Planning Was Never the Top of the Stack

The most expensive mistake in enterprise AI is not a model. It is a layer. For two years the industry poured money into making prediction and planning faster, and planning was never the top of the stack. The two layers where earnings are actually decided sit above the plan, and almost nobody has built them. That is the whole story behind the numbers everyone keeps quoting.

And the numbers are brutal. Gartner expects more than 40% of agentic AI projects to be canceled by the end of 2027, on escalating cost, unclear business value, and weak risk controls. MIT's 2025 study of enterprise AI put 95% of generative AI pilots at no measurable return on the P&L. McKinsey calls it the gen AI paradox: nearly eight in ten companies have deployed gen AI, and more than eight in ten report no material impact on earnings, with just 1% calling their programs mature. Two years of spend, and the bottom line has not moved.

The house explanation is immaturity. Another turn of the models, cleaner data, deeper integration, and the returns will arrive. There is some truth in it, and it is still the wrong diagnosis, which the analysts' own findings quietly confirm. Gartner names agent washing, chatbots and RPA relabeled as agents with nothing agentic beneath. McKinsey finds the wins are horizontal copilots whose benefits are real but too diffuse to show on the P&L, while the vertical, high-value use cases die in the pilot. MIT traces the failure to tools that never embed in the real workflow. None of that is a story about immature technology. It is a story about software aimed at the wrong layer.

Point software at prediction and planning and you get faster prediction and faster planning, which is precisely what the enterprise already had and precisely what did not move earnings. To see why, separate four things the industry has collapsed into one word. A forecast estimates what will happen. A plan is a sequence you follow if nothing surprises you. Above them sit the two layers nobody built: an engine that chooses the resilient enterprise action across hundreds of futures, and the posture that tells that engine what winning means. ML forecasting, deterministic planning, the decision engine, the posture. The market just spent billions making the first two faster. The last two are where the earnings live.

What you already have

ML demand forecasting is the first layer, and it is a commodity with a blind spot no accuracy will close. Extrapolation carries the past's commercial drivers forward as if they were fixed: the prices you charged, the promotions you ran, the share you held, the customers you had. But demand is conditional on those drivers, and the drivers are decisions you have not made yet. Reprice and the curve moves. Win a large account or lose one and the series you projected no longer describes the business you are in. This is not the tired point that you estimate unconstrained demand and then ration it to supply. It sits further upstream: there is no single unconstrained demand to estimate, because demand is a surface that shifts with every commercial lever you pull. A forecast that projects yesterday's drivers forward describes the one future you can be certain will not happen, the one where you change nothing.

Deterministic supply planning is the second layer, and it is where the vocabulary gets abused. Take the point forecasts, apply rules and constraints, return one plan under one set of assumptions, treated as certain. Useful, and now routinely relabeled as decision intelligence, which it is not. A deterministic planning engine with a conversational front end is NOT decision intelligence, it is still a deterministic planning engine now with a conversational front end. The interface got friendlier and the core did not move: fixed inputs, a single answer, and no way to represent the one material decisions are made of, which is uncertainty. Worse, a deterministic plan honors every assumption absolutely, so the moment one breaks, and one always breaks, the whole plan lurches, and planners quietly rebuild by hand the buffers the system was supposed to remove. A chatbot on that core does not make a decision engine any more than a steering wheel makes an engine. Everything here answers "what is the plan," and a plan is not a decision. It is the residue of a decision someone already made in silence, when they set the objective and the constraints the planner never questioned.

That silent act is the whole game, and the current state has nowhere to put it. And the conflation costs you a second time, after the fact: if the plan is the decision, the only thing you can grade is plan against actual, which can never separate a good decision that got unlucky from a bad one that got lucky. You end up unable to tell your own judgment apart from the weather, and a company that cannot measure its decisions cannot improve them.

Where the value begins

The two layers worth building rest on one foundation: demand, supply, and risk treated as three levers that continuously reshape one another, not three forecasts produced in sequence. Push price and demand moves and margin moves. Tighten capacity and the profitable shape of that same demand changes. Raise the odds of a disruption and both the price and the capacity you should want change with it. Model them apart and you miss the only thing that matters, which is how they trade against each other. On that foundation sit the two missing layers.

The third layer is the enterprise decision engine, and it is the one the stack skipped. Not one plan against one forecast, but the plan that holds up across hundreds of plausible futures drawn from your own history, chosen by a stochastic solve against a single enterprise objective, economic value, that still decomposes cleanly into every KPI beneath it, so service, inventory, expediting, capacity, and cash stay visible and, where you want them, act as constraints on the solve. Because there is one enterprise objective instead of a scoreboard per function, no silo can improve its own number at the enterprise's expense without the cost surfacing where the engine can see it. And here is the shift worth carrying out of this article: the deterministic instinct is to predict one future and plan against it, while the resilient instinct is to accept that you cannot, and to choose the action that performs well across the futures you are unable to rule out. Not be right about the future. Be hard to hurt across it. That single move is the difference between a plan and a decision.

This is also the layer where autonomy belongs, and belongs completely. Translating a chosen posture into a fully solved enterprise plan, and re-solving it hour after hour as the world moves, is a matter of speed and scale no human team can perform or should be asked to. A person cannot hold hundreds of futures and thousands of interacting levers in working memory. A machine can, every hour, without fatigue. Handing this to software is not a loss of control. It is the first time the control was real.

The fourth layer is the decision posture, and it is the one almost nobody names, even though it is where the enterprise actually decides. The engine finds the best plan under a posture. The posture is the choice the engine can never make for you: the objective itself, whether to maximize expected economic value, to minimize the damage at the fifth percentile, or to run a deliberate hybrid that defends the downside to a chosen percentile while chasing upside above the median; which constraints are hard and which are tradable; the service floors you guarantee and at what confidence; the risks you price in and the scope across which you carry them. That posture, once chosen, is encoded as the decision policy the engine solves against. Move it and the entire optimal plan changes while not one forecast has moved. This is the layer that decides whether a good quarter was earned or borrowed against the next bad one, and in most companies it is invisible. Someone fixed an objective in a configuration screen, or in a meeting no one remembers, and every downstream plan inherited a risk appetite nobody consciously chose. The organization has a posture. It just discovered it by accident.

Making it deliberate is the real prize, and it is where two traditions finally marry. Drawing the decision at the executive level, the actions, outcomes, levers, and dependencies laid out as a shared picture, has always produced clarity. But clarity is not determination. A map of every lever does not tell you where to set them. The marriage is exactly this: the decision model supplies the structure, the engine sweeps it, and every posture is priced and laid on an economic frontier with the KPI tradeoffs visible underneath. The engine can show you, to the dollar, what defending the fifth percentile costs in expected value, or what a tighter working-capital band does to service. What it cannot do is choose among the postures, because that choice is not a computation. It is a statement of what the enterprise is willing to lose and what it will protect at any cost, and you cannot optimize your way to a value.

Which yields a cleaner division of labor than the human-in-the-loop cliché. The machine solves when the objective is fixed. The human owns the objective, and is called back only for the rare calls where the stakes are high and the engine itself is unsure. Predict, plan, solve, govern: automate where the goal is given and the search exceeds human bandwidth, keep the one place where the goal itself is in question, and spend scarce human judgment on the few exceptions that earn it. This is not a human in the loop approving every move. It is a human above the work, holding the posture, and stepping in only where the machine raises its hand.

Planning was never the top

Strip it back and the whole mistake is one sentence: the industry automated prediction and planning, called the result intelligence, and stopped exactly where the value begins. ML forecasting extrapolates a past your own decisions will overwrite. Deterministic planning turns that into one certain answer and mislabels it. The decision engine, the resilient enterprise solve, is the layer missing from nearly every stack in the market, and the one a machine should run at a speed no room can match. The posture, the objective the enterprise is actually optimizing for, is the layer leaders should refuse to surrender even after the solving is no longer theirs to do.

Better forecasts and faster plans were never going to change the decisions, because the decision was never in the forecast or the plan. It was always two layers up, waiting for someone to build the engine beneath it and then choose, on purpose, what the enterprise is willing to risk. Planning was the part everyone could see. It was never the part that mattered.

On August 25 we'll show what Decision Engine and Decision Posture looks like, live, in working software: "Autonomous, Where It Counts — The Five Levels of Earned Decision Rights."

Registration link: vyan.ai/resources/webinars/autonomous-where-it-counts