The Closed Door — product data as a condition of entry

A product line rarely dies at the shelf. It dies at the gate.

Before a shopper ever sees an item, that item passes through thresholds it does not control. A retailer’s item setup process asks a defined set of questions, in a defined vocabulary, with defined tolerances. Answer them and the door opens. Leave one required attribute blank, or supply it in a form the receiving system rejects, and the door stays shut. The product waits outside the room where merit is judged.

That is the part leaders tend to underweight. Product data reads like an administrative burden and behaves like a condition of entry. Every week an item sits in rejection status is a week of shelf presence forfeited, a week of ad spend pointed at a page that does not exist, a week of a launch calendar quietly slipping. Companies measure the cost of a wrong price with precision. The cost of the closed door usually goes unmeasured.

This problem has a structure. Three lines separate what belongs under governance, what must be translated at the boundary, and what can be delegated to a machine. In my experience across nearly thirty years of SAP implementations and retail systems work, drawing those lines in the wrong place is the source of most of the expensive failures I have watched unfold.

Line one: what belongs under governance

Some facts about a product are singular by nature. They have exactly one correct value, one accountable owner, and a legal or financial consequence when they are wrong. These facts belong in your master data system — SAP MDG, a PIM, or a governed record of any kind — and they should not be touched downstream.

Identity. What is this thing, and how do we know it is the same thing everywhere? Global trade item numbers, internal item numbers, and the resolution rules connecting a manufacturing SKU to a selling SKU to a case to a pallet. Identity is the anchor everything else hangs from.

Hierarchy. How items roll up into products, brands, categories, and reporting structures for planning and financial consolidation. Hierarchy is what makes a number addable.

Regulated attributes. Allergens, nutrition panels, ingredient statements, country of origin, hazardous materials classifications, certification claims. These carry statutory exposure. They belong to a named owner, and they change under a controlled process.

Lineage. Who changed which value, when, on whose authority, and what the prior value was. Lineage is the difference between a recall you execute in hours and one you execute in weeks.

Governance here means ownership and process before it means software. Most companies already license a system capable of holding these facts — SAP’s Material Master alone can carry most of this if the fields are populated and owned. What they lack is a named accountable owner and a written definition. The platform is rarely the gap.

Line two: what gets translated at the edge

Here is where enterprise thinking usually goes wrong, and where the most expensive SAP implementations I have seen have produced the least return.

Retailers will never converge on one taxonomy. That bears stating plainly, because a great deal of architecture gets built in quiet hope that they will. A retailer’s category tree, required attribute set, controlled vocabularies, character limits, image specifications, and copy rules are competitive instruments. They encode how that retailer believes its shoppers search and decide. Convergence would mean surrendering a difference each of them actively defends.

So, the same governed fact arrives at each destination in a different form, and each of those forms is correct in its own context. A pack configuration described one way in one retailer’s spec and another way elsewhere. A certification claim permitted in one copy environment and restricted in another. A single item that legitimately sits in different category nodes at different retailers, because those retailers organize shopping differently.

The instinct — and I have seen this across hundreds of implementations — is to resolve this upstream: pick one representation, force it into the master record, push it everywhere. That instinct produces a record that is wrong for every destination in a slightly different way, and it buries the wrongness where nobody can see it.

The sounder architecture holds one governed fact and produces plural outputs, one per destination, each valid when it arrives. Translation becomes a first-class function with its own rules, its own version history, and its own owner — separate from the master record, not collapsed into it. When a retailer changes a spec, one mapping changes. The governed record stays untouched.

The organizational consequence matters more than the technical one. Once translation is explicit, you can measure it. Rejection reasons become a data set. You learn which retailer costs you the most days, which attribute breaks most often, and whether a given failure originated upstream in governance or downstream in mapping. That visibility is the first return, and it arrives well before any efficiency gain does.

Line three: where AI earns its place

AI is genuinely good at a narrow band of this work, and the band is worth naming precisely.

It is good at proposing. Mapping a new retailer’s attribute set against an existing dictionary. Suggesting a category node from a title and an image. Flagging a value that falls outside the historical distribution. Drafting copy variants inside a hard constraint set. These are pattern tasks over messy text, and a competent model performs them faster and more consistently than a coordinator working a spreadsheet at eleven at night.

Adjudication is a different act. Deciding that a proposed mapping is correct, that an anomaly is acceptable, that a regulated attribute may be published: each of these is a decision with an owner. Keep it human and keep it auditable.

In practice, that means four requirements on every automated proposal: a stated rationale, a confidence value, a reversible record, and a named approver for the classes of data where being wrong is expensive. This is not a new governance concept — it is exactly the escalation logic that well-run SAP change control has always required. What changes is the volume and the speed at which proposals arrive.

Confidence thresholds are the practical control. High confidence on a low-consequence attribute can flow through with sampling. Anything touching a regulated attribute stops at a person, every time, with no exception granted under volume pressure. Write that rule down before you are under volume pressure.

Where a translation layer is the wrong answer

Honesty about scope is the price of being useful here.

A brand selling through a single retailer faces a single spec. The answer there is discipline and a checklist, applied consistently by someone whose job it is.

A brand whose real failure sits upstream will get no relief from better translation. Where nobody owns item identity, where no written attribute dictionary exists, where marketing and supply chain hold different values for net weight and each side believes it is right, automating the boundary distributes the confusion faster and at greater cost. This is the most common failure mode I encounter: companies investing in translation tooling before their master data is governable. The tool accelerates the error.

Regulated attributes should be resolved under governance and left alone at the edge. If a retailer’s spec appears to require transforming an allergen statement, the answer is a conversation with that retailer, not a mapping rule.

What to build, regardless of what you buy

Five things. Four of them cost nothing beyond decisiveness.

Name owners. One accountable owner for item identity, one for regulated attributes. Names, in writing, known to the organization.

Publish an attribute dictionary. Definition, format, source system, owner, per attribute. A spreadsheet is an acceptable first version.

Instrument rejections. Capture reason codes by retailer. Review the pattern monthly with the owners present.

Turn on lineage. Most licensed platforms support change history already and ship with it dormant. This includes SAP. The capability is there.

Set escalation rules before automation arrives. Confidence thresholds, sampling rates, and the list of attributes that always stop at a person.

Any of these raises the ceiling on what a tool can later achieve. Skipping them lowers it permanently.

The threshold, again

Product data is the set of answers a company gives at someone else’s door. Govern the facts that are singular and consequential. Translate at the boundary, because the doors differ and will keep differing. Let machines propose and let people decide, with a record that survives an audit.

Do that, and the door opens on the first attempt — which is the outcome the market actually rewards.

Rej Pathania is Co-Founder and Chief Revenue Officer of BridgeCommAI Inc. He writes here in a personal capacity, drawing on nearly thirty years of retail and consumer goods systems experience. Articles argue from the mechanism, never the product, and are vendor- and firm-neutral. This piece is published as a disclosed pair with What Arrives at the Door by Mohamed Amer — the two authors are co-founders of BridgeCommAI, writing from opposite ends of the same door.

Next: meet Rej Pathania and his column Condition of Entry · browse the sapperment library · or explore all sapperment experts.