Replenishment Intelligence

Overview

Replenishment Intelligence: Overview

What is Replenishment Intelligence?

Replenishment Intelligence is the Retail AI module that answers one operational question for a Walmart marketplace seller:

"What should I reorder this week, for which stores, how much, and why?"

It is not a single model. It is a decision system that fuses three layers of intelligence into one ranked, ready-to-act list of purchase order recommendations:

• Demand Signals - the forward-looking forecast of what each item will sell over the next 26 weeks.
• Store Signals - the store-by-store health and risk picture, derived from anomaly detection (low weeks of supply, stockouts, overstock, sales drops, and more).
• Agent collaboration (Agent2Agent) - the orchestration that brings demand and store signals together automatically and explains the result.

The output is a set of PO recommendations and the decision trace behind every number, so the recommendation is defensible all the way back to source data.

Why it matters for business

A Walmart replenishment manager loses money in two directions at once:

• Stockouts - shelves empty, sales are lost, and item performance scorecards drop.
• Overstock - working capital sits frozen in inventory that is not turning.

Most teams manage this with spreadsheets, manual weeks-of-supply math, and gut feel across thousands of item-store combinations. Replenishment Intelligence replaces that manual stitching with a system that:

• looks at every item and every store at once,
• weighs future demand against current store-level risk,
• and produces a prioritized reorder plan with the reasoning attached.

Core business capabilities

• One ranked action list - thousands of item-store combinations reduced to the products that actually need a PO this week.
• Risk-aware reordering - reorder quantities reflect not just the forecast, but store-level stockout and overstock risk.
• Distribution-center aware - recommendations roll up from store level to DC level, with full-truck-load efficiency in view.
• Fully explainable - every recommended quantity carries a decision trace from source signal to suggested units.
• Conversational - ask in plain language and get a defensible plan back.

Where it fits in the product experience

Replenishment Intelligence lives under Retail AI as the Replenishment Intelligence Center. It sits downstream of two foundational capabilities documented separately:

• Forecasting feeds the Demand Signals layer.
• Anomaly Detection feeds the Store Signals layer.

Replenishment Intelligence consumes both and turns them into action.

Typical stakeholders

• Replenishment and inventory managers
• Supply chain and demand planners
• Category and account managers serving Walmart
• Leadership reviewing service levels and working capital

Typical decisions it supports

• Which products to reorder this week, and in what quantity
• Which stores are at the highest risk of stocking out
• Where inventory is overstocked and capital can be released
• How to consolidate orders into efficient truckloads
• How to defend a reorder decision to leadership or to the retailer

Quick value summary

Replenishment Intelligence turns demand forecasts and store-level risk into a single, prioritized, fully explainable reorder plan - so a seller protects sales, frees working capital, and can defend every number.

Demand Signals

Replenishment Intelligence: Demand Signals

What are Demand Signals?

Demand Signals are the forward-looking layer of Replenishment Intelligence. They answer:

"How much of this item will sell, where, over the coming weeks?"

Demand Signals come from the forecasting engine and become the baseline need that every reorder quantity starts from. Without a credible forward demand number, replenishment is just reacting to the past.

What the business user sees

• 26-week demand forecast - projected units for each product over the next 26 weeks, the planning horizon used for ordering decisions.
• Projected revenue - the forecasted dollar value of that demand, so prioritization is by business impact, not just unit count.
• Forecast accuracy - a measured quality score for the forecast, so the user knows how much to trust each number.
• High-confidence coverage - how many SKUs are forecast with high confidence versus those that need a closer look.
• New product introductions (NPI) - newly launched items that have little or no sales history and are handled with special care.

Demand segmentation

Not every product behaves the same way, so Demand Signals are segmented before they drive replenishment:

• Demand segments group products by how predictable and how fast-moving they are.
• Performance tiers separate high-volume drivers from the long tail.

This segmentation lets the business user filter the reorder plan down to the products that matter most to them - for example, high-volume, high-confidence items first.

How Demand Signals drive replenishment

1. The forecast establishes expected demand per item over the planning horizon.
2. That demand is converted into a baseline reorder need.
3. The need is then adjusted by Store Signals (see next tab) for store-level risk before a final PO quantity is produced.

In other words: Demand Signals say how much you will need; Store Signals say where reality is diverging from plan; together they produce what to actually order.

Why it matters for business

• Reorders are anchored to future demand, not last week's sales.
• Prioritization is by forecasted revenue, so attention goes where the money is.
• A visible accuracy score keeps the system honest and builds trust.
• NPI items are flagged so they are not under-ordered just because they lack history.

Related documentation

For the modeling detail behind these signals, see the Forecasting documentation under Demand AI. This tab focuses on how the forecast is used inside replenishment.

Store Signals

Replenishment Intelligence: Store Signals

What are Store Signals?

Store Signals are the risk layer of Replenishment Intelligence. They answer:

"Where is reality diverging from plan, store by store, right now?"

Demand Signals tell you what should sell. Store Signals tell you where the network is unhealthy - where an item is about to stock out, where inventory is piling up, where sales suddenly dropped or spiked. These come from store-level anomaly detection and sharpen every reorder decision.

The store-level signals we surface

Each item-store combination is checked against a set of supply-and-sales health rules. The business-facing signals are:

• Low Weeks of Supply - inventory is running thin relative to demand; high stockout risk.
• High Weeks of Supply - far more inventory on hand than demand justifies; overstock risk.
• Zero Sales with Inventory - stock is on hand but nothing is selling; a possible shelf, listing, or localized demand problem.
• Sales Drop - a meaningful, unexpected fall in sales versus the recent baseline.
• Sales Spike - an unusual surge in sales that can drain supply faster than planned.
• Inventory Jump - a sudden, unexpected change in on-hand inventory.
• Traited with Zero Inventory - the store is set up to carry the item but currently has none; a direct lost-sales situation.

Each signal carries a severity so the most urgent situations rise to the top.

How the business user sees the impact

Store Signals are rolled up into business terms the user acts on:

• Revenue at risk - dollars exposed by stockout-type conditions.
• Inventory at risk - dollars exposed by overstock and slow-moving conditions.
• Stockout risks - count of item-store situations heading toward empty shelves.
• Overstock exposure - count of item-store situations carrying excess.
• Stores affected - how wide the issue spreads across the network.
• Critical alerts - the highest-severity situations that need attention first.

Weeks of Supply, in plain terms

Weeks of Supply (WOS) is how many weeks current inventory will last at the expected rate of sale. It is the core health metric:

• Too low - you will run out before the next delivery; reorder, and reorder more.
• Too high - capital is frozen on the shelf; slow down or stop ordering.

Replenishment Intelligence uses WOS on both sides - to push orders up where supply is short and to hold orders back where supply is long.

How Store Signals drive replenishment

The baseline need from Demand Signals is adjusted by Store Signals:

• Stores trending toward stockout get reorder quantities pushed up.
• Stores carrying overstock get reorder quantities pulled down or suppressed.
• Zero-inventory-but-traited situations are surfaced as immediate lost-sales recoveries.

Why it matters for business

• Protects sales by catching stockouts before the shelf is empty.
• Frees working capital by flagging overstock instead of reordering into it.
• Quantifies exposure in dollars and store counts, not abstract scores.
• Focuses scarce attention on the critical few among thousands of combinations.

Related documentation

For the detection methodology and the full rule set, see the Anomaly Detection documentation in Retail AI. This tab focuses on how those anomalies become replenishment risk signals.

PO Recommendations

Replenishment Intelligence: PO Recommendations

What is a PO Recommendation?

A Purchase Order (PO) Recommendation is the actionable output of the module - the answer to "order this much of this product." It is where Demand Signals and Store Signals converge into a single, ready-to-act number.

Each recommendation tells the business user:

• which product needs ordering,
• how many units to order,
• where it is needed (store and distribution center), and
• how urgent it is.

What the business user sees

• Products needing a PO - the count of products that require action this week, filtered out of the full catalog so attention goes only where it is needed.
• Total suggested PO units - the aggregate order quantity across the recommendation set.
• Must-replenish list - the prioritized subset that cannot wait, ranked by urgency and business impact.
• Per-store recommendations - order quantities resolved down to the individual store level.
• DC cascade - store-level needs rolled up to the distribution centers that serve them.

How the recommended quantity is built

A recommendation is not a raw forecast number. It is constructed in stages:

1. Baseline need from the 26-week Demand Signal.
2. Net of position - subtract what is already on hand and in transit.
3. Risk adjustment from Store Signals - push up where stockout risk is high, pull down where overstock exists.
4. Prioritization - rank by urgency and forecasted revenue so the must-replenish items surface first.

The result is a quantity that reflects both what will sell and what is actually happening in stores.

Distribution-center awareness and truck efficiency

Replenishment does not stop at the store. Store-level needs cascade up to distribution centers, and the system looks at order consolidation:

• DC rollup - aggregate the stores served by each DC into a single replenishment picture.
• Full-truck-load (FTL) efficiency - highlight where consolidated orders align to efficient truckloads, reducing freight cost per unit.

This means a recommendation is not just "right for the shelf" but "efficient for the supply chain."

What the user can do with it

• Review the ranked PO list and the suggested units.
• Drill into any product to see its per-store breakdown and the reasoning behind the quantity (see Agent2Agent).
• Filter by segment, tier, severity, and channel to focus the plan.
• Export the recommendations to act on them in the ordering workflow.

Why it matters for business

• Converts analysis into a concrete order quantity, not just a chart.
• Reflects current inventory position, so the system does not over-order.
• Is DC- and freight-aware, protecting margin as well as service level.
• Surfaces the must-replenish few so nothing critical is missed.

Related documentation

See Agent2Agent for how each recommendation is produced on demand from a plain-language request and how its reasoning is made transparent.

Agent2Agent

Replenishment Intelligence: Agent2Agent

What is Agent2Agent?

Agent2Agent (A2A) is how the specialist agents behind Replenishment Intelligence work together so the business user does not have to stitch the analysis by hand.

Behind the scenes, replenishment is not one model - it is a team of agents:

• a Forecasting agent that produces Demand Signals,
• an Anomaly agent that produces Store Signals,
• a PO agent that turns those into recommendations, and
• a Report agent that summarizes the outcome.

Agent2Agent is the collaboration layer that lets these agents hand work to one another and assemble a single, coherent answer.

The reorder loop

The core A2A flow is the reorder loop. A business user asks, in plain language:

"What should I reorder?"

and the agents collaborate to answer:

1. Forecast - the Forecasting agent establishes demand over the planning horizon.
2. Anomaly - the Anomaly agent overlays store-level risk on that demand.
3. PO - the PO agent reconciles demand, inventory position, and risk into recommended quantities.
4. Report - the result is summarized with its decision traces, ready to review and act on.

The user gets a defensible plan back without manually running forecasts, pulling anomaly reports, and reconciling spreadsheets.

Self-awakening on fresh data

A2A also means the system reacts to new data rather than waiting to be asked. When the underlying sales and inventory data refreshes:

• the relevant signals are recomputed, and
• the replenishment view reflects the latest picture.

The user opens the module to a current plan instead of stale numbers.

The Decision Command Center

As agents collaborate, their activity is surfaced rather than hidden. The user can see the chain of work - which signal was produced, which risk was found, which recommendation resulted - so the collaboration is transparent, not magic.

This same transparency produces a decision trace on every recommendation: the demand input, the inventory position, the store-level risks that adjusted the quantity, and the rule path that led to the final suggested units. Because the agents pass structured signals to one another, the reasoning is preserved end to end - so the user can always answer "why this number?" and defend it to leadership or the retailer.

How the business user experiences it

• Ask in plain language - a conversational panel inside the Replenishment Intelligence Center answers reorder and risk questions against the live data.
• One coherent answer - demand, risk, and recommended action arrive together, already reconciled.
• Always defensible - because the agents pass structured signals to one another, the reasoning is preserved end to end and shown as a decision trace.

Why it matters for business

• No manual stitching - one question replaces a multi-tool, multi-spreadsheet workflow.
• Faster decisions - the loop from "what should I reorder?" to a defensible plan is minutes, not days.
• Consistency - every user gets the same rigorous flow, every time, with the reasoning attached.

Related documentation

See PO Recommendations for how the recommended quantity is built, and Overview for how the three signal layers fit together.