Your Best Salesperson Is a PDF Nobody Owns

How we built a production product-data and asset system for a three-brand furniture manufacturer — 610 products, 4,342 digital assets — with AI in the build seat and humans on the decisions.

Somewhere right now, a sales rep is attaching a spec sheet to an email. It has last year’s price on it. He doesn’t know that. Neither does the customer. Both of them will find out about six weeks from now, in a conversation nobody enjoys.


That is a marketing problem. It will never appear on a marketing budget.


The category nobody owns Marketing owns campaigns. IT owns the ERP. Sales owns the relationship. Product owns the roadmap.


The product record belongs to nobody. So it lives in eleven places, and all eleven disagree. Every manufacturer we talk to has a version of this. Most have stopped calling it a problem. They call it “how it is.”


But look at what sits downstream of the product record: your website, your spec sheets, your price sheets, your dealer portal, your rep’s quote, your shopping feed — and now your visibility in AI-generated answers, which is decided almost entirely by whether machines can read structured facts about your products.


All of it runs on one thing. And that thing is usually a spreadsheet.


What we built Beaufurn is a contract furniture manufacturer running two brands — Beaufurn and Shelter Outdoor — into hospitality, restaurant, senior living, education, and A&D. Deep catalog, heavy attributes: finishes, textiles, COM options, dimensional variants, certifications.


In production today. Not a pilot.


How we did it We inventoried before we built. Crawled the live catalog end to end — every product page, image, swatch, and linked PDF — then pulled the ERP export and the internal pricing files and reconciled all three against each other. The output was a written gap report: what exists, what’s actually published, what has drifted. That report was the business case.

We modeled the product, not the page. Families and variants as an explicit relationship. Dimensions, materials, construction, application, and certifications as queryable fields instead of sentences buried in copy. Finishes and textiles as records, not JPEGs. Bill of material and cost attached to the same record as the marketing language.


We built the asset layer as infrastructure. Thousands of images, swatches, brochures, and spec PDFs ingested, de-duplicated by content hash, and bound to product records. One rule: the storefront never holds its own copy. One asset, one home, many consumers.


We made documents a function of the data. Spec sheets and price sheets stopped being files somebody maintains and became outputs generated on demand. Change a dimension once and everything produced after that moment is correct.
 The most expensive object in most marketing departments is the PDF that was accurate the day it was made.

We put pricing behind a service and roles in front of it. Admins see cost. Inside sales sees list, discounts, and volume breaks. Reps see what reps should see. Proposals carry a frozen price snapshot, so a quote stays a quote.


We made the website a consumer of the record. Products, variants, attributes, pricing, and media publish automatically to the storefront — which means the catalog is now recoverable independent of the website. Re-platform, redesign, change agencies. The product truth doesn’t move.

We built the ERP swap seam. The client was mid-ERP evaluation. Conventional advice says wait. That advice costs companies fiscal years. We routed all pricing through one abstraction layer instead, so when the ERP lands we change a layer, not a system.


Where AI actually did the work AI did the extraction, normalization, and reconciliation at a scale nobody would fund by hand — thousands of pages and documents parsed into structured fields. It drafted the schema, wrote application code, and checked its own results against written acceptance criteria.

AI did not decide what a product family means for this business, who is allowed to see cost, or how discounts should work in a channel with reps, dealers, and specifiers.
 AI compressed the labor. It did not make the decisions. Anyone telling you otherwise is selling a demo.

That is a better story than the hype anyway. A scope like this is conventionally a six-figure, multi-quarter integration. The judgment-light 70% collapsed, which left budget and calendar for the 30% that required people who understand the business. Spend the savings on judgment.


Why a marketing agency built this You can buy the best media plan in your category and it will still underperform if the product page it lands on is missing dimensions, showing a discontinued finish, and carrying a price from two years ago. At some point, optimizing the ad is malpractice when the problem is the catalog. More importantly is the next phase of customer self service and relationship management. This is where the long term marketing strategy takes place.


We are a marketing agency that ships systems. That sentence should not be unusual. It is.

If the answer to “where does our catalog come from?” is a person’s name, you are about to pay for the same problem twice. We run a fixed-scope product data audit that reconciles your master catalog against what is actually published. Days, not quarters — and you own the output whether or not you do anything else with us.