CSV Pilot uses AI to draft product data, but the final call is yours. Understanding what the AI does well, what it guesses, and where the preview protects you makes the difference between a great catalog and a draft that looked finished. This guide explains the division of labor: what the model writes, what deterministic checks handle, and what only you can decide.
What the AI does
The AI turns natural language into structured products:
- Names and descriptions — readable, consistent product copy.
- Categories and tags — assigned from the vocabulary you name.
- Variants and add-ons — size and topping structures built from your description.
- Prices — the values you state, or plausible ones when you do not.
You say "create 8 pizzas with sizes, ₹299–₹599"; the AI produces 8 product rows with Size option groups and prices in that range. The quality of the output tracks the quality of the input: "10 pizzas" produces ten generic pizzas; "10 pizzas — 4 margherita, 3 pepperoni, 3 farmhouse, ₹299–₹599" produces a catalog you could almost ship.
What the AI infers
Anything you do not state, the model fills in — and marks as inferred. If you say "create 5 pizzas" without prices, the AI picks typical pizza prices. If you do not name a category, it invents one. These inferred values are highlighted in the preview grid.
The preview is where inference becomes review. A highlighted price means "I guessed this — confirm it." A highlighted category means the same. You can click any inferred cell and replace it with the real value. This design exists because inference is the AI's biggest strength and its biggest risk: it makes drafts instant, and it makes silent wrongness possible. Surfacing the guesses is how both benefits are kept.
What deterministic checks handle
The AI drafts; the engine enforces. Structural things never depend on the model:
- Price calculations — percentage price changes are computed in code, not estimated by the AI. For example, 15% off ₹499 is calculated as ₹424.15 before any configured rounding.
- CSV structure — column names, quoting and the Hyperzod option-group format are generated deterministically and then shown for review.
- Validation — duplicate SKUs, missing required fields and broken variants are flagged by the validator, not by the model.
- Data handling — uploaded CSV cells are passed through the catalog workflow as product data. Review the generated result before export, especially when source values contain unusual instructions or markup.
The review workflow
- Generate or convert the catalog.
- Scan the preview for highlighted (inferred) cells.
- Fix anything wrong by clicking or by chatting ("make the pizzas ₹100 cheaper").
- Check the validation count and export.
Three judgments AI should not make alone
- Truth. Whether a fact is real — materials, certifications, origin claims. If the source data does not say it, do not let the draft say it.
- Pricing strategy. What a discount should be; the tool computes it exactly, but you choose it.
- Taxonomy decisions. Which category a product belongs to, and when the category tree itself needs restructuring.
AI makes the first draft cheap; the preview supports a safer review. Use both — generate quickly, review carefully, and test a small import before replacing a live catalog.