Concept buildA worked example of how we would build this for a typical business. Not client work: no client, no client data, no achieved results.
AI Agents
Concept buildA concept build for a US online marketplace where sellers upload products with thin titles and half-empty attributes. An agent reads each new listing and the supplier’s spec sheet, fills category attributes and a description from that data only, validates the result against the category schema and sends it to a merchandiser for approval. Nothing publishes without a person.
The brief
A typical multi-seller marketplace receives a steady stream of new SKUs every week through spreadsheet uploads. Titles are inconsistent, required attributes such as material, dimensions and compatibility are blank or buried in a PDF spec sheet, and descriptions are copied from the manufacturer or missing. Incomplete listings rank poorly in on-site search and filters, so shoppers never find them. A small merchandising team can clean only part of the backlog, and the obvious shortcut, letting a model write the specs, risks publishing attributes nobody has checked.
Explainer
Explainer · 7 slides · about 46 s
Concept buildConcept build: Catalogue Enrichment Agent for an Online Marketplace
Concept build, explained in 7 slides: a concept build for a US online marketplace where sellers upload products with thin titles and half-empty attributes.
Slide 1 of 7
01 · AI agents
Catalogue Enrichment Agent for an Online Marketplace
Concept build: Catalogue Enrichment Agent for an Online Marketplace.
02 · The problem
What a typical online marketplace is up against
A typical multi-seller marketplace receives a steady stream of new SKUs every week through spreadsheet uploads.
03 · The approach
How we would build it, in 5 steps
We would build it in 5 steps, from category schemas and a 200-SKU audit to evaluation before scale.
04 · Architecture
How the pieces connect
How the parts connect, from the first input to the person who reviews the result.
05 · Demo · sample run
Enriching one cookware listing
A sample run, step by step, including where a person steps in.
06 · Typical targets
What we would measure it against
Typical targets, each with the basis it would be measured against.
07 · Next step
Scope a build like this for your business
Book a strategy call to scope a build like this.
Architecture
Architecture of the Catalogue Enrichment Agent for an Online Marketplace concept: New listings → Enrichment agent (product); Supplier sheets → Enrichment agent (spec data); Enrichment agent → Schema validation (draft); Category schemas → Schema validation (rules); Schema validation → Merchandiser (valid draft); Merchandiser → Store catalogue (approved listing).
Diagram with 7 parts: New listings (source), Supplier sheets (source), Category schemas (source), Schema validation (processing), Enrichment agent (model), Merchandiser (human), Store catalogue (system of record).
Connections:
Sample run
Fetch productRoutine
Listing “Steel Skillet 11”: material, size and stove compatibility empty.
Read supplier sheetRoutine
The PDF states stainless steel, 11-inch and induction-ready; no weight given.
Fill attributes + descriptionRoutine
Three fields filled with source links; weight marked “needs seller input”.
Schema checkRoutine
Units and allowed values pass for the cookware category.
Approval gateA person takes over
The merchandiser compares old and new and approves with one edit.
Publish via the Admin APIRoutine
Only the approved version is written to the store.
Build plan
We start from the marketplace’s category taxonomy and write a JSON schema for each priority category: required and optional attributes, allowed values and units. A 200-SKU audit sample shows which fields the supplier data can actually fill, and that sample becomes the evaluation set the agent has to pass.
The agent can do five things: fetch a product, read the attached supplier sheet, look up the category schema, write a draft and submit it for review. It cannot publish, change prices or touch inventory. Every tool call and its result is recorded in a trace the merchandiser can open.
Attributes are filled only where the supplier data states them, and each filled field links to the line it came from. Anything the data does not state stays empty and is marked “needs seller input” rather than guessed. Descriptions are written from the filled attributes and the marketplace’s style guide.
Every draft is validated against its category schema, covering allowed values, units and required fields, before a person sees it. The merchandiser reviews the old and new listing side by side and approves, edits or rejects. Only an approval writes to the store, through the commerce platform’s admin API.
On the 200-SKU set we score fill rate, field accuracy against the supplier data and how often reviewers edit a draft. The agent moves to a new category only after it meets the threshold agreed for the current one, and reviewer edits feed back into the prompts and the evaluation set.
Typical targets
Typical target
70–90% of attribute fields pre-filled for review
Basis: Fields derivable from supplier data in the 200-SKU audit sample
Typical target
Nothing publishes without approval
Basis: Design property: the agent has no publish permission; only the approval step writes to the store
Typical target
Every filled attribute traceable to its source line
Basis: Design property: a source link is stored with each field in the draft
Your project
Every action the agent may take, with its read or write scope and which ones need a person’s approval first.
Each run records the plan, every tool call with its inputs and outputs, and the approval decisions.
Real examples from your backlog scored before every prompt, tool or model change goes live.
Step and cost ceilings per run, so an agent that loops or meets bad data stops and asks instead of carrying on.
Where this fits
FAQ
No. It is a concept build, and the marketplace it describes is a typical one rather than a client. The skillet listing and the agent trace in the demo are sample data, and the fill-rate range comes from what an audit sample would show, not from a live deployment.
It is designed not to. It may fill a field only when the supplier data states it, every filled field links to its source, and empty fields are marked for seller input. The schema check and the reviewer catch what slips through, and the evaluation set measures how often that happens.
Yes, because each category gets its own schema and its own slice of the evaluation set. We roll out one category at a time, usually the one with the largest backlog, and move on only when the current category meets its agreed threshold.
The marketplace does, inside the catalogue it already controls. Drafts and their source links are stored with the listing, and approved text is the marketplace’s to use like any other copy. Descriptions are written from verified attributes and the style guide, never copied from a manufacturer’s page.
Each listing keeps a list of fields marked “needs seller input” with the reason, such as a weight the spec sheet does not state. The marketplace can show that list in the seller portal or send it as a weekly digest, so sellers fill gaps with real data instead of the agent guessing.
Worked examples
Concept builds are worked examples, not client work.
Concept buildAI Agents
A concept build showing how we would build an AI agent that qualifies property leads on WhatsApp and website chat at any hour. It would ask about budget, area and timing, score each lead against the agency’s own rubric and book viewings for qualified buyers, so agents spend their time with people who are ready to view.
Concept buildAI Agents
A concept build for a research agent that prepares a cited company brief before a sales call, so reps start from checked facts, not a pile of browser tabs. It plans its own research, works only through allow-listed tools, links every line to a source, and drafts a CRM note that a rep approves before anything is saved.
Concept buildAI Agents
A concept build for the purchasing team of a UAE trading and distribution company that chases overseas suppliers by hand for order confirmations, ship dates and shipping documents. Four agents split the work: a planner picks the orders that need chasing, a researcher gathers the facts, a writer drafts one email per supplier and a reviewer checks every figure against the sources. A buyer approves each email before it goes, so nothing reaches a supplier on its own.
Next step
Book a free 30-minute strategy call. Bring one process you want to automate and we will talk through scope, integrations and a sensible first step.
Free 30-minute call · no commitment
