A merchandiser opens a catalog of 80,000 products. Somewhere in there are the 400 that came back from enrichment with a missing spec, a suspicious weight, or a description that reads like it was written for a different product entirely.
Finding those 400 is the job. And for twenty years, the answer has been the same: a grid, a filter panel, a et of saved views, and a lot of clicking.
Software vendors have made that grid as good as a grid can get. Sortable columns. Faceted search. Bulk select. It works — as long as you already know what you're looking for and how to express it in the filters somebody built for you last year.
That's the ceiling. You can only ask the questions the interface was designed to answer.
The interface should adapt to you
Agentic AI moves that ceiling. Instead of translating your question into whatever filters happen to exist, you just ask.
"Show me products from Vendor X that are missing a UNSPSC code."
"Why did this one come back with no dimensions when the one next to it worked fine?"
"Set the brand to Milwaukee on everything in this filtered set where it's empty."
The Dyver assistant now works directly inside the catalog view. It sees what you see, changes what you're looking at, and acts on the products in front of you. No workflow to configure first, no separate screen to switch to.
Four things it does
It navigates. Filter the table by any field value — vendor, missing data, matching text. Jump to a single product to see it in detail. Reorder or hide columns so the ones that matter are the ones on screen. Switch between To Review and Approved. Narrow to one job or one project. The view reshapes itself around the question you asked, not the other way around.
It explains. This is the part a grid can never do. Ask the assistant to run diagnostics and it surfaces validation errors, warnings, and data quality problems across the whole set. Open one product and it will tell you which enrichment steps ran, and why each field came out the way it did. Go a level deeper and you get the actual trace: the search queries that ran, the pages that matched, which fields fed which steps.
That turns the most frustrating question in catalog work — why is this one wrong? — into something answerable. Point at two products and ask why one was found and the other wasn't, and you get a real comparison instead of a guess.
It acts. Approve products or send them back for review. Edit a field on one product, or bulk-edit an entire filtered set: set values, clear fields, fill only the empty cells, find and replace text. Re-run the enrichment workflow on a selection to recalculate fields once you've fixed an upstream input. Ask for stats and it tells you how many products sit in each quality state.
It helps. It searches the product documentation when you want to know how something works. And when a request is ambiguous — which brand field, which of the two vendors with similar names — it asks rather than guessing.
Why the trace matters more than the chat
Plenty of products have added a chat box. A chat box that can only rephrase what's already on screen is a novelty.
What makes this useful is that the assistant is wired into the enrichment engine underneath it. It isn't reading your catalog — it's reading the record of how your catalog got that way. Every field Dyver produces carries its provenance: the step that generated it, the source it came from, the inputs it depended on.
So when a merchandiser asks why a description is thin, the answer isn't a plausible-sounding paragraph. It's the actual reason: the supplier page had no spec table, the fallback source didn't match, here's what ran.
That's the difference between a catalog you argue with and a catalog you can trust.
What this changes day to day
The review queue stops being a scroll. You ask for the problem set, you get the problem set. You ask why, you get why. You fix it in bulk, re-run the affected products, and move on.
For teams running catalogs in the tens or hundreds of thousands of SKUs, most of the work was never the fixing. It was the finding, and the explaining, and the re-explaining to whoever asked why product 44,912 looks wrong.
That part now takes a sentence.