Every retailer now has an AI strategy

Walk into almost any large retailer today and you will find the same paradox. The data lake is bigger than it has ever been. There is a forecasting model, a price optimizer, an assortment tool, a workforce planner, a last-mile router — and now, a generative AI pilot or three. Most of them work, in the technical sense. And yet the decisions that actually move the P&L — what to buy, where to send it, how to staff the store, when to promote, what to do when the customer is not home — still feel reactive, still depend on the same five people, and still produce surprises every Monday morning.

The instinct is to blame AI maturity. More data, better models, more agents. We think the diagnosis is wrong. After working with retailers across grocery, fashion, electronics, DIY and quick commerce, we see a different pattern: the AI works. It is just pointed at the wrong question, in the wrong way, for the decision at hand.
‍

Most retailers don’t have an AI problem. They have an AI-fit problem.

The conversation around AI in retail has collapsed into two extremes: generative pilots that produce demos, and predictive models that produce dashboards. Neither directly moves the operational decisions that determine cost-to-serve, availability and customer experience.

Analytical AI is the third category, and the one that matters most for operations. It is the disciplined use of data, models and learning systems to make recurring operational decisions better — procurement, replenishment, allocation, fulfillment, last-mile, returns. It includes machine learning, but also stochastic modeling, optimization, causal inference and the increasingly important glue between them.

The question retail executives should be asking is rarely “do we need an AI model?” It is closer to: “what role should data play next to the decision we are trying to make?”

Drawing on a recent framework we developed, we find that Analytical AI initiatives in retail fall into four very different modes — each with a different definition of success, a different cost profile, and a different reason to fail.

Confusing the four modes is the single most common reason retail AI programs underdeliver. It is also the most fixable.

‍

The four modes of Analytical AI in retail
‍

Whether the data sits outside the decision model (shaping orevaluating it) or inside the decision model (feeding it), and whether you are designing the model or enabling its use — these two axes produce four distinct modes. The names are technical; the implications are commercial.

‍

These modes are not a maturity ladder. A retailer running sophisticated input-augmenting personalization can still be flying blind on whether its replenishment policy is structurally right. A retailer with a beautifully calibrated parametric routing engine can have no idea whether it actually moved cost-to-serve. The modes are complements, not stages — and a serious AI strategy in retail allocates capital across all four, not just the most fashionable one.
‍

Three patterns that cost retailers real money
‍

1. Treating a structural problem as a tuning problem. A grocer convinced its forecast is broken keeps re-training the model on more recent data. The forecast is fine. The replenishment logic assumes shelf life behaves the way it did before the assortment widened by 30%. That is an assumption-oriented question, and no amount of parametric tuning will surface it.

2. Buying an optimizer when the bottleneck is data plumbing. A fashion retailer invests in a sophisticated allocation tool. Eighteen months later, two regions are live and the rest are running on spreadsheets. The model was never the constraint. The constraint was getting clean store-cluster, size-curve and markdown-history data into the model on a rhythm the business could trust. That is parametric work, and it is unglamorous, and it is where most of the value sits.

3. Letting unstructured data sit on the floor. Returns notes, customer service tickets, in-store video, search queries, supplier emails. Most retailers have all of it, and almost none of it touches a decision. Input-augmenting AI — turning that signal into a live input for inventory, assortment, fulfillment or last-mile decisions — is where the next cycle of competitive separation will happen, particularly as omnichannel and quick commerce compress the cost of being wrong. This, not chatbots, is where generative models earn their keep in operations.

And underneath all three: most retailers cannot tell you, with credible evidence, what their last AI deployment actually changed. Impact-assessing analytics — the discipline of treating a model rollout like a clinical trial rather than aproduct launch — is still rare. It is also what separates organizations that compound returns on AI from those that simply accumulate tools.
‍

A 60-second diagnostic for your ExCo committee

Before approving the next AI initiative, listen for the language in the room. The phrasing of the problem usually betrays which mode is actually needed — and which one the team is about to deliver instead.
‍

If the language in the room and the mode of the proposed initiative don’t match, you are about to fund a tool that will work as advertised and disappoint as deployed.

Where we see the next wave of value

Across our recent retail engagements, four shifts keep showing up:

■     From forecasting to sensing. Demand signals from search, weather, local events and competitor pricing are mature enough to move from “nice-to-have feature” to a primary input — particularly for short-shelf-life and fashion categories.

■     From network design to network choreography. Stores as fulfillment nodes, dark stores, urban micro-DCs and cross-docks are no longer separate strategic decisions. They are one continuous flow problem, and the retailers solving it are the ones treating it parametrically end-to-end rather than as four siloed projects.

■     From returns as cost to returns as signal. Return reason codes, photos and free-text notes are some of therichest, least-used data in retail. Feeding them into assortment, suppliers corecards and demand models has measurable impact — and almost nobody is doing it well yet.

■     From rolling out tools to proving them. The retailers who institutionalize causal evaluation — holdouts, staggered rollouts, synthetic controls — are the ones that will win the budget conversation in 2027. Everyone else will be defending feelings against a CFO who has stopped accepting them.

The retailers pulling ahead are not the ones with the most AI. They are the ones who match the mode of the AI to the mode of the decision — and who can prove, after the fact, that it worked.

‍
How to use this

If you are a CEO, COO or Chief Data Officer in retail, three questions are worth asking your team this quarter:

■     Of our last five AI initiatives, can we name which mode each one was, and whether the mode matched the problem?

■     For each live model in production, do we have credible evidence —not dashboards, evidence — of what changed when we deployed it?

■     Where is unstructured data we already collect sitting outside every operational decision it could improve?

If the answers are uncomfortable, that is a useful signal. It usually means there is significant value sitting one reframe away from where the team has been looking.