19 December Session: Three AI Jobs Inside Product Data

The 19 December 2024 session split AI in product data into three jobs a team can staff separately: images, discovery, and error detection. Each one needs its own owner.

The recording:

Speakers:

• Stephan Spijkers, Co-Founder, PIMvendors, on AI in product data management and discovery
• Morten Næss, EVP of Technology at Bluestone PIM, on AI already usable inside a PIM
• Hendrik van Zwol, "The AI Magician", on hyper-realistic imagery, including the environmental side of AI in product photography

Hendrik raised the environmental cost of AI in product photography.

What the session covered:

Fashion came first. Realistic product images and virtual models, and a path to a more personal shop with less studio time. That is an asset workflow. It fails when the garment attributes, the size run and the colour are wrong, because the picture then sells a product you do not have.

Discovery came next. Search that reads a normal sentence, and suggestions that pull from more than one marketplace. The shopper-facing trick is only as good as the attributes behind it. A vague record produces a vague answer.

Error detection came last, and it is the one most teams skip. The speakers said analytics inside the PIM can spot errors and inconsistencies and correct them. "Correct" is the word to test. A suggestion a merchant accepts is a different control from a silent overwrite.

Staff them as three decisions:

• Images: who approves a generated shot before it replaces a studio photo.
• Discovery: which attributes the search is allowed to trust.
• Corrections: which fields a model may change, and which stay with a person.

Fashion variants, seasons and the speed of onboarding are the catalog shape in PIM for fashion and apparel. Generated images are assets, and the system that should hold them is covered in what DAM stands for. Detecting and correcting errors is a governance choice, set out in product data quality and governance. Discovery aimed at an agent, rather than a classic search box, depends on the record described in AI-ready product data.