AI connectors for manufacturer catalogs: who is authoritative

The app-store precedent

When mobile app stores opened, the businesses that did well were not always the ones who built the best interface. They were the ones who owned something a store could not replicate: a catalog, a customer relationship, a verified identity, a fulfillment network. The interface was a thin layer on top of that.

A connector between an assistant and a product catalog repeats the pattern. The assistant supplies the conversation. What the connector exposes, and whether the buyer can trust it, depends on the catalog behind it.

What a connector exposes

At its simplest, a connector is a defined interface through which an assistant can ask a catalog for a part, a specification, an availability or a quote, and receive a structured answer instead of scraping a page. Whether the interface is a purpose-built one or an established API, the questions it forces are the same.

• Which identifiers does it accept: manufacturer part number, distributor SKU, GTIN, or a cross-reference?
• What does each returned value carry: its source, its revision, the date it was last confirmed?
• What happens when the value is unknown: an explicit unknown, or an empty field the assistant fills?
• What can the assistant do next: read, request a quote, place an order? Who approves?

Authority is the hard part

A buyer asking about a part may reach a manufacturer's connector, a distributor's and a dealer's. Each answers about the same item with different data: the manufacturer with the engineering values, the distributor with its SKU and branch stock, the dealer with price and lead time. If they disagree, the assistant needs a rule. The rule cannot live in the assistant. It has to be carried by the data.

That means each value declares its origin and its owner. A connector that returns a dimension from a distributor's catalog without saying it was transcribed from the manufacturer's data sheet at a given revision is passing along a copy and calling it a fact.

Read before write

The prudent sequence is read access first. A connector that only reads, with every value carrying its source and date, can be measured against the record and corrected. Write access, such as a quote request or an order, comes after the read path is trusted, and carries a human approval until the failure rate is understood.

This is the order Lehvel's work follows: a checked record first, then publication to the places machines read, then a controlled action. See Business Truth and the AI Observatory.

Questions