External-source research synthesis. This is not Sander Gansen’s client implementation.
The real problem is not a missing chatbot
A furniture finish is not merely a colour name. It is associated with a product family, a supplier, textures, image assets, markets and approved combinations. When three regional SKU records describe the same underlying material differently, a buyer, dealer or assistant can receive a plausible but inconsistent answer. Faster writing does not resolve which record is correct.
Akeneo’s Steelcase customer story describes spreadsheet-based surface-material work, manual image and texture associations, and three regional records per SKU. O2 Commerce led the Product Cloud implementation, configured Supplier Data Manager, automated asset linking and aligned taxonomy and governance across internal and external users. The operating changes—not an isolated model prompt—are the useful executive lesson. Read the source case.
Keep each reported result in its own unit
The case reports annual supplier-data loading falling from 33.36 to 12.14 hours, described as a 62% reduction. Time per product family fell from 22 to 8 minutes, described as 63.64%. Product-release work fell from 16.68 to 6.07 hours per release.
These are provider-published operational outcomes. The source does not give a measurement period, sample count, control or independent audit. Annual load hours are not hours saved every week. Product-release time is not customer conversion, and a reduction in one task cannot be extrapolated across an entire department. The case is useful precisely because the tasks are specific enough to discuss—not because it proves a universal revenue improvement.
One source of truth is often the wrong slogan
A PIM governs product descriptions, attributes, assets, locales, variants and publication workflow. It does not automatically own stock, financial records, released drawings, available production capacity or the price a customer is entitled to receive. Those authorities may reside in ERP, PLM, approved commercial rules or a named decision-maker. Akeneo’s PIM explanation and API documentation describe the product-information layer.
The practical requirement is a source hierarchy and clear ownership. An assistant should know where an approved finish description comes from. It should not infer that the same record can approve compatibility, certification, delivery or a special price. A machine-readable field is a representation of a decision, not permission to make that decision.
A useful first family for a manufacturer
Consider a furniture exporter with finish choices distributed across its own catalogue, two dealer portals and a supplier spreadsheet. This is an illustrative application, not a reported deployment. Choose one family with frequent discrepancies. Record its canonical identity, approved attributes, material and finish, asset links, locale, channel, revision, effective date and approving owner.
Separate public descriptions from customer-specific prices and availability. Define required fields and permitted units before importing supplier data. If two files disagree on a finish code or permitted combination, quarantine the record rather than asking a language model to decide which sounds more likely. A source-linked draft may help the owner resolve the conflict; publication needs the agreed approval.
The same discipline applies to a control-cabinet manufacturer whose voltage option, drawing and translated dealer name describe different variants. A buyer needs the correct identity and conditions, not a more persuasive description of the wrong item.
Measure the work where it changes
Track supplier-load minutes, family-release minutes, correction rate, rejected imports, stale attributes and contradictions in downstream dealer descriptions. When testing AI extraction, add source-linked attribute accuracy and unsupported-claim rate. Keep the original source and revision so corrections can be traced.
A complete record can still produce a wrong answer when relationships or exclusions are absent. Test near-identical variants, missing units and expired revisions, not only the cleanest records. Ask whether an incorrect compatibility claim reaches publication, whether a conflict triggers review and whether the owner can revoke a stale fact across channels.
Questions for the executive sponsor
Who confirms the technical attributes? Who owns translations? Which system owns price and stock? What is the expected correction time in dealer channels? What should happen when the owner is unavailable? What work will remain after the implementation partner hands over?
Start with a manageable family and operating rules that the team can maintain. The outcome is a reusable commercial information foundation, not a guarantee that an external assistant recommends the product. Explore OfferSignal, the focused Business & AI Workshop and the related research on industrial knowledge and quoting.
Hands-on product, business and AI leadership for executives of established companies. Working on agentic commerce through OfferSignal.
About Sander