantee improvement in another.
A workable model usually starts with the decisions that the organization can influence and the observable outcomes they affect. If late delivery is a major source of complaints, model the delays and service windows directly. If reliability matters more than average speed, consider worst-case lateness or the share of commitments met. If fairness across customers matters, examine whether the optimizer repeatedly sacrifices the same group.
A model should not claim to maximize happiness because it maximizes a proxy. Its value is in making a chosen, testable relationship between decisions and outcomes clear.
A decision worth revisiting
A penalty is a hypothesis about what matters. The best way to test it is to compare plans. How much extra transport cost would eliminate the last late delivery? Does the answer change if the penalty doubles? Does a small set of customers experience most of the disruption?
Optimization then becomes more than a machine for ranking options. It gives the business a way to examine its priorities and decide what its commitments are worth.