How retail data becomes useless without a decision layer

There's a question coming up more and more often in retail executive committees, one nobody dared formulate a few years ago. "We've invested so much in data over the past five years — data warehouse, data lake, Snowflake, Databricks, data engineering teams, modern BI platform, predictive models... where's the return?". The question is polite. Rarely frontal. But increasingly present, and increasingly hard to dodge.

Because the observation is troubling. Chains that massively invested in their data infrastructure find themselves, ten years after the great wave started, with remarkably sophisticated stacks — petabytes of collected data, trained ML models, dashboards of impressive graphic beauty. And with operational KPIs that have, on average, barely progressed. Markdown rates stay high. Stockouts persist. Overstocks accumulate. Sector net margin stays tight.

This paradox — mass of data accumulated but performance unchanged — isn't a technical failure of data. It's the expression of a truth few vendors and consultancies want to say out loud: retail data, on its own, produces no value. It only becomes useful the moment it feeds an executed decision. Until that last step is crossed, all the upstream investments — collection, storage, transformation, modeling — remain cost without benefit.

This article looks straight at that blind spot: why accumulating data isn't enough, why BI and data science layers don't bridge the gap, and what to build so your data investments stop being a cost center and finally become a performance lever.

The 2015-2020 data promise: revisiting what was actually sold

To understand where we are, you have to revisit what was promised. From roughly 2015, retail shifted into a "data-driven" logic carried by three converging narratives.

First narrative: "data is the new oil". An appealing formula suggesting it was enough to extract, refine and store data for it to release value. Second narrative: "break the silos, create a single source of truth". Hence the wave of data lakes, then data lakehouses, meant to become the enterprise's single source of truth. Third narrative: "AI and ML will transform retail". Hence massive data scientist hiring, algorithmic POCs, proprietary model announcements.

These three narratives aren't false. Data does have potential value. Breaking silos is a good idea. AI can produce value. But they share an implicit bias: they treat data as an end in itself, when it's only a means. And that bias structured a decade of investments that accumulated upstream without building downstream — the layer that turns data into executed decisions.

The result is a situation I'd call "data rich, decision poor". Chains now have everything they could have dreamed of in raw material terms. And they still don't know how to exploit it to steer their network.

The "data-driven" misunderstanding: what it doesn't mean

You have to be precise on vocabulary, because that's where the costliest confusions hide. Being "data-driven" isn't having lots of data. It's not having pretty dashboards. It's not even having sophisticated predictive models. All these things are necessary but largely insufficient conditions.

Being truly data-driven is having built a complete chain where data feeds decision, decision feeds execution, execution produces a measured return, and that return feeds data again — without systematic intermediate human breakage. It's an end-to-end property, not a characteristic of the upstream alone.

In most chains, however, "data-driven" stops upstream. Data is collected, stored, transformed. It's possibly modeled. It's surfaced in dashboards. And then... nothing. Or rather: a human opens the dashboard, interprets it their way, manually formulates a decision, and has it executed by teams not connected to the initial data system. The execution feedback, in turn, almost never returns to the data chain.

The result is an organization that has invested tens of millions of euros to produce quality descriptive data, which it consumes at 5% of its potential because it didn't build the layer that would turn this data into operational performance. That's exactly what "data rich, decision poor" means.

The four stages of data value — and why almost nobody gets past stage 2

To make this more precise, here's a simple four-stage frame. Each stage corresponds to a type of value produced, and each stage corresponds to a type of investment required.

Stage 1: data exists and is accessible. The founding stage. Your sales, stock, e-commerce, supply chain data are collected, cleaned, stored, accessible via queries. Most retailers have reached this stage, sometimes imperfectly, but the initial investment is done. At this stage, data has no intrinsic value. It's just there.

Stage 2: data is surfaced and understandable. The classic BI stage. Your data is presented in dashboards, reports, ad-hoc analyses. Teams can browse, explore, extract observations. The majority of retailers sit at this stage. The value produced is descriptive: you know what happened. But this value stays largely potential — it depends entirely on what humans do with the understanding.

Stage 3: data feeds an automatic prescription. The stage where a system no longer just shows data, but automatically draws an action recommendation from it. Not a human opening a dashboard and thinking — an engine that scans data and formulates, for each SKU/store, what to do (mark down, transfer, replenish, return). Very few retailers have reached this stage, and those who claim to often confuse it with parameterized alerts that aren't real contextualized recommendations.

Stage 4: the prescription is executed and the loop learns. The ultimate stage where the recommendation propagates automatically to execution systems (ERP, WMS, pricing, e-commerce), where its effect is measured, and where this return feeds the system to adjust subsequent recommendations. At this stage, data finally produces the economic value it was collected for. A very small fraction of retailers — typically global leaders — actually operates at this level.

This grid is useful for a simple reason: it reveals that going from stage 2 to stage 4 is a qualitative leap, not a quantitative improvement. You don't cross that leap by adding more dashboards, more data scientists, or more storage. You cross it by building a layer of a different nature — the decision and execution layer — which has almost nothing to do with what was deployed upstream.

Why the descriptive layer never becomes prescriptive "on its own"

One of the most costly mistakes retail leaders make is believing that by continuously perfecting the descriptive layer (BI, dashboards, models), they'll eventually mechanically produce prescription. This belief is comfortable because it justifies ongoing investments and indefinitely postpones the moment to change paradigm. But it's false, for three structural reasons.

Reason #1: BI doesn't ask questions. It answers the ones you ask. As long as a human has to open the dashboard and ask "what should I look at today?", the overwhelming majority of situations deserving action aren't looked at at all. No improvement of BI changes this property — it's baked into its architecture.

Reason #2: predictive models predict, they don't prescribe. A forecast at 92% accuracy still doesn't say whether to mark down, transfer, or do nothing. These decisions require arbitration logic models don't carry. You can stack increasingly sophisticated models, the prescriptive leap won't happen spontaneously.

Reason #3: the descriptive layer doesn't execute anything. Even if it managed to produce a recommendation, there's no automatic channel between that recommendation and execution systems. The decision has to be re-entered by a human into another tool — which produces delay, attrition, loss of consistency. The chain stays structurally broken.

Until these three architectural properties are addressed by a dedicated layer, the descriptive layer will stay what it is: an infrastructure cost useful for the organization's data culture, but with no directly measurable impact on net margin.

The hidden cost of unexploited data

How much does this situation — data accumulated but not turned into decision — cost? The exercise is rarely done honestly in chains, but when it is, the orders of magnitude are striking.

You have to add up several lines. First, the direct cost of data infrastructure: cloud licenses, ETL, BI platforms, model subscriptions. For a mid-size retailer, regularly between €2 and €5 million per year. Second, the cost of data teams: data engineers, data scientists, BI analysts, project managers. Several more million euros. Third, the opportunity cost of projects that don't ship: algorithmic POCs that never make it to production, dashboards built then abandoned, data collected and never consulted.

But the biggest cost isn't in the data project bill. It's in operational opportunity cost. A chain that has all the data needed to optimize its decisions, but keeps making them by hand because it doesn't have the prescriptive layer, absorbs an opportunity cost amounting to several net margin points per year. On a retailer generating €500M in revenue, that's €5-15M per year of recoverable margin staying on the table.

Cumulated over a decade of data investments without a decision layer, you're talking tens of millions of euros of unrealized potential. And the bill keeps running as long as the prescriptive layer isn't in place.

The paradox: the more data you have, the more the absence of decision layer costs

Here's the dynamic that makes the current situation so particular. The more quality data a chain accumulates, the more the absence of a decision layer becomes economically absurd — because the unrealized potential value grows as the raw material gets richer.

A chain with little data and no decision layer has little to gain from building one — it lacks raw material. A chain with vast data and no decision layer has enormous gains to capture — its raw material is underexploited at industrial scale. That's exactly the situation most large retail chains find themselves in today: they spent ten years enriching upstream, and every year that passes without building downstream makes the gap more costly to absorb.

This inverted dynamic — "the more we invested, the more we lose by not taking the next step" — isn't intuitive. But it's mathematically implacable. And it explains why the most lucid data leaders are starting today to redirect their investments away from upstream (where marginal return collapses) toward the prescriptive layer (where marginal return explodes).

What to build: a dedicated decision layer

Building a decision layer isn't a natural extension of your current data stack. It's a project of a different nature, requiring specific architectural choices.

First, continuous prescription logic. The system must permanently scan all available data, identify situations deserving action, and formulate the adapted recommendation. Not wait for a human to ask. Not limit itself to situations hitting a parameterized threshold. Continuous, contextual, prescriptive surveillance.

Second, business rules as first-class citizens. Recommendations must respect by construction the chain's operational constraints — margin floors, supplier constraints, commercial calendars, specific handling. Without this integration, recommendations are theoretically optimal but practically inapplicable.

Third, frictionless execution. Validated decisions must propagate into existing execution systems without re-entry. The delay between formulating a recommendation and its real effect on the floor must be compatible with commercial dynamics — hours, not days.

Fourth, a measurable return loop. The effect of each executed decision must be measured and integrated into the system to adjust subsequent decisions. It's this learning loop that turns a static decision layer into an increasingly intelligent system at every cycle.

A platform carrying these four capabilities radically transforms the value of your existing data stack. Your data stops being an infrastructure cost and becomes the fuel of a decision system that measurably produces margin. And the awkward question of data ROI finally finds a positive answer.

The trap to avoid: rebuilding your data stack before adding the decision layer

A final warning, because it costs retailers who fall into it dearly. When data leaders become aware of the gap between their current stack and the decision layer they need, the temptation is often to rebuild upstream before adding downstream. "Before building the decision layer, we first need to modernize our data warehouse, rebuild our models, improve our data quality."

This logic is almost always a mistake. For two reasons.

First, it indefinitely postpones the moment value is actually produced. A data lake rebuild typically lasts 18 to 36 months. During all that time, the organization keeps absorbing the opportunity cost of having no decision layer — which amounts to millions per year.

Second, and this is the key point, perfect data quality isn't a prerequisite for the decision layer. A well-designed modern decision layer knows how to deal with imperfect data, identify low-quality zones, request complements, prioritize decisions where data is reliable enough. It doesn't wait for everything to be perfect to start producing value — and by starting to produce value, it often finances the data improvements it needs.

The right approach is therefore inverse to the usual intuition: start with the decision layer on what exists, identify zones where it hits data quality issues, and evolve data based on the prescriptive layer's needs. Not the other way around.

The Solya approach: turning your data stack into a decision lever

That's precisely Solya's philosophy. Not a new data lake. Not yet another BI tool. Not yet another ML model. But the decision and execution layer that turns your existing data investments into measurable operational performance.

Concretely, Solya connects to your data sources — POS, ERP, e-commerce, supply chain, existing data lake — and rebuilds an exploitable view at the SKU/store level, continuously updated. The decision engine permanently scans that view to identify situations deserving action, formulates for each a contextualized recommendation embedding your business rules and operational constraints, and propagates validated decisions to your execution systems without re-entry. Observed effects feed the engine back to adjust subsequent decisions.

The paradigm shift is simple to state. You invested to have lots of data. Solya adds the layer that turns that data into margin. Your existing data infrastructures aren't replaced — they finally become profitable, because they're now wired to a system that exploits their potential all the way down the chain.

For data leaders, it's also an answer to the awkward ROI question. Instead of having to keep justifying data budgets on promises of future use, you can now justify them on measurable recovered margin. It's a positioning shift that turns the data function from a tolerated cost center into a recognized performance lever.

The real question to ask

What share of the data you collect today actually produces recovered margin for your organization? Not how much is stored. Not how much is consulted in your dashboards. How much feeds an executed operational decision that wouldn't have been as good without it.

If the answer is "a marginal fraction", you're living the paradox this article describes: a data-rich but decision-poor organization. And every year that passes without building the prescriptive layer worsens the opportunity cost you absorb.

Exiting this paradox isn't through more data, more models, or more dashboards. It's through crossing the threshold that turns data into executed decision. And it's this threshold, more than any other transformation, that separates today's retailers who actually monetize their data investments from those who keep accumulating them, hoping that one day they'll produce the promised value.

---

Is your data actually producing margin?

At Solya, we offer retail leadership teams a personalized 30-minute diagnostic to assess, on your own context, the real exploitation rate of your data investments — and quantify the margin potential recoverable by adding a prescriptive decision layer on top of your existing stack.

👉 [Book your Solya diagnostic] — 30 minutes, by video, with one of our retail experts.

You'll walk away with:

• A map of your data investments' real exploitation rate
• A quantified estimate of margin potential recoverable through a decision layer
• The first high-ROI use cases to turn your data stack into an operational lever