Why generation count is a poor KPI

As a catalog grows, employees copy supplier descriptions, normalize names, fill specifications, and find duplicates. A model can speed up drafts, but a wrong material, size, or compatibility claim can lead to a return or complaint. Measure the cost of a card that passes required checks and receives human approval.

Structure is not truth

Ollama supports constrained output, including JSON Schema. This helps produce expected fields but does not prove they are true. Separate fact extraction, normalized suggestions, schema checks, reference-data validation, human approval, and writing through the catalog service.

For dimensions, voltage, composition, compatibility, and legal claims, require a source. If the supplier file lacks a value, leave it blank rather than letting the model guess.

An illustrative calculation

Assume a manual card takes 8 minutes and fully loaded labor costs RUB 900 per hour: that is RUB 120 per card. In an AI workflow, model and infrastructure are assumed to cost RUB 5 per draft; three minutes of review add RUB 45. If 10% need another four minutes of correction, that adds an average RUB 6. The modeled total is about RUB 56 per accepted card.

This is an example, not a company measurement; it excludes integration, labeling, licenses, QA, and returns. If review takes seven minutes or many cards need a full rewrite, savings disappear. Track generation and retries, review time, acceptance, critical errors, and support separately.

A one-category pilot

Choose a category with structured data and a moderate cost of error. Take 300–500 SKUs, remove personal and contractual details, and reserve an independent test sample. Record the source and allowed values for each attribute.

Generate suggestions only at first. Compare manual and AI groups for speed and quality; test missing values, rare variants, contradictions, units, and similar SKUs. Automatic writes should be limited to fields with reliable sources and deterministic validators.

Next step

Record minutes per card, correction frequency, and losses caused by errors, then run the same sample through structured drafting and human review. The management metric is the cost of a correct card ready to publish. If the pilot does not lower it without reducing quality, it is too early to scale.