Macy's Moves AI Forecasting Out of Pilot, but the Numbers Come Later

On 10 September 2026, Macy's takes AI forecasting out of pilot

On 10 September 2026, on Macy's Inc.'s earnings call, COO and CFO Tom Edwards described a precise operational move: an AI forecasting overlay applied to replenishment, transitioning from pilot to a broader rollout.

The stated objective covers two items: on-shelf availability levels and inventory efficiency. The capability sits within a wider initiative, tied to the principle of the «right product in the right place at the right time», as Edwards put it.

The account comes from Retail Dive^[1], which published its report on 18 September 2026. The deployment announcement comes with a date, a scope and a named executive. The metric capable of measuring the result, as things stand, has not been published.

The original idea: an AI layer on top of a plan already in motion

Since 2024, Macy's has been working on a three-pillar transformation plan called «Bold New Chapter», with the supply chain at its centre of gravity.

The plan involves closing fulfilment centres judged to be underperforming and opening an automated facility in North Carolina. Stated savings expectations amount to 235 million dollars by 2026, according to the Retail Dive report.

The design choice deserves attention. The company grafted forecasting on top of a logistics network already being redrawn, rather than treating it as a standalone project. The pattern recurs in successful enterprise deployments: the advantage stems from architectural decisions taken years earlier. The technology arrives afterwards, onto prepared ground.

The numbers that are there and the ones still missing

The call offers one inventory figure: in the second quarter, stock levels grew 2.5%, in line with sales growth.

It is a measure of operational balance. But the figure describes the entire commercial operation. Nothing in the public record links it to the forecasting system that has just been extended.

The economic benefit is explicitly deferred. «In total we expect to realise supply chain efficiencies in the second half of 2026, with a positive effect on gross margin», Edwards told analysts. The sentence covers the whole plan: closures, automation, replenishment, forecasting. Isolating AI's contribution, as of today, remains impossible for an outside reader.

The distinction matters. A result announced on a call is a statement from management. A verified result is a number a third party can recalculate. The Macy's case belongs to the first category, with a named executive and a precise date, which already puts it ahead of the average AI announcement.

The point of friction: scale declared before measurement

Every useful story contains a point of friction. In this case the friction lies in the sequence.

The scale threshold is communicated to the market before the number that justifies it. The pilot closes, the rollout starts, and the public proof will arrive later, inside an aggregated gross margin line.

This sequence carries an informational cost. It makes it hard to distinguish a technical success from a programme decision. It also shifts verification to a horizon where other levers will be acting at the same time. The evaluation criterion stands firm: ROI is measured with a before/after on a metric that has a denominator. «Significant improvement» remains communications until a comparison between two periods backs it up.

Why replenishment is the right ground to try it on

Replenishment is an almost pure forecasting problem: how much stock, in which store, in which week.

An overlay works on top of the existing system and corrects the demand estimate with signals that the classic statistical model struggles to handle: weather, local promotions, digital traffic, returns. The expected gain shows up in availability levels and in merchandise sitting idle in the warehouse.

The chosen scope also reduces exposure to the customer. The warehouse stays far from the end customer. A model error therefore translates into an internal cost, not a compromised shopping experience. An analysis by MIT Sloan Management Review^[2] on customer resistance to AI is a reminder of how much acceptance weighs when the technology touches the moment of purchase. Starting from the back office reduces that risk surface.

Dirty data beats a brilliant model

An overlay inherits the quality of the underlying data: master records, prices, store mappings, clean sales histories.

Jim Webber of Neo4j, in an interview published by diginomica^[3], ties data accuracy directly to the quality of AI systems' answers, and observes that cutting budgets does little to solve the cost problem. The reasoning holds well beyond the world of knowledge graphs.

Applied to Macy's, the principle points to what to watch in the coming quarters.

• In-stock levels by category, before and after the rollout
• Stockouts in the stores covered by the extension
• Markdowns applied to slow-moving merchandise
• Days of cover per store

A gross margin improvement, taken in isolation, leaves open the question of where the result came from. The careful reader looks for the disaggregated line, and asks for the comparison with a prior period.

What you can take away

The reading changes depending on the reader's role.

For an SME founder, the playbook remains replicable at modest cost. The method calls for selecting an internal process that is measurable and far from the customer. Onto that process you graft a forecasting layer, without rewriting the core system.

For a CTO, the message is about architecture. An overlay coexists with the existing ERP, so rollback risk stays low and the test cycle shortens. For a board, the bar moves elsewhere: a retailer with billions in revenue declares the scale of the deployment before the proof, and the market receives the move as a signal of progress.

For a department manager, the translation is immediate. The forecast handled by hand on a spreadsheet today takes in two external signals, the ones the department knows better than anyone else in the company. The comparison of results runs over eight weeks.

The open question

One question remains valid for any organisation, from the global chain to the shop with three locations: which metric have you decided to publish before extending the technology to every department?

Whoever answers with a number and a date has a case study in hand. Whoever answers with an objective has a project in hand. The project is still a good start, provided the measure is set now.

The next verification date for Macy's is already fixed: the second-half 2026 results. Gross margin will provide an indication. The quality of the answer will depend on the level of disaggregation of the line items the company chooses to publish.