The Easy Assumption

There is a common assumption around AI and B2B product content: if the source catalog already exists as a PDF, the rest should be easy.

Upload it into a chat. Ask for translation, cleaner tables, better images, localized copy and a more modern layout. Get a usable sales asset back.

It sounds reasonable. Until you try it on a real industrial catalog.

What We Started With

The source was a fully rasterized industrial tooling catalog. Mixed languages, dense tables, pictograms, product photos, coating marks, article-code logic and drilling schematics appeared on the same pages.

At first, a small pilot looked promising. We rebuilt a fragment, cleaned the layout, redrew weak visuals and checked the result. Then we tried to scale it. That is where the just-use-AI assumption broke.

Where The Catalog Logic Was Hidden

The catalog was not a document in the practical sense. It was a compressed product system presented as pages.

Availability was not stored as a clean SKU list. It was shown through table markers. Some information lived in pictograms, some in labels, and some in visual differences between tools.

Similar series could differ mainly by shank. Coatings could look close but mean different things. A regenerated product image could look better than the original and still be technically wrong.

Why Looking Finished Is Not Enough

AI can make a page look cleaner. It can make text more fluent and tables feel organized. But in industrial product content, looking finished is not the same as being correct.

A wrong coating, shank, availability marker or image matched to the wrong series is not a cosmetic issue. It is a trust issue.

What Actually Worked

We stopped treating the catalog as a PDF to improve. We treated it as a system to rebuild.

First came structure: page order, series logic, table rules, what belongs in the catalog and what does not. Then source data: final tables were generated from structured data, not edited page by page.

Terminology had to be locked before generation. Every image needed a source crop, product series, role, review status and technical check. QA compared data, images, tables, coating labels, terminology and final output against the source.

AI was part of the work. It was useful for rebuilding poor raster crops into clean catalog-ready visuals. It did not replace the source of truth or the responsibility for technical accuracy.

The Useful Questions

  • What is the source of truth?
  • How is availability encoded?
  • Which decisions are market decisions?
  • Which terminology is locked?
  • Which generated assets need technical review?
  • Who signs off before this reaches a customer?

If those questions are missing, AI can still produce an output. It may even look good. That is exactly why it can be dangerous.

AI is not universal. It is a strong tool inside a controlled workflow. The mistake is treating it as the whole workflow.