Demand volatility has become a structural feature of modern supply chains, not a temporary disruption to be managed. Weather swings, viral product trends, geopolitical shocks, and sudden shifts in consumer behaviour can reshape demand patterns in days rather than quarters. Traditional forecasting systems, built primarily on historical sales data and updated on weekly or monthly cycles, were not designed for this pace. By the time a conventional model detects a shift, the stockout has already happened or the overstock is already accumulating.
Artificial intelligence is changing that dynamic. By combining machine learning with generative AI, companies can now integrate real-time signals from weather feeds, social media, IoT sensors, and point-of-sale systems and simulate multiple demand scenarios before disruptions materialize. It is a fundamental change in what forecasting means: from predicting a number to understanding demand dynamics as they unfold.
In case you want to jump ahead
The Methologies Behind the Shift
To understand what AI is changing in demand planning, it helps to understand what the new methods actually do and how they differ from the statistical models that preceded them. Four approaches currently define the state of the art.
MACHINE LEARNING FORECASTING PLATFORMS
Classical statistical models like ARIMA, exponential smoothing, and linear regression forecast demand by extrapolating past patterns, making them effective in stable environments but less reliable when conditions change rapidly. Machine learning models improve on this by analyzing much larger datasets and incorporating multiple variables simultaneously, allowing them to identify complex demand patterns.
TRANSFORMER BASED MODEL & PROBABILISTIC FORECASTING
Transformer-based models like the Temporal Fusion Transformer, iTransformer, and Informer use self-attention mechanisms to identify the most relevant historical patterns and external variables for predicting demand. Their main advantage is probabilistic forecasting: instead of producing a single demand estimate, these models generate multiple demand scenarios with associated probabilities.
GENERATIVE AI FOR SCENARIO SIMULATIONS
Generative AI simulate plausible future scenarios that haven’t occurred before, like tariff changes or viral demand spikes. This is increasingly applied through digital twin simulations, where virtual supply chain models are tested against thousands of generated scenarios to evaluate risks and operational responses. It can also solve the challenge of forecasting new products, by using GANs to generate synthetic demand patterns.
AGENTIC AI FOR AUTONOMOUS ACTION
Agentic AI enables systems to continuously monitor supply chain conditions, detect disruptions or demand changes, evaluate response options, and execute actions with minimal human intervention. These systems already manage routine replenishment decisions, safety stock adjustments, and supplier communications, while escalating only higher-risk or strategically significant decisions to human planners.
Ice Cream and the Case for Real-Time Forecasting
Few products expose the limits of traditional demand forecasting as clearly as ice cream. And few companies have gone further in rethinking how demand planning works at scale.
Demand for ice cream is simultaneously predictable in aggregate and deeply volatile at the local level. Seasonal patterns are well understood. But a sudden heatwave can trigger a regional surge within 48 hours, while a rainy week suppresses sales just as rapidly. For a product sold almost solely through impulse purchases at the point of a freezer cabinet, with a cold chain that cannot be improvised overnight forecasting errors translate directly into lost revenue, excess inventory, and inefficient logistics.
This is the structural challenge faced by Unilever, whose ice cream supply chain spans 60 countries, 35 factories, and roughly 3 million freezer cabinets worldwide. At that scale, even a 5% improvement in forecast accuracy is worth hundreds of millions of dollars annually. And the cost of a stockout on a hot afternoon when a customer reaches into a cabinet and finds it empty is not just a lost sale. It is a brand interaction that fails at the most critical moment of the purchase journey.
From Historical Extrapolation to Real-Time Demand Sensing
Traditional forecasting at Unilever, as at most large CPG companies, relied primarily on historical sales patterns updated on a periodic cycle. For weather-sensitive products, that approach has an inherent flaw: the model learns from past summers, but it cannot know that this Tuesday afternoon will be the hottest day in three years and that demand will spike accordingly.
The intervention was to add real-time signal integration. Unilever’s forecasting systems now ingest weather data, retail insights, and operational inputs. Rather than producing a single demand projection updated weekly, the system continuously models multiple demand scenarios, weighting them by probability as signals evolve.
AI-Powered Demand Intelligence
The Freezer Cabinet as a Demand Sensor
The most operationally significant change has happened at the retail edge. More than 100,000 of Unilever’s freezer cabinets are now equipped with AI-powered image recognition systems that monitor stock levels in real time. When a camera detects that a cabinet is running low, the system generates a replenishment recommendation automatically, not based on a projected depletion curve calculated last week, but based on what is actually in the cabinet right now, cross-referenced with the current weather forecast and local sales velocity.
The practical consequence is that roughly 5% of seasonal orders are now generated directly from these AI-driven recommendations rather than from traditional planning cycles. That percentage will grow as coverage expands. For operations teams, this means fewer manual ordering decisions, more efficient delivery routes, and a fundamentally different relationship between demand signal and supply response.
Retail Sales Growth
8%
TURKEY
Retail Sales Growth
12%
USA
Retail Sales Growth
30%
DENMARK
Why Ice Cream Case Matters Beyond Ice Cream
The ice cream case is valuable precisely because it is not a special case. The same pattern applies to beverages, fresh food, seasonal apparel, and any product where the gap between demand signal and supply response carries a real cost.
What Unilever demonstrated is that closing that gap requires three things working together: a real-time data layer (weather, POS, cabinet sensors), a forecasting model capable of processing those signals faster than the planning cycle, and an execution system that can act on the output without waiting for human review. None of those three elements alone is sufficient. The competitive advantage comes from their integration.
The freezer cabinet is not just a storage unit. For Unilever, it has become the primary demand sensor, generating signals that the planning system acts on within hours, not weeks. That is a different model of forecasting.
Additional Industry Cases
The ice cream model is not an isolated experiment. The same underlying shift is playing out across industries, each with its own version of the volatility problem.
Click on the use case you want to learn more about:
Fast Fashion: When a Trend Goes Viral Faster Than Your Factory Lead Time
H & MSocial Listening & ML Forecasting
Fast fashion’s structural tension is a time mismatch. Consumer trends now move at social media speed. A colour, silhouette, or style can go from niche to mainstream on TikTok within a week. Traditional production planning, built on months-long design-to-shelf cycles, has no mechanism to respond at that speed.
H&M pursued a parallel approach. Facing persistent overstock driven by bulk seasonal planning, the company built ML models trained on over a decade of historical data, augmented with real-time signals from social netowrks. Natural language processing detects emerging patterns, specific color preferences, style surges, with reported trend detection accuracy of 95%. Models forecast at item level per store, including size and color attributes. The outcome: a 30% reduction in lead times through AI-assisted scenario simulation, and a meaningful reduction in overstock that had previously cost the business millions annually.
95%
Trend detection accuracy via NLP on social signals
30%
Reduction in lead times with AI-assisted scenario simulation
↓M$
Overstock reduction, saving millions annually
Pharmaceuticals: When a Stockout Is Not a Lost Sale, It Is a Patient Without Medication
Amazon PharmacyAWS Supply Chain ML
Pharmaceutical demand forecasting carries stakes that retail cannot match. A stockout for a prescription medication means a patient without a drug they need. At the same time, many medications have strict shelf lives and cold-chain storage requirements, making overstock equally costly.
Before its AI overhaul, Amazon Pharmacy managed demand through siloed, team-by-team processes: different functions running separate forecasting workflows, with longer planning cycle times and limited data granularity. After deploying AWS Supply Chain ML-based demand planning, the system generates daily-level forecasts, refreshing automatically with the latest data and publishing outputs directly to downstream and upstream processes. Two forecast horizons drive different decisions: T-1 (one week out) guides immediate staffing to meet incoming prescription volume; T-5 (five weeks out) informs capacity planning for labor adjustments. The ML system detects correlations and seasonality patterns across prescriptions that were previously too complex to identify manually, including regional disease prevalence cycles and promotional event effects.
50%
Trend detection accuracy via NLP on social signals
~13%
Reduction in lead times with AI-assisted scenario simulation
77%
Pharma companies investing in AI demand sensing
Food & Beverage: A Sporting Event and a Replenishment Window Measured in Hours
Global Beverage LeaderEvent-Driven Demand Sensing
Food and beverage demand forecasting illustrates the consequences of the gap between demand signal and supply response more viscerally than most categories. When a major sporting event drives a regional surge in snack and beverage demand, an AI-based demand sensing platform can detect the spike through point-of-sale data as it begins, and reroute shipments autonomously, increasing supply to high-demand zones while reducing allocations elsewhere. Under a traditional weekly replenishment model, the stockout would have already happened before anyone noticed.
A leading global beverage company applied this principle systematically, deploying an AI forecasting system integrating weather forecasts, local event calendars, and social media activity alongside historical sales data. The outcome: a 75% reduction in out-of-stock incidents.
75%
Reduction in out-of-stock incidents
Hours
Replenishment response time vs. weekly traditional cycles
3 x
Data sources integrated: weather, events & social media
Organizational Readiness Check
Before committing to a technology investment, supply chain leaders should assess their organisation’s readiness across four dimensions. The table below provides a structured framework. Most large organizations entering AI demand planning in 2025–2026 sit at the Developing stage on data and at the Early Stage on governance, the latter being the more consequential gap to close.
| Dimension | Early Stage | Developing | Advanced |
|---|---|---|---|
| Data Infrastructure | Siloed ERP systems, limited external data feeds, manual data reconciliation | Integrated ERP with some external feeds; basic data lake in place | Real-time multi-source data lake; IoT streams; automated data quality monitoring |
| Model Capability | Statistical baselines only (ARIMA, exponential smoothing); no ML in production | ML forecasting deployed for major SKUs; limited scenario simulation | Transformer-based probabilistic models; generative AI scenario simulation; digital twin operational |
| Governance & Trust | No AI governance framework; ad-hoc planner overrides; no audit trail | Defined human-in-loop rules for key decisions; basic explainability in place | Full decision-rights taxonomy; explainability layers; autonomous execution within defined limits |
| Organisation & Skills | Planners use spreadsheets as primary tool; no data science capability | Small data science team; planners beginning to work with AI outputs | Cross-functional AI literacy; planners as orchestrators; dedicated AI centre of excellence |
Key Implementation Steps: Choose Your Perspective
The developments described above are not sequential innovations. They are converging simultaneously.
Click on the perspective you want to analyze – executives or analysts.
Key Implementation Steps (Executive View)
1. Define the Value Objective Before Selecting Technology
✔ Clarify whether the primary goal is inventory reduction, service level improvement, revenue growth through higher availability, or operational agility.
✔ Decide explicitly what level of AI autonomy you are targeting: monitoring and alerting only, scenario simulation with human decision, or autonomous execution within defined limits.
✔ Without this clarity, AI forecasting programmes produce dashboards rather than decisions.
2. Establish a Single Source of Demand Truth
✔ Ensure AI models have access to unified data across ERP, CRM, marketing campaigns, retail data, and external inputs such as weather and macroeconomic indicators.
✔ Break the silos between procurement, manufacturing, and logistics.
✔ Data integration consumes 42% of AI implementation timelines (Kumar et al., 2024). Budget for it accordingly.
3. Set Governance Guardrails Before Granting Autonomy
✔ Define which decisions can be automated (routine replenishment, safety stock adjustments) and which require human validation (major production changes, strategic inventory shifts).
✔ Given that 40%+ of agentic AI projects are expected to fail by 2027, establishing explicit decision rights before deployment is not a compliance exercise.
✔ Build an audit trail for every autonomous decision the system makes.
4. Align Organizational Incentives with the New Model
✔ Shift planning KPIs toward forecast accuracy, responsiveness to demand signals, and cross-functional collaboration.
✔ Planners become orchestrators of autonomous systems rather than spreadsheet operators. Role definitions and performance criteria must evolve accordingly.
✔ Resistance to AI adoption in planning teams is most frequently a governance problem, not a technology problem.
5. Start with a Focused Pilot in a Volatile Product Category
✔ Launch a 90-day proof-of-value in one product category or region with measurable demand volatility (high weather sensitivity, seasonal spikes, or short shelf life).
✔ Define 3–5 KPIs before launch: forecast accuracy (MAPE), stockout rate, inventory days, replenishment cycle time.
✔ Demonstrate measurable improvement in those KPIs before scaling architecture across sites. Unilever’s approach of piloting AI-enabled freezer cabinets in select markets before global rollout is a model worth replicating.
Key Implementation Steps (Analyst View)
1. Integrate Structured and Unstructured Data Sources
✔ Combine traditional sales history with weather feeds, promotional calendars, social sentiment signals, and IoT data streams from retail or warehouse sensors.
✔ Establish data quality standards before model training: ML models require large, clean datasets.
✔ For new products with limited history, consider synthetic data generation using generative AI (VAEs, GANs) to bootstrap training sets.
2. Benchmark Model Performance Rigorously Against Baselines
✔ Run back-testing against historical data to compare AI models with existing statistical forecasting baselines (ARIMA, exponential smoothing).
✔ Measure MAPE, weighted MAPE, and forecast bias by product category — not just aggregate accuracy figures that can mask SKU-level failures.
✔ Transformer-based models with explanatory variables achieve up to 12.4% NRMSE reduction vs. sales-history-only models.
3. Implement Natural Language Interfaces for Planners
✔ Enable planners to query forecasting systems directly in natural language: “Why did demand increase 15% in Region X last week?”
✔ Explainability is not a nice-to-have, it is the primary determinant of planner trust and adoption. A model that cannot explain its recommendations will be overridden.
✔ Multi-agent AI studies show that AI-assisted consensus planning across supply chain partners reaches agreement 80% faster than human-led cycles.
4. Build Explainability Layers into Every Model
✔ Ensure models generate interpretable insights by identifying the key drivers behind forecast changes.
✔ Require concise rationales for all algorithmic suggestions. Document counterfactuals: what would have happened under a different decision.
✔ For agentic systems, maintain a full audit log of every autonomous action taken, the signal that triggered it, and the threshold parameters that governed it.
5. Create Continuous Learning Loops
✔ Capture manual forecast overrides and feed them back into training pipelines so models learn from planner judgment over time.
✔ Monitor model drift: geopolitical shocks, tariff changes, and structural market shifts invalidate prior training data and require retraining triggers.
✔ For agentic AI: introduce autonomy gradually simulation mode first, shadow execution second, live autonomous decisions last.
Conclusion: Sense Faster, Respond Earlier
Traditional forecasting systems were designed to predict numbers. AI-driven systems are designed to understand demand dynamics as they unfold. The combination of machine learning, transformer-based probabilistic models, generative AI scenario simulation, and agentic autonomous execution is transforming demand forecasting from a periodic planning activity into a continuously learning, real-time system.
The Unilever ice cream case makes this concrete. A freezer cabinet that detects its own depletion, cross-references the local weather forecast, and generates a replenishment order without human intervention is not a futuristic scenario. It is already operating at scale across 100,000 retail locations.
For supply chain executives, the strategic implication is clear. Better forecasts improve inventory efficiency, reduce waste, and ensure product availability when demand spikes. But the deeper competitive advantage is structural: organisations that can sense demand faster and respond earlier than their competitors gain a durably superior position.
The organisations that gain structural advantage in the next five years will not be those that deployed AI first. They will be those that built the data foundations, governance frameworks, and organisational capabilities to use it reliably at scale.
In case you missed it – Previous Pulse Editions
A Fleet Executive’s Playbook for the Next Fuel Shock
Supply Chain Analytics PulseA Fleet Executive's Playbook for the Next Fuel Shock July 2026 | Supply Chain Analytics PulseIn late February 2026, the Strait of Hormuz, which carries roughly a fifth of the world's seaborne oil, effectively shut down as a sudden regional...
What Is Hyperautomation, and Why It’s Becoming Supply Chain’s Operating System
Supply Chain Analytics PulseWhat Is Hyperautomation, and Why It's Becoming Supply Chain's Operating SystemJuly 2026 | Supply Chain Analytics PulseBy 2034, the global hyperautomation market is projected to reach $235 billion, up from $55.54 billion in 2025 - a 17.38%...
Future of Warehousing: RFID, IoT, and Vision Picking
Supply Chain Analytics PulseFuture of Warehousing: RFID, IoT, and Vision PickingJuly 2026 | Supply Chain Analytics PulseThe warehousing industry is at the forefront of a technological transformation. Technologies such as radio-frequency identification (RFID), the...
Recent Comments