AI Forecasting Unifies Replenishment Across Banners

Neighborhood Intelligence, the parent company of Bed Bath & Beyond, is deploying AI to unify forecasting and replenishment across a network of more than 335 stores and multiple online brands in North America. The program consolidates demand signals that previously sat in separate planning systems, with the goal of producing one replenishment decision per SKU rather than one per banner. For multi-banner retailers, the interesting question is not whether the model is accurate but whether the organization can act on a single number.

Key Conclusions

The scale of the rollout is what makes the case instructive. Supply Chain Dive reports the parent company operates more than 335 stores (Supply Chain Dive) plus multiple online brands, and is using AI to unify forecasting and replenishment across them. Consolidating forecasting at that scale means inventory decisions move from banner-level ownership to a shared planning layer, which changes incentives as much as it changes algorithms.

Unified forecasting usually produces its first visible gain in exception handling. When every banner submits its own plan, planners spend their time reconciling differences rather than resolving genuine anomalies. A single demand signal lets the planning team triage exceptions by financial impact, which typically improves service levels before it improves forecast accuracy metrics.

What Unification Changes in Practice

From banner plans to one SKU view

A shared planning layer requires a common product hierarchy, a common calendar and a common definition of service level. Retailers often discover during implementation that their banners define fill rate differently, which makes consolidated reporting impossible until the definitions are reconciled. This data cleanup work is unglamorous but it determines whether the model can be trusted.

The second change is decision rights. If a single system produces one replenishment number per SKU, banner teams lose the ability to override it locally without justification. Organizations that leave override authority ambiguous tend to see the model's recommendations bypassed during peak periods, which erodes the data quality that feeds the next forecast cycle.

Evidence and Guardrails for AI Planning

Measuring outcomes, not model scores

Forecast accuracy is a useful diagnostic but a poor success metric on its own. The measures that matter are in-stock rate, inventory turns and markdown expense, reported by banner and category. A model that improves accuracy while inventory turns stall has not delivered value, and reporting both together prevents that ambiguity from persisting.

Guardrails matter equally. Set explicit bounds on how far a single planning cycle can move replenishment quantities, and require review when a recommendation breaches them. Bounds limit the damage from a bad data feed while preserving the efficiency gains from automation during normal operations.

Best Practices

Reconcile metric definitions before rollout

Before switching on a unified forecast, standardize fill rate, in-stock and lead-time definitions across banners and document them in one place. Reconciliation is the slowest part of the program and the easiest to skip under time pressure, yet it determines whether consolidated reporting can be trusted by the planners who use it daily.

Set explicit bounds on automated changes

Define maximum movement thresholds per planning cycle and require human review when a recommendation breaches them. Bounds protect the business from a corrupted feed while still allowing automation to deliver most of its efficiency benefit. Review outcomes should feed back into threshold tuning rather than prompting one-off manual overrides.

Common Mistakes

The most common mistake is treating unified forecasting as an IT project. The model depends on reconciled hierarchies, shared metric definitions and clear override rules, all of which are organizational decisions. Retailers that deploy the software first and resolve governance later typically see adoption stall in the first peak season.

A second mistake is measuring success by forecast accuracy alone. Accuracy can improve while in-stock rates and inventory turns stay flat, leaving the business no better off. Reporting commercial outcomes alongside model diagnostics keeps the program focused on value rather than on technical performance.

Summary

Unifying forecasting and replenishment across more than 335 stores is primarily an organizational program with a modeling component. Shared hierarchies, reconciled metric definitions and explicit override authority determine whether the single demand signal is actually used. Commercial metrics such as in-stock rate and inventory turns are the honest test of value.

The transferable lesson for multi-banner retailers is to sequence governance before automation. Reconciling definitions and setting movement bounds takes time, but it is what allows a consolidated forecast to survive its first peak season instead of being overridden store by store.