AI-driven supply-chain and inventory optimization

That works for stable, non-perishable goods. It breaks in FMCG for three reasons:

• Short shelf life. A buffer of yogurt is not free insurance. It is a spoilage risk with a clock on it.
• Promotions and seasonality. A weekend price promotion can triple demand for one SKU (stock keeping unit, a single product-size-flavor combination). Fixed buffers cannot flex fast enough.
• Network complexity. Product flows through multiple stages: factory to central warehouse to regional depots to stores. Each stage holds stock. Optimizing them one at a time wastes cash and freshness.

That last point has a name: multi-echelon inventory optimization (MEIO). An echelon is one stage in the network. MEIO optimizes stock across all echelons at once, instead of each depot hoarding its own buffer.

Multi-echelon thinking, in plain terms

Imagine each regional depot independently keeping five days of yogurt "just in case." Multiply that across twenty depots and you are holding a mountain of perishable stock, most of which will never be needed in any single location.

MEIO asks a smarter question: where in the network should each unit of buffer actually sit?

Often the answer is to hold less at the edges (stores, depots) and slightly more centrally, where one buffer can serve many locations. This is called risk pooling: variability averages out when you combine demand across regions. A central buffer of 1,000 units can cover the same service risk as 3,000 units spread thinly across depots, because the depots rarely all spike at once.

For fresh dairy, MEIO adds a hard constraint: freshness. A unit held centrally must still reachreachThe number of unique people exposed to your message in a given period. Unlike impressions, reach counts each person once, no matter how often they see it.View full definition → the shelf with enough remaining shelf life. So the model balances pooling benefits against transit time and remaining days of life.

The MIT Center for Transportation and Logistics publishes accessible research on these network trade-offs if you want to go deeper.

Where reinforcement learning comes in

Forecasting tells you what demand might be. It does not tell you what to *do*. Replenishment is a decision problem: how much to order, for each SKU, at each location, every day, given uncertain demand and perishability.

Reinforcement learning (RL) is a branch of AI where an agent learns a policy (a rule for choosing actions) by trial and error against a reward signal. Think of it as learning to play a game where the "score" is your business objective.

For dairy replenishment, the setup looks like this:

• State: current stock at each location, product age, day of week, upcoming promotions, weather, recent sales.
• Action: how many units to order or ship to each location.
• Reward: service level achieved (availability), minus spoilage cost, minus holding cost.

The RL agent runs millions of simulated days against historical and synthetic demand. It learns, for example, that shipping extra to a specific store before a hot weekend pays off, but only for products with more than four days of shelf life left.

Crucially, RL learns policies that classic rules cannot express. A fixed safety-stock formula says "keep X days." An RL policy says "keep X days, unless product age exceeds Y and it is midweek with no promotion, in which case ship less and let the central depot absorb the risk." That conditional logic is exactly what the freshness-versus-availability trade-off requires.

A simplified reward function

Here is the core idea of the reward, in pseudocode. This is illustrative, not production code.

The agent tunes its ordering policy to maximize the sum of these daily rewards over time. Change the penalties, and you change behavior: raise SPOILAGE_COST and the agent gets more conservative on fresh SKUs.

Making it real: data and guardrails

RL and MEIO are only as good as their inputs and constraints.

Data you need:

• Point-of-sale (POS) data by store and SKU, ideally daily.
• Current inventory and product age at every echelon.
• Promotional calendars and price changes.
• Lead times (how long stock takes to move between stages).

Guardrails that matter:

• Shelf-life constraints must be hard rules, not soft penalties. You never want a policy that ships expired product to hit a service target.
• Supplier and production limits. The agent cannot order more than the factory can make.
• Human override. Planners should be able to inspect and adjust recommendations, especially during unusual events. AI proposes; the supply chain team stays accountable.

A practical rollout usually starts in shadow mode: the AI generates recommendations, humans keep making the real decisions, and the two are compared for several weeks. Only after the AI consistently beats the incumbent method on spoilage, service, and stock does it take control, often one product category at a time.

What good looks like

For our dairy producer, a well-tuned system does not chase one metric. It reports on all three at once:

• Service level: share of demand met from stock (fewer out-of-stocks).
• Spoilage rate: share of units perished before sale.
• Working capital / inventory days: how much cash is tied up in stock.

The win is moving all three in the right direction together, or accepting a tiny, deliberate trade in one to gain a lot in another. Vendors sometimes advertise large improvements; treat headline figures as estimates and insist on your own shadow-mode test before believing them.

Common failure modes to avoid

Optimizing forecasts instead of decisions. A more accurate forecast that still feeds a rigid ordering rule leaves most of the value on the table. The decision policy is where the money is.

Ignoring the retailer's constraints. Many grocery chains reject deliveries below a minimum remaining shelf life (a "days-to-expiry" requirement). If your model does not encode the retailer's acceptance rules, its "optimal" shipments get bounced at the dock.

Training on a world that no longer exists. Demand patterns shift: new competitors, price inflation, changed consumer habits. RL policies need periodic retraining, and monitoring for drift (when live conditions diverge from training data).

Over-centralizing perishables. Risk pooling saves cash, but for very short-life products, holding stock too far from the shelf just relocates spoilage. The model must respect transit time as a first-class constraint.

Bringing it back to the dairy

Return to that producer with two complaints. Under an AI-driven system, the morning looks different. Overnight, the agent saw a heatwave forecast for one region and a promotion starting in another. It pushed fresher stock toward the promotion region, held slightly more at the central depot to cover heatwave uncertainty, and trimmed shipments of older stock to the chain that enforces strict shelf-life acceptance.

Spoilage and out-of-stocks stop being a tug-of-war. They become two terms in a single objective the system solves every night.

Key Takeaways

• Freshness, availability, and working capital are one problem, not three. FMCG supply chains fail when they optimize these separately. AI lets you balance them in a single objective.
• Multi-echelon optimization plus risk pooling reduces total perishable stock by holding buffers where they cover the most locations, subject to shelf-life and transit-time limits.
• Reinforcement learning optimizes decisions, not just forecasts. Its reward function directly trades stockout penalty against spoilage and holding cost, producing conditional policies fixed rules cannot express.
• Guardrails are non-negotiable: hard shelf-life constraints, production limits, and human override. Roll out in shadow mode and prove results on your own dataown dataData collected directly from your own customers and prospects through your own channels: your most reliable and privacy-compliant source.View full definition → before ceding control.
• Treat vendor improvement claims as estimates. Measure service level, spoilage rate, and inventory days together, and retrain models as demand patterns drift.