The Step Change Supply Chain Planning Has Been Waiting For

I have spent 30 years inside supply chain planning systems, solutioning them, selling them, implementing them, and watching each generation deliver real value. Inevitably, the benefits flatten with time and the limitations of the approach emerge. Every approach is bounded by the computing power of its era. It is time for the next step.

Supply chain planning has plateaued because nearly every planning system on the market still runs on deterministic math: one set of inputs, one solve, one optimal plan, with scenario copies on the side for what-if questions. They respect uncertainty only to the extent that it can be fit to a modeled distribution. Reality often does not cooperate. Demand, supply, and lead times arrive as ranges of possibility, and a plan built on single numbers starts decaying the moment it is published. Simulation and probabilistic methods model that reality directly. The math is eighty years old. What changed is that the compute and AI models needed to run it every day have quietly become affordable. That is why I believe the plateau is finally about to break.

Why do generations of planning technology plateau?

Every generation of advanced planning has delivered a genuine step change and then flattened, because every generation refined the same core method instead of replacing it. MRP put the bill of materials on a computer. The APS engines of the 1990s added optimization and constraint solving and created enormous value. S&OP and integrated business planning platforms connected the plan to the money. Concurrent planning made the whole model recalculate in seconds. Each wave was real progress. Each wave flattened.

The pattern is too consistent to blame on any one vendor, and I say that having carried a bag for several of them. Each generation produced a better answer to the same question: given one set of assumptions, what is the single best plan? Some supply chain prophets are questioning the question.

Each generation gave a better answer to the same question. The question itself was the ceiling.

How do today's planning systems actually handle uncertainty?

Almost all of them handle uncertainty the same way: a deterministic solve at the core, with scenarios and what-if copies alongside it. A recent independent review of the major planning platforms found deterministic cores across the mainstream, with probability admitted only at the edges, as a safety stock module here or a scenario workspace there.¹ Scenario planning is the industry's standing concession to uncertainty: run the base case, run a high case, run a low case, and pick. Three hand-built snapshots of a future that has thousands of ways to unfold.

I know the add-on path first-hand. At SmartOps we built multi-echelon inventory optimization that treated demand and supply variability honestly, with real distributions and real stochastic math. And for the results to be usable, they had to be translated back into safety stock targets and service levels that a deterministic plan could digest, because that is what the surrounding systems understood. SAP acquired SmartOps in 2013, and that engine still does its work inside a deterministic planning core today. We got probability into the building. We could not get it into the driver's seat.

To be fair to the field, a probabilistic minority exists. A few vendors carry distributions deeper into their platforms, and simulation-native tools do excellent work in network design, stress-testing a supply chain before you commit capital to it. But across operational planning, the daily decisions about what to buy, make, and move, the mainstream core remains a single solve. Joannes Vermorel, the CEO of Lokad, reached the same conclusion: most of the market now says the word probabilistic, and almost none of it has rebuilt the stack underneath.¹ His assessment carries a self-serving motive, since it points to Lokad's own commitment to the new approach, but the review reads directionally accurate. I imagine some vendors will point out areas where they feel they have tried to become more "probabilistic".

If probabilistic methods are better, why hasn't planning used them?

Probabilistic simulation stayed out of daily planning for one reason: it was computationally expensive, so for eighty years it was rationed for occasional, high-stakes analysis. The math itself is anything but new. Stanislaw Ulam worked out the modern Monte Carlo method in 1946, pondering solitaire odds while recovering from an illness, and developed it with John von Neumann for the nuclear weapons work at Los Alamos. Nicholas Metropolis named it after the casino where Ulam's uncle liked to gamble, and the first computerized simulations ran on ENIAC in 1948.² From there the method spread wherever the stakes justified the compute. Physicists modeled neutrons with it. Oil companies modeled reservoirs. Meteorologists built ensemble forecasts. By 1964 the Harvard Business Review was teaching executives to simulate capital investments instead of trusting a single-point projection.³

Supply chain got the same treatment, on the same terms. Simulation was reserved for the occasional strategic study: redesign the distribution network every few years, stress-test a new plant, model a major sourcing shift. If the model took a week to run, nobody minded, because nobody was going to run it again next Tuesday. The one place it never reached was the daily planning desk, where decisions cannot wait a week. The math was rationed by the cost of computing it, and for decades the deterministic single solve was the reasonable compromise.

What changed to make probabilistic planning practical?

The compute constraint that rationed simulation for eighty years collapsed in the last three. Simulating thousands of possible futures across a full supply chain network every night used to carry a supercomputer budget. Today that class of computation rents by the hour: cloud GPU prices fell roughly 44 percent in a single cut last year, with capable hardware available for a few dollars an hour, and simulation workloads parallelize across cheap cloud capacity almost perfectly.⁴ At the same time, modern machine learning models changed what a forecast can be. They produce full probability distributions natively, a range of outcomes with likelihoods attached, where the previous generation produced a single number and an apology.

Which leads to the uncomfortable part. The barrier to probabilistic planning is no longer computational. The same market review that mapped the deterministic mainstream concluded that the incumbents simply have not rebuilt their architectures, even though the compute now permits it.¹ After 30 years, the plateau has become a choice.

Should every planning horizon use the same method?

No. To use a maritime analogy, long and midterm planning set the general bearing of the business, daily execution steers it through the shallows, and after 30 years I am convinced the two horizons deserve different methods. Over the long haul you are pointing the ship: where to build capacity, how to position inventory, which lanes to contract. Deterministic optimization has always done that work well, and it should keep doing it. The execution horizon is different. Day to day you are fine-tuning around obstacles, guiding the ship through the channel at the destination port, and this is exactly where traditional methods have always fallen short.

Some vendors have claimed the answer is a single solve across every horizon. I have implemented enough of these systems to disagree. Attempting it produces unstable solutions, because assumptions that hold at one horizon do not stick at the other. The objectives may look similar on paper, but the near term faces acute realities and constraints that deserve their own approach.

The World Cup final kicks off this Sunday, so consider a soccer version of the problem. Back in June, picking which teams would survive the group stage was a long-horizon call built on form, squad depth, and the draw. Today I have the luxury of knowing the final two are Spain and Argentina, and this weekend's question runs on entirely different information: the injury report, card status, even the latest TikTok trash talk. None of that mattered for the group-stage call, and almost none of it can be priced into an optimization objective. It can, however, be simulated. That is the execution horizon in one sentence: the factors that decide the outcome are the ones a single solve cannot hold.

What does planning look like when it starts from the odds?

Planning built on probabilistic simulation gives the planner the current odds on every decision, updated as reality unfolds, instead of one plan and a wall of alerts. Keep the deterministic bearing for the long haul. On the execution horizon, every purchase order, production run, and stock transfer is a bet, and the odds on each one shift as actual events land. A system that frequently simulates thousands of futures can tell a planner the current probability of a stockout, an excess position, or a missed service commitment, and which action moves those odds most.

I am not the only one who sees the industry this way. Lora Cecere has argued for years that planning technology was built for transactional efficiency rather than for deciding what to do next, and that market signals should feed probabilistic, economically grounded decisions.⁵ Trevor Miles, who coined the term concurrent planning at Kinaxis, put it bluntly when he joined a probabilistic planning company in 2023: supply chain decisions are bets, and the fatal flaw of current planning systems is deterministic math that leads executives to make bad bets without understanding the risks.⁶ When the people who built the last generation start saying the math must change, the plateau is ending.

This conviction is why, when I chose a technology to help commercialize, I chose River by Pillar AI, which runs frequent probabilistic simulation as its core operating method rather than as a what-if module on a deterministic engine. I have written about the treadmill it replaces in The Replan Trap, and about why wiring agents to yesterday's logic just automates bad decisions faster in Agents Are Not Enough.

Proxigee Services, based in Pittsburgh, partners with technology companies to bring innovations like this to market, guides implementation firms navigating the planning landscape, and helps companies put AI to work on their hardest operational problems. Commercializing a new category is a team sport, a case I made in Making Waves with Partner Ecosystems.

The math is eighty years old. The compute arrived in the last three years. The only thing still deterministic about supply chain planning is the industry's habit of asking for one right answer. Break the habit, and the plateau breaks with it.