Melvin Conway famously observed:
“Organizations which design systems… are constrained to produce designs which are copies of the communication structures of these organizations.”
In the realm of Decision Intelligence, this law manifests in the traditional divide between Machine Learning (ML) and Operations Research (OR). Currently, most organizations maintain a communication structure where ML teams provide a single point forecast, which the OR team then treats as an absolute truth to find an "optimal" solution. This "Predict-then-Optimize" workflow is a direct mirror of siloed departments that only exchange narrow, deterministic data, ignoring the vital uncertainty inherent in the real world.
Does this mean we must merge ML and OR into a single department?
Conway’s Law does not necessitate a total organizational merger to achieve better results, despite common misreadings of the text. We can preserve specialized departmental responsibilities - allowing ML to focus on patterns and OR on constraints - while fundamentally changing what, how, and when they communicate.
The failure of current systems is the transmission of point forecasts that strip away critical context. Consider Gurobi's avocado pricing and distribution example: a point forecast might predict demand, implying an "expected zero waste" scenario. However, optimizing for this single number is fundamentally flawed. It results in decisions that neither account for the risk of low demand - leading to significant spoilage - nor the opportunity to sell much more than the average prediction during a spike. The decision fails because the communication structure suppressed the reality of uncertainty, not because the teams were separate or unskilled.
Modern Solvers Enable Richer Communication
The solution lies in evolving the "handshake" between departments without the friction of a formal reorganization or the time-intensive upskilling of teams. InsideOpt's Seeker solver enables this shift by allowing ML specialists to communicate full distributional forecasts rather than flattened point estimates. In Seeker, OR specialists can even embed the ML models they receive from another department to make decision-dependent forecasts.
By passing the posterior distribution or the ML model itself, the OR team can finally account for the true landscape of uncertainty. This change in communication structure enables massively better decisions without requiring ML engineers to become optimization experts or OR practitioners to retrain in deep learning. We don't need a new org chart or mass upskilling.
We simply need a communication protocol that carries the weight of reality. And a solver that supports it.