It’s a scenario familiar to anyone involved in demand planning: In the third week of every month, a demand planner opens a spreadsheet, pulls a statistical baseline out of the planning system, and begins overriding it line by line. That’s because the planner knows things the system does not: this customer is depleting stock, that competitor moved on price, the last promotion pulled volume forward. Based on that knowledge, the planner revises the numbers. The results are reconciled across several review meetings, and the new plan is approved.
Analyze the distribution center that the plan supplies, and a mismatch is immediately apparent. Much of the inventory on the shelves is there not to satisfy customer demand, but because nobody was confident enough in the forecast to carry less. Some of this inventory is safety stock against uncertain demand; some reflects forecasts that consistently overestimate what customers will buy. That makes forecast accuracy a working-capital and growth issue—not simply a planning KPI.
Companies have spent decades trying to improve demand planning with more sophisticated statistical models, advanced planning systems, and increasingly elaborate sales and operations planning (S&OP) processes. Yet the basic pattern has proven remarkably stubborn: The planners are not the problem. They are compensating for a system that never knew it was broken, yet the planners absorb the pain.
Recent advances in AI offer a way to attack both sides of the problem: the quality of the forecast and the cumbersome process built around it. More sophisticated machine learning models can draw on a much wider set of signals—from competitor pricing and freight costs to shifts in consumer behavior—to produce higher-quality forecasts. Agentic AI can then take on much of the work surrounding those forecasts: gathering and structuring information, investigating anomalies, testing assumptions, recalibrating models, explaining what changed and why, preparing scenarios and recommendations, and eventually carrying approved decisions back into planning systems.
That distinction matters. Large language models are not forecasting engines, and machine learning models are not communicators. The opportunity comes from bringing the two capabilities together. Predictive AI can produce a forecast that merits greater trust; agents can help people understand, challenge, and act on it. But the larger opportunity is not simply a better forecast or a more automated S&OP process. It is a different planning system—one that continuously senses changes in the market, predicts their implications, explains what has changed, recommends a response, and—within defined guardrails—helps execute it. AI can begin to close the loop from signal to decision to action.
This also explains why simply adding agents to today’s S&OP process is not enough. Automating a process that produces a mediocre forecast simply produces a mediocre forecast faster. Yet improving the forecast without changing the work around it leaves value on the table as well. The opportunity is to improve the intelligence underneath S&OP while redesigning the process around what that intelligence can now do.
That opportunity is substantial, but it is also early. Better forecasting has been demonstrated and quantified; fully agentic S&OP remains frontier territory rather than established best practice. This article draws on both: what companies have already learned about building forecasts that materially improve business performance and what early deployments suggest about using agentic AI to create a new model for enterprise planning. The combination of agentic AI and predictive AI therefore produces a forecast worth trusting; agents help people understand, challenge, and act on it.
The value of getting the forecast right
The financial stakes are considerable. A poor forecast can tie up millions of dollars in inventory, create costly waste, or leave a company without enough product to meet demand. Improving accuracy can therefore affect working capital, margins, growth, and customer service—not simply the performance of the planning function.
The value comes through three channels: Better predictions can improve decisions relating to inventory, service, production, and capacity; faster sensing and decision-making allow companies to respond sooner to changes in demand; and automation can reduce the substantial effort required to turn forecasts into coordinated action. The prize is not forecast accuracy itself—it is better business performance.
The most immediate and measurable opportunity is usually inventory. Companies hold extra inventory to protect themselves against forecast uncertainty, while systematic overforecasting can lead to additional excess stock. The less reliable or more biased the forecast, the more inventory companies tend to carry. By improving forecast accuracy and reducing bias, companies can lower inventory without sacrificing service levels. Our research, consistent with external benchmarks, puts the relationship at roughly 2.7 percentage points less finished-goods inventory for every percentage point of forecast accuracy gained. This relationship, however, can vary significantly across industry and level of company maturity.
In chemicals businesses where we have tracked the impact of demand-planning transformations, the results have been substantial. We have seen forecast accuracy improvements of 15 to 20 percentage points and, in the first wave, on-hand inventory reduced by 20 to 30 percent. For a business unit with $1 billion in revenue, that can release an eight-figure sum of cash in the same quarter.
In theory, the inventory reduction could be considerably larger. In practice, companies rarely capture all of it at once. A more accurate forecast may indicate that a company can safely hold much less inventory, but planners and business leaders need to trust the new forecast before they are willing to make that change. They typically reduce inventory marginally at first, watch how the model performs against actual demand, and make further reductions as confidence builds. This distinction is important: improving accuracy creates the opportunity; changing business decisions converts that opportunity into cash.
In businesses that sell perishable goods, excess inventory creates a different cost. A manufacturer can often hold extra finished goods against future demand; a grocer that orders too much fresh food may have to throw it away. Underforecasting creates the opposite problem: When a product is unavailable, the loss may never appear on a management report. It is a sale that did not happen and a potential lost customer.
Better forecasts, therefore, give companies a strategic choice: maintain the same service level with less inventory, or use greater accuracy to improve availability without adding inventory. That choice has implications for working capital, revenue, and customer service—and belongs with business leaders rather than being embedded, often implicitly, in a planning parameter.
Inventory is only one part of the opportunity. Better forecasts can help companies allocate production across plants and lines, adjust the product mix to expected demand, and plan capacity with fewer last-minute changes. They also give procurement teams a clearer view of future material needs, supporting earlier supplier commitments and, when combined with price and supply market insights, more informed strategic buying. Capturing these benefits requires connecting the demand outlook to decisions that reflect costs, margins, and operational constraints—not simply producing a more accurate number.
Companies are already capturing some of this value. McKinsey analyzed roughly 100 companies that have implemented advanced planning systems and found a median inventory improvement of 11 to 15 percent. A typical implementation returned about 2.6 times its investment, with all-in investment averaging $20 million to $30 million per $15 billion of revenue. But the results vary enormously: Top-quartile implementations returned about 13 times their investment. And around 42 percent of organizations report no positive or clearly measurable return from AI in planning.
Why do some companies capture so much more value than others? In our experience, much of the answer lies in choices that are easily overlooked: what information the forecast can see, the level of detail at which it operates, the time horizon it predicts—and what the organization actually does with the result.
Rewired, Second Edition
This updated edition offers brand-new insights into cutting-edge AI solutions—and what it takes to implement them—as well as the new economics of digital and AI transformations.
Building a forecast worth trusting
Capturing more of this value starts with improving the forecast itself. Many planning systems still rely heavily on models that extrapolate from a product’s own historical sales. The systems capture patterns such as trend and seasonality, but they cannot account for changes in competitor pricing, macroeconomic factors, or changing consumer behavior. For stable, high-volume products, that may be adequate. For more volatile portfolios, it can put a ceiling on accuracy.
Machine learning can help overcome some of these limitations, but doing so requires more than choosing a sophisticated algorithm. A useful forecast needs to draw on the signals that truly reflect demand, operate at the level where the business makes decisions, and look far enough ahead to inform those decisions.
McKinsey recently worked with a multinational tire producer whose low forecast accuracy was driving high inventory costs. Its existing planning solution relied on relatively simple models across a complex international supply chain. The company’s first step in addressing the problem was to broaden the scope of its forecasts. Eighteen expert interviews and two cross-functional workshops identified more than 250 potential drivers of demand; quantitative testing narrowed those to roughly 30 with predictive value. These included some nonobvious signals, such as long-distance freight costs, raw material indices, passenger-car registrations, competitor share prices, and consumer search volume for used tires.
The company also reconsidered where and how far ahead it needed to forecast. Demand forecasts needed to be useful at the level where production decisions were actually made—down to a particular material and plant—and over horizons that reflected physical lead times rather than simply the monthly reporting calendar. The team tested more than 30 machine learning models, including univariate, multivariate, and ensemble approaches, and selected among them based on measured out-of-sample accuracy.
The results were substantial. Forecast accuracy improved by ten percentage points at the material-code level and by about 25 percentage points at the material-code-by-production-plant level. Annual operational cost savings reached over $12 million.
But achieving those results required overcoming another challenge: People had to trust the new forecasts enough to use them. A planner confronted with an unexpected number still needed to understand why the forecast had changed and whether there was a reason to override it. The team therefore added an agentic layer for forecast analysis and explainability, using SHAP (Shapley additive explanations) values, a method for showing how individual variables contribute to a prediction. A planner could see, for example, that the forecast for a particular product and region had moved because freight costs rose and registrations softened. The model was no longer simply handing planners a number; it was giving them something they could examine and challenge.
The tire company’s experience shows how machine learning and agentic AI can complement each other. Machine learning produces a more informed forecast; agents help do the work required to make that forecast usable: gathering and structuring information, investigating anomalies, testing assumptions, validating outputs, and turning results into explanations and recommended actions. Humans retain authority over decisions that exceed agreed-upon risk, confidence, or business thresholds. Connecting this intelligence to enterprise resource planning systems can close the loop, while actual outcomes and human overrides provide the feedback needed for continuous learning.
From better forecasts to a different planning process
In a traditional S&OP process, planners spend much of each cycle assembling data, investigating discrepancies, preparing materials for review meetings, reconciling different views of demand, and then translating the decisions made in those meetings back into planning systems. A better forecast improves the starting point without necessarily changing that work.
Agentic AI creates the possibility of redesigning it. Rather than waiting for planners to assemble the information needed to interpret a forecast, agents can retrieve and structure those inputs themselves. They can investigate unusual results, compare outputs against benchmarks, surface potential problems, prepare explanations, and recommend actions. An orchestrating agent can coordinate that work across the planning cycle, escalating decisions that require human judgment.
McKinsey worked with a life sciences client whose planning process involved nearly 100 meetings and multiple revision cycles per monthly forecast.
The client’s original process illustrates why a better model alone is insufficient. Planning teams updated assumptions, revisited models and schedules, filled reporting templates, and refreshed presentation slides; stakeholders reviewed the forecast; and assumptions were revised through repeated cycles, after which the approved forecast was loaded into the system. Interviews revealed multiple reviews, adjustments, and Excel workarounds, including eight to ten rounds of manual changes in parts of the planning process. External factors were considered qualitatively, but there was no systematic flow of driver data into the forecast, and the process lacked automated postmortem analysis to identify recurring methodological issues.
The redesigned process targeted these areas for improvement. External drivers were sourced systematically and incorporated quantitatively with greater granularity. AI-based models generated the forecast, automated back testing could compare forecasts with outcomes, and explainability could show which drivers shaped the result. Automated postmortem analysis identified recurring errors and supported continuous learning. Rather than using meetings to assemble inputs and reconcile different versions of the forecast, teams entered reviews with a common, evidence-based view of the forecast and its drivers.
The results illustrate both sides of the value equation. The redesigned process reduced forecast error by approximately 50 percent. At the same time, by automating transactional steps such as data collection, forecast generation, and reporting, the new process led to an 83 percent reduction in people hours required. This changes the purpose of the planning meeting. Routine consolidation, accuracy checks, and preparation of explanatory material can increasingly happen beforehand and, where data refreshes allow, continuously. Meetings can focus on decisions that require cross-functional judgment: whether to accept a model’s recommendation, how to respond to an unusual external signal, how to resolve trade-offs across operational and commercial teams, and when business context justifies an override. Human judgment remains essential where data are incomplete, decision rights cross functions, or implications cannot be reduced to a model parameter.
The more consequential shift comes when agents move beyond analysis and recommendations into execution. An agent that identifies an inventory problem but leaves a planner to enter the resulting decision into another system has made the existing process faster. An agent that can write an approved parameter back into the planning system, release a replenishment proposal, or trigger procurement within defined limits begins to remove some of the handoffs on which the old process depends.
How far companies take that autonomy will vary. Most deployments today remain at the advisory end of the spectrum: AI proposes, and people approve. The next step is conditional autonomy, where agents can make or execute specified decisions when they fall within agreed parameters and escalate exceptions or higher-risk decisions to people. In regulated environments, the boundary may deliberately remain more conservative; pharmaceutical and medical-products companies, for example, use human-in-the-loop architectures to satisfy quality requirements.
The larger question is what happens to S&OP once more of this work can occur continuously. The traditional process is organized around a monthly cadence partly because people need time to assemble information, reconcile competing views, prepare materials, and convene decision-makers. If agents can perform much of that work as new information arrives, the monthly meeting no longer has to carry the same burden.
That does not necessarily mean the S&OP meeting disappears. It could become smaller and more focused, with routine reconciliation happening beforehand and senior attention reserved for consequential exceptions and trade-offs—for example, a capacity constraint that cannot be resolved automatically, or a demand shift requiring a commercial rather than purely operational response. Planning, in other words, can move from a periodic process toward a more continuous, exception-driven decision system.
What changes for the planner
A more automated planning process changes the planner’s job. Today, demand and supply planning teams spend much of their capacity assembling data, producing baselines, investigating discrepancies, and preparing review materials. As agents assume more of that work, planners can spend more time on decisions requiring judgment: challenging an unexpected forecast, assessing market signals, preparing demand scenarios, or deciding when the system should not act on its own.
The potential shift is substantial. In the target state we design, the people hours required for many of today’s planning activities can fall by up to 50 percent. (This figure is a target state for a full agentic redesign including the execution layer, not a realized cross-industry average; organizations that simply add advisory AI to a conventional planning process have seen much smaller effects.) These savings come mostly from automating transactional steps in the demand-forecasting life cycle, such as data collection, forecast generation, and reporting. That can free up capacity for evaluation of forecast results, handling exceptions, and improving the AI-led methodology.
The roles that remain require a different mix of skills. Planners need to understand the models well enough to challenge their outputs, recognize when an exception deserves attention, and bring commercial and operational judgment to decisions the system cannot resolve. Data scientists and engineers, meanwhile, are needed to maintain the forecasting models, agents, integrations, and the underlying infrastructure.
This is not a trivial workforce transition. Presenting it simply as an augmentation understates what a full agentic redesign can mean for planning organizations. Some activities—and potentially some roles—may require substantially less human capacity. At the same time, the work that remains becomes more focused on judgment, supervision, and decision-making. The scarce capability is no longer producing a number. It is supervising an AI-enabled decision system and knowing when that system is wrong.
Both the tire manufacturer and life sciences client undertook sustained efforts to build internal capabilities, combining online courses and in-person workshops. They raised general knowledge of AI capabilities among managers and business stakeholders while building deeper knowledge of AI forecasting and orchestration methodologies among demand planners.
Building toward agentic S&OP
Our work suggests that capturing this opportunity requires capabilities that have traditionally been treated separately: planning expertise, predictive modeling, agentic AI, data and software engineering, enterprise system integration, and operating-model change. Bringing these together is what turns an AI use case into an end-to-end planning transformation. McKinsey’s Rewired framework for AI transformation provides a broader model for this transformation: linking the business road map to talent, operating model, technology, data, adoption, and scaling. In demand planning, that means treating AI not as an isolated technology deployment but as a redesign of the decisions, workflows, systems, and roles that collectively produce the plan.
Seven questions can help leaders understand where their planning organization stands today:
- In which area would better forecasting create the most value—inventory, service, production efficiency, or sourcing cost—and have we quantified it?
- Can our forecasting approach see anything outside our own history—and, if not, what is it structurally blind to?
- At what granularity do our planning decisions get made, and do we forecast at the same level or at a more or less aggregated level?
- Are our forecast horizons set by our physical lead times or by our reporting calendar?
- What share of our planners’ time goes to assembling data rather than exercising judgment, and where do we sit today between advisory AI and conditional autonomy?
- If the forecast improved by ten points tomorrow, which parameters, targets, and decision rights would have to change for that improvement to reach the balance sheet—and who owns them?
- When a planner overrides the system forecast, do we track whether the override improved accuracy?
The last question is often particularly revealing. In many organizations, nobody knows the answer, which means one of the most frequent human interventions in the planning process has never been systematically measured.
Putting human judgment where it matters
For decades, S&OP has been built around an unavoidable constraint: No model could see enough of the world to be trusted on its own. Companies responded rationally. They surrounded the forecast with people who could supply missing context, meetings where competing views could be reconciled, and buffers that protected the business when the resulting plan was wrong.
AI begins to loosen that constraint. As predictive models become better at forecasting, and agents become better at interpreting those forecasts, orchestrating the work around them, and acting on approved decisions, companies can begin to reconsider much of the scaffolding built around the old system’s limitations.
The goal is not an autonomous planning process for its own sake—or is it to remove human judgment from S&OP. It is to become much more deliberate about where that judgment belongs. People are extraordinarily valuable when the future departs from what the system knows: when a customer behaves unexpectedly, a market turns, competing objectives collide, or a decision requires accepting a risk that cannot be reduced to a parameter. They are considerably less valuable when copying data between systems and reconciling numbers that machines can reconcile themselves, which brings us back to the planner and the spreadsheet.
The planner was never the problem; the planner was the solution to the system’s limitations. The opportunity now is to build something different: an AI-enabled planning system that can sense, predict, explain, support decisions, execute within appropriate guardrails, and continuously learn from outcomes. As that system matures, S&OP itself can evolve—from a periodic process for reconciling forecasts into a more continuous enterprise decision system, with human attention concentrated on the exceptions, trade-offs, and risks that matter most.
