What a Catalog Should Propose, and What It Should Leave to You
A line between what a furniture catalog should prepare or propose for review (mappings, variant grouping, copy, images, channel versions) and what must stay a human decision (facts the supplier didn't give, prices, what goes live).

The promise of an AI catalog is that it does the preparation your team used to do by hand. The worry is that it does the deciding too. Both are reasonable. The way to judge any tool is to sort its features into what it prepares or proposes for your review and what it decides outright, and to check that the second list is short.
What the catalog can prepare or propose for review
These take judgment, which is why they should be proposals you confirm rather than decisions made silently. The value is in preparing more of the range for review without removing the merchandising decision.
Where the header is and what the columns are. Title rows above the header are recognized; the assistant previews the sheet and suggests where each column goes [1]. "DESCRIPTION" is probably the product name, but that's a proposed mapping until you confirm it.
How the rows group into variants. For a range with variants, the assistant proposes or asks about the grouping before it builds the structure [1]. Whether a different seat height is a size variant or a separate product is your merchandising call, and your channels' rules may have a view too.
Copy written from facts you have. An Autofill column writes from the fields you allow it to read, under an instruction you wrote, with a preview and a ten-cell sample before it continues [2]. If the supplier's specification says oak legs and bouclé upholstery, the description can say so. A photo alone can suggest a wood-like finish and textured fabric; it cannot confirm construction, species or composition.
Images made from a reference. Image fields on the product can be filled from product references across selected products [3]. When the Furniture Connect Agent generates or edits imagery using product references, its workflow includes a visual comparison of product presence, silhouette and proportions, material, finish and color. Identified mismatches or uncertainty are flagged for review. This check does not guarantee accuracy or replace a person's approval before publication, and a generated image is not proof of a product specification.
Each channel's version from the record. Next's readiness score and template export, Shopify's content sync, per-channel titles and descriptions, retailer-sized image exports, all from one record [4][5][6].
A record of what happened. Which products were added and updated on import; for Autofill, a per-run undo that keeps your later edits and a per-cell revert [1][2]. That covers Autofill runs; it isn't a universal undo for imports, media changes or edits made in connected systems.
What should stay a human decision
A fact nobody supplied. Material, dimension, weight, country of origin, certification, warranty, safety or performance claim, delivery promise. If it isn't in the supplier's data or an approved source, no column should fill it, and the instruction and the sample run are where you make sure of that.
Prices, SKUs and barcodes. A derived retail price belongs in a formula column, which is arithmetic on a figure you supplied [2]. Nothing generated should touch the fields that make a product sellable and reorderable.
Whether something goes live. New products arrive as drafts [1]. New Shopify products are created only when you select them and choose Create Shopify drafts, and stay unpublished until you publish them in Shopify [5]. Content sync to existing Shopify listings is a separate, continuous thing, and conflicting edits are yours to resolve [5]. Channel exports are files you hand to the channel's own process [4].
What a change should cost. Credit estimates and a budget before an Autofill run, pause during it, undo after [2]. Larger jobs in the Agent need your approval before they run [7].
Why the line sits here
The first list is preparation: work where a good proposal plus a quick review beats an hour of typing. The second list is consequence: a return, a rejected listing, a mis-priced product, a customer told something untrue, or a listing live before anyone meant it to be.
Software that quietly moves items from the second list to the first is selling speed at the price of the thing you're responsible for. It's usually sold as "fully automatic", which is a phrase to read carefully, as what 'assisted' import means sets out.
What this means for the team
The job shifts from making the product data to reviewing it and making the calls: confirm the mapping, answer the grouping questions, read the sample, decide what's live where, and send the supplier one email listing the facts only they have. Repetitive entry can be reduced; the judgment stays.
For the whole journey from supplier file to buyable on every channel, see getting a new furniture range online.
Sources
- Furniture Connect Help Center: How to import products into the Agentic Catalog
- Furniture Connect Help Center: How to fill product data with AI Autofill
- Furniture Connect Help Center: How to create image fields for your products
- Furniture Connect Help Center: Connecting the Agentic Catalog to sales channels and ERPs
- Furniture Connect Help Center: How to connect your Shopify store
- Furniture Connect Help Center: How to export product images at any size
- Furniture Connect Help Center: Approving larger jobs in the Agent




