Choose bounded tasks
Good tasks have clear inputs and verifiable outputs: mapping a source category to a controlled taxonomy, extracting stated dimensions, normalizing units or drafting a localized description from approved facts. Do not ask a model to invent missing GTINs, compatibility, materials, compliance claims, prices or availability.
Preserve provenance
Store the source field or document, transformation version, generated value, confidence or review status and final approved value. Provenance makes corrections possible and prevents generated text from silently replacing authoritative data.
Validate structure and meaning
Validate types, permitted values, required fields, length and taxonomy fit before publishing. Then apply semantic checks: a valid number can still use the wrong unit, and a fluent translation can still alter a product claim. Google’s Merchant API distinguishes submitted product input from the processed product and its validation state; check both submission and downstream result.
Use human review where risk is high
Review regulated claims, safety information, compatibility, technical specifications and low-confidence mappings. Sampling alone is unsuitable when one wrong value can create legal, commercial or customer-service risk.
Measure the workflow
Track acceptance, rejection reasons, correction categories and drift by model or prompt version. Measure the quality of approved fields, not just the number of generated records.