Case Study 1: Procter & Gamble's Demand Sensing — From Forecast to Sense-and-Respond

Introduction

For most of the twentieth century, Procter & Gamble ran one of the most sophisticated demand planning operations in the world. The company's supply chain managed over 300 brands across more than 180 countries, serving approximately 5 billion consumers daily. Its traditional forecasting approach — built on historical shipment data, promotional calendars, and years of institutional knowledge — was the industry benchmark.

By the mid-2010s, that benchmark was no longer good enough.

The fundamental problem was not that P&G's forecasts were inaccurate — by industry standards, they were excellent. The problem was that "accurate by industry standards" still meant forecast errors of 30-50 percent at the store-SKU-week level. For a company shipping billions of dollars of consumer goods, those errors translated into two costly outcomes: excess inventory (overstock) and empty shelves (stockouts). P&G estimated that these twin problems cost the company and its retail partners hundreds of millions of dollars annually.

This case study examines how P&G transformed its demand planning approach from traditional forecasting to what it calls "demand sensing" — a near-real-time, multi-signal system that fundamentally changed how the company aligns supply with demand.

The Limitations of Traditional Demand Planning

P&G's traditional demand planning process, like most consumer goods companies, followed a monthly cycle:

1. Baseline forecast. Statistical models (primarily exponential smoothing and ARIMA variants) generated baseline demand forecasts using 2-3 years of historical shipment data.

2. Demand review. Category planners adjusted the statistical baseline based on known future events: planned promotions, new product launches, price changes, competitive activity.

3. Consensus planning. Sales, marketing, finance, and supply chain leaders met monthly to review and agree on a single demand plan. This process, known as Sales & Operations Planning (S&OP), typically took 5-10 business days.

4. Execution. The agreed plan drove production scheduling, raw material procurement, warehouse allocation, and distribution.

This process had three structural weaknesses:

Latency. By the time the monthly S&OP process concluded, the demand plan was based on data that was 3-6 weeks old. In a world where consumer behavior could shift overnight — due to weather, viral social media, a competitor's surprise promotion, or a sudden health scare — a plan that was weeks old was already stale.

Granularity. The traditional process planned at a relatively high level of aggregation — typically brand-by-region-by-month. But supply chain execution happens at a much finer granularity: SKU-by-distribution-center-by-week. The disaggregation from planning level to execution level introduced additional error.

Signal poverty. The traditional process relied almost exclusively on two data sources: historical shipments and the promotional calendar. But consumer demand is influenced by dozens of factors — weather, local events, social media trends, macroeconomic sentiment, competitor actions — that the traditional process ignored.

Business Insight: P&G's experience illustrates a paradox common in large organizations: the planning process itself becomes a bottleneck. A monthly S&OP cycle that takes 10 business days to complete and uses data that is weeks old cannot respond to a demand signal that is hours old. The process was designed for a world that moved slowly enough for monthly planning to be adequate. That world no longer exists for most consumer goods categories.

The Demand Sensing Vision

In the mid-2010s, P&G's supply chain leadership articulated a new vision: move from predict-and-plan to sense-and-respond. Instead of generating a monthly forecast and hoping it held, the company would build systems that continuously monitored demand signals and adjusted supply chain responses in near-real-time.

The shift required three capabilities:

1. Multi-Signal Demand Sensing

P&G began ingesting demand signals far beyond historical shipments:

• Point-of-sale (POS) data. Actual consumer purchases at the register, shared by retail partners (Walmart, Target, Kroger, and others). POS data is the closest available signal to true consumer demand — unlike shipment data, which reflects what retailers ordered, POS reflects what consumers bought.

• Retail inventory positions. How much P&G product was on retailer shelves and in retailer warehouses. Combined with POS data, this allowed P&G to estimate true demand even when shelves were partially stocked out (a consumer who would have bought Tide but found an empty shelf represents demand that POS data does not capture).

• Weather data. Temperature, precipitation, and severe weather forecasts at the zip-code level. Certain categories — laundry detergent, beverages, cold/flu remedies, sunscreen — have strong weather correlations.

• Web search and social media. Google search trends and social media mentions provided early signals of shifting consumer interest. A spike in searches for "cold remedy" in a particular region could predict a demand surge 1-2 weeks before it appeared in shipment data.

• Promotional execution data. Not just the planned promotional calendar, but real-time data on which promotions were actually in-market, at what price points, and with what level of retail compliance.

• Economic indicators. Consumer confidence indices, unemployment claims, and retail sales data — particularly useful for anticipating shifts in consumers' willingness to trade down from premium to value brands.

2. Machine Learning for Signal Fusion

Ingesting these signals was necessary but not sufficient. The challenge was combining them into a coherent demand estimate. P&G built machine learning models that:

• Weighted signals dynamically. POS data was the strongest signal for short-term demand (1-2 weeks out). Historical patterns were stronger for long-term forecasting (3+ months). The models learned to weight signals differently depending on the forecast horizon.

• Detected anomalies. When a signal was inconsistent with other signals — for example, POS data showing a sudden spike but no corresponding promotion or weather event — the system flagged it for human review rather than automatically adjusting the forecast.

• Learned category-specific patterns. The relationship between weather and demand varies enormously by category. A temperature drop drives cold medicine purchases within days but affects laundry detergent purchases with a one-week lag. The ML models learned these category-specific dynamics from data rather than requiring manual specification.

3. Automated Response Protocols

The most radical change was connecting demand sensing to automated supply chain responses:

• Short-term (0-2 weeks): If sensed demand exceeded the current plan by more than a specified threshold, the system automatically triggered expedited shipments from the nearest distribution center. No human approval required for adjustments within predefined guardrails.

• Medium-term (2-8 weeks): Significant demand shifts generated revised production schedules, subject to planner review and approval.

• Long-term (2+ months): Sustained demand signal changes triggered revision of the baseline forecast and upstream raw material procurement adjustments.

Implementation: The Hard Parts

P&G's demand sensing transformation took approximately four years from initial pilot to broad deployment. The technical challenges, while significant, were not the hardest part. Three organizational challenges proved more difficult:

Data Sharing with Retail Partners

POS data is extraordinarily valuable — and retailers know it. Persuading Walmart, Target, Kroger, and dozens of other retailers to share granular, timely POS data required multi-year relationship building and clear demonstrations that better P&G demand sensing would reduce stockouts on their shelves (benefiting the retailer) as well as reduce P&G's inventory (benefiting P&G).

The data-sharing agreements included strict governance provisions: how the data could be used, who could access it, how long it could be retained, and explicit prohibitions against using one retailer's data to benefit another retailer. These negotiations involved legal teams, IT teams, and senior commercial leaders on both sides.

Changing Planner Behavior

P&G's demand planners — hundreds of experienced professionals worldwide — had built their careers on a specific way of working: review historical data, apply judgment, produce a monthly plan. Demand sensing asked them to trust a machine-learning system that updated continuously and that they could not fully understand.

Resistance was predictable and significant. P&G addressed it through several approaches:

• Transparency. The demand sensing system was designed to show why it was adjusting the forecast — which signal drove the change, by how much, and with what confidence. Planners could see the logic, not just the output.

• Guardrails. Automated adjustments were limited to a predefined range. Extreme shifts still required human approval. This gave planners veto power while reducing the routine manual burden.

• Gradual transition. Demand sensing was initially positioned as an "enhancement" to the existing process, not a replacement. Planners could use it as a supplementary input. Over time, as they saw it improve accuracy, many voluntarily shifted to it as their primary tool.

• Metrics that mattered. P&G tracked not just forecast accuracy but also bias (systematic over- or under-prediction) and value-added (the percentage of planner adjustments that actually improved the statistical forecast). The data showed that planner adjustments improved the forecast only about 50 percent of the time — meaning planners were adding noise as often as they were adding value. This was a humbling but motivating finding.

IT Infrastructure

Processing POS data from thousands of retail locations, weather data at the zip-code level, and web search trends at daily frequency required a data infrastructure that P&G's existing systems could not support. The company invested in cloud-based data platforms, streaming data architectures, and modern ML deployment infrastructure — a multi-year, multi-hundred-million-dollar investment that extended well beyond the demand sensing initiative.

Results and Impact

By the early 2020s (pre-pandemic), P&G's demand sensing system was producing measurable results:

• Forecast accuracy improvement. Short-term forecast error (1-2 weeks out) improved by 30-40 percent compared to the traditional monthly planning process.

• Inventory reduction. Finished goods inventory, measured in days of supply, decreased by approximately 15 percent across categories where demand sensing was deployed.

• Service level improvement. On-shelf availability at major retail partners improved by 1-2 percentage points — a seemingly small number that, at P&G's scale, represented hundreds of millions of dollars in prevented lost sales.

• Response time. The time from demand signal detection to supply chain response decreased from weeks (under the monthly planning process) to days for short-term adjustments.

• Planner productivity. Experienced planners spent less time on routine forecast adjustments and more time on high-value activities: new product launches, strategic promotions, and supply disruption management.

Lessons for Chapter 16

P&G's demand sensing journey illustrates several principles from this chapter:

1. Forecasting is a means to an end. P&G did not pursue better forecasts for their own sake. The goal was better supply chain outcomes: lower inventory, fewer stockouts, faster response. Forecast accuracy was a leading indicator of supply chain performance, not the final objective.

2. External regressors are transformative. Moving from a model that used only historical shipments to one that incorporated POS data, weather, and search trends produced dramatic accuracy improvements. The regressors must be causally plausible, empirically validated, and available in near-real-time.

3. Uncertainty and automation coexist. P&G's system does not produce a single forecast. It produces a range with confidence levels, and the automated response protocols are calibrated to the confidence level. High-confidence signals trigger automatic responses; low-confidence signals trigger human review.

4. Organizational change is the bottleneck. The technical system — data ingestion, ML models, automated responses — took two years to build. Changing planner behavior, building retail partner trust, and upgrading IT infrastructure took four years. The organizational transformation was twice as long as the technical development.

5. The "model" is the smallest part. The ML models at the core of demand sensing are important but represent a small fraction of the total system. Data pipelines, data quality monitoring, anomaly detection, planner interfaces, automated response protocols, governance frameworks, and performance monitoring constitute the majority of the work. This echoes Ravi Mehta's observation in the chapter: "The model is maybe 20 percent of the work."

Discussion Questions

1. P&G's demand sensing system relies heavily on POS data shared by retail partners. What happens if a major retail partner (say, Walmart) decides to restrict or monetize access to this data? How should P&G mitigate this risk?

2. The chapter emphasizes that forecasts should communicate uncertainty. How does P&G's automated response protocol — which triggers supply chain adjustments based on confidence levels — implement this principle?

3. P&G found that demand planner adjustments improved the statistical forecast only about 50 percent of the time. Does this mean human judgment is unnecessary in forecasting? What types of situations might still benefit from planner intervention?

4. How might P&G's demand sensing approach differ for a new product launch (with no historical demand data) versus an established brand like Tide?

5. P&G's transformation took four years. A startup building a similar capability today would have access to cloud-native data platforms, pre-built ML tools, and modern forecasting libraries like Prophet. How much faster could a startup build a comparable system? What advantages does P&G's approach have that cannot be replicated quickly?

This case study is based on publicly available information from P&G investor presentations, supply chain industry publications, academic research partnerships, and industry conference presentations. Specific financial figures are estimates based on publicly reported metrics and industry analysis.