Predictive systems learn from recorded outcomes. A task duration, a delay, a completed move. But those records do not capture how the work actually unfolded and the difference between what happened and how it happened is where prediction still falls short.
What predictive systems learn
In most distribution centers, predictive analytics relies on structured data captured through systems such as WMS, TMS, and ERP, where tasks are logged, timestamps are recorded, and durations are measured. Over time, this data becomes the foundation for forecasting throughput, estimating task performance, and identifying potential delays, allowing models to improve continuously as more historical outcomes are accumulated.
A typical record may show that a task took 12 minutes to complete, and that information is captured, stored, and used to refine future expectations. As more data is collected, the system becomes increasingly confident in its estimates, even though what remains unclear is how those 12 minutes were actually spent.
What the data doesn’t capture
A task may have been delayed before it began, traffic may have built up in the area, slowing movement midway through execution, an operator may have been waiting for equipment, or movement may have slowed due to temporary constraints in the area. In other cases, the same task may have been completed faster than expected, without any friction at all. From the system’s perspective, these scenarios appear identical, since they produce the same recorded outcome despite reflecting very different operational conditions.
Where learning falls short
This creates a gap in how predictive systems learn from the operation, not because the models themselves are insufficient, but because operations do not behave as a series of isolated events. They unfold continuously, with delays building gradually, flow slowing before it breaks, and capacity becoming constrained over time, while improvements emerge in the same way, as work accelerates or clears more efficiently than expected.
When models learn only from final outcomes, they capture the result but not the process that produced it, turning delays into numbers rather than sequences, and longer task durations into data points without preserving the conditions that led to them. Over time, this limits how prediction improves, as the system becomes more accurate in estimating averages but remains less capable of distinguishing between different underlying causes that lead to similar results.
Expanding what feeds the model
Improving prediction, in this context, is not a matter of collecting more historical data, but of representing the operation more faithfully. Traditional systems capture discrete events, recording when work starts, when it ends, and how long it takes, which is essential, but ultimately describes execution only after it has been completed.
What is missing are signals that reflect what happens in between.
Much of that information is physical in nature, relating to movement, waiting, interaction, and flow across space, all of which shape how work is performed but are not explicitly represented in event-based systems. Technologies such as computer vision and image processing make it possible to observe these aspects directly, capturing how movement evolves, how congestion forms, and how work progresses through the operation in ways that are not reflected in recorded events.
Rather than replacing existing systems, this layer extends them by adding context to what is already captured, allowing recorded outcomes to be understood as part of a broader process instead of isolated results. A task duration is no longer just a number, but an expression of how time was distributed across waiting, movement, and execution under specific conditions.
From outcomes to understanding
Over time, this changes what predictive systems are able to learn. Instead of training only on how long tasks take, models begin to learn how different conditions influence performance, making it possible to distinguish between delays caused by local traffic build-up, delays driven by resource constraints, and variability that is part of normal operations.
The result is not a different model, but a different understanding of the operation.
As distribution centers become more dynamic, this distinction becomes increasingly important, as systems that learn only from recorded results will continue to improve incrementally, while systems that learn from how operations actually unfold will improve in a more meaningful way.
The difference is not in the model. It is in what the model is able to learn from.
