Santanu Boral is the associate director at Cognizant, driving SAP-led supply chain transformation, ERP modernization and AI innovation.
In healthcare and pharmaceutical supply chains, drug products become unusable after expiration dates are reached. Balancing demand volatility and replenishment uncertainty increases the complexity of inventory decisions. U.S. hospitals discard an estimated $3 billion in medications each year.
Life sciences organizations have too much inventory and still fail to fulfill customer demand. Traditional planning models often operate in silos, separating demand forecasting from safety stock decisions. As a result, organizations struggle to adjust inventory levels in response to real-time changes in customer demand.
Traditional inventory planning techniques are mostly reactive and fail to detect shelf-life alerts early enough to support proactive intervention, and this limitation is significant because customer-specific shelf-life requirements at different customer service levels vary by geography. A product may exist somewhere in the network yet have a very short shelf life or regulatory condition to fulfill a customer need.
The core problem is not that traditional machine learning forecasting has become obsolete; rather, it’s that it’s frequently disconnected from the decisions it’s supposed to improve. So I want to present a framework that puts the implementation argument for a stronger operating model that connects prediction, optimization, shelf-life risk and execution into one governed decision loop.
How The Framework Works
This implementable framework has four connected layers, each with its own defined capability, outcomes and owners:
Layer 1 – The Data Foundation: This is the input layer, and it consists of master data (e.g., location, product, demand signals, lead time shelf-life attributes). It’s owned by the data and planning leads.
Layer 2 - Prediction And Risk Sensing: Algorithms generate a forecast accuracy model, uncertainty signals and expiry-risk scores, which demand and supply planners use to inform planning decisions.
Layer 3 – Risk Optimization: This layer recommends a target inventory position (TIP) for each product in each geographic market, accounting for service-level requirements, shelf-life constraints and applicable local regulations. Inventory and network planners are responsible for these recommendations across the network.
Layer 4 – Execution Governance: This is the approval layer where users prioritize exception rules. It is owned by the operation, quality and regulatory teams.
Implementation Approach For The Framework
This framework provides early-warning alerts that identify potential shelf-life waste several planning periods in advance, giving planners time to take corrective action. It integrates shelf-life constraints, projected inventory levels and financial parameters into a unified planning process, using cloud infrastructure and AI to support decision-making.
The cloud-based component is implemented through enterprise planning capabilities like time-phased inventory projections, attribute-based logic, real-time analytics and custom alert generation. The shelf life is typically defined at the product location level, and projected inventory is monitored against minimum remaining shelf-life requirements based on customer, market or regulatory environment.
The AI component is used to create early warning signals. Projected inventory is analyzed using relevant shelf-life attributes and forward-looking shortage logic. A business-defined horizon can be used to identify inventory that is likely to become waste before actual expiration occurs. Financial exposure can also be estimated by linking projected expiring quantity to product valuation and quantifying the potential economic impact of expiry-related waste.
We adopted a five-step approach to implement this framework in the life science industry:
1. Start With One Decision And Build A Decision Model: Begin with a focused use case, such as managing short-dated inventory at two distribution centers or addressing safety-stock exceptions for a critical product family. Once validated, expand the model to additional product families within the market.
2. Expand The Decision Model: Extend the model to a product-location dataset that reconciles demand forecasts with available inventory, accounts for batch-level expiry dates and lead times, verifies product quality and release status, and applies market-specific shelf-life requirements.
3. Account For Forecast Uncertainty: Replace a single-point forecast with a central estimate, uncertainty bands and risk triggers. Use scenario-based planning to identify potential demand that exceeds the upper planning threshold and generate warnings when inventory or supply may be insufficient.
4. Embed Constraints In The Optimization Logic: Incorporate forecast uncertainty, shelf-life limits, product release status, lead-time variability and service-level requirements into inventory optimization to ensure that recommendations are operationally feasible and commercially viable.
5. Close The Loop Through Governance: Route material exceptions to planners with recommended actions, and track the resulting decisions and outcomes. Record whether inventory was redeployed, demand was reprioritized, supply was expedited or an alert was rejected, including the reason for rejection.
The framework supports dashboard-based visibility and alert-driven monitoring to notify planners when projected shelf-life exposure crosses defined thresholds.
The following pilot timeline is a practitioner recommendation. During weeks 1–4, establish baseline performance. Over weeks 4–8, develop an enterprise-wide risk-sensing and recommendation model. By weeks 8–12, deploy the model in production with human oversight and a robust governance framework.
A pilot should establish baseline performance before implementation and measure subsequent changes against consistent product, location and time-horizon parameters. Improvements in model accuracy alone do not demonstrate enterprise-level impact; the pilot must also measure business outcomes.
Why AI-Driven Implementations Commonly Fail
AI-driven inventory optimization initiatives commonly fail when organizations start with a broad AI platform rather than a specific inventory decision they want to improve. They may also rely on forecast accuracy as the sole measure of success, overlooking service levels and financial performance. Another common mistake is treating shelf-life limitations, product quality requirements and regulatory rules as downstream checks rather than constraints that must be incorporated into the optimization process itself.
Implementation can also falter when organizations automate decisions before planners have validated the data and established appropriate controls and governance processes. Generating hundreds of alerts without prioritization or a clear decision framework further undermines the system’s effectiveness, overwhelming planners rather than helping them focus on the most consequential risks.
Conclusion
The future of inventory optimization will not be determined by the most sophisticated AI driven forecasting model. It will be determined by which organization can connect uncertainty signals to governed action and do the risk-pooling through the network.
In pharmaceutical supply chain planning, AI is moving beyond isolated pilots and becoming embedded in core processes such as demand forecasting, supply planning and production scheduling.
The model enhances organizations’ digital planning capabilities and minimizes the shelf-life risk, making the supply chain resilient and sustainable.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?