The problem

A supplier sends a spreadsheet of products. You need them in Odoo.

What is in the file: product names, maybe a code, maybe dimensions, a price.

What Odoo needs: a name, an internal reference, a category, a unit of measure, a cost, a sales price — and for anything customer-facing, a description worth reading.

The gap is the problem. A spreadsheet with 400 rows becomes 400 records, each needing fields the file does not contain and a description somebody has to write.

Done by hand, that is weeks.

Where AI helps and where it does not

Be specific, because this is what decides whether the workflow is safe.

AI helps with:

Descriptions. Turning attributes into readable text, in a consistent tone, at volume.

Category suggestion. Reading a product name and proposing which of your categories it belongs in.

Reference generation. Producing a consistent internal reference from the name and category.

AI must not decide:

Prices. Ever. Cost and sales price come from the file or from your pricing rules.

Units of measure. A guess here produces wrong stock and wrong costs, silently, forever.

Specifications. If the file does not say the bracket is stainless steel, the description must not say so either.

That last one is the risk that matters. A model given gaps fills them, plausibly and confidently. Invented specifications reach customers, and in a product catalogue that is a real problem.

The workflow

Read the file. Excel or CSV into structured rows.

Validate each row. Does it have a name? A price? Anything required?

Check whether the product already exists. By supplier reference or SKU. Skip or update — do not create a second one.

Generate the missing fields. Description, category suggestion, reference.

Create the product — as a draft or unpublished, not live.

Report what happened. Created, skipped, failed, and why.

Give the model your real data

The single biggest factor in output quality.

Supply the actual attributes from the row — dimensions, material, colour, size, whatever the file contains.

Supply your category list. A model choosing from categories you gave it produces usable values. A model inventing categories makes a mess of your product configuration that somebody has to unpick.

Supply examples. Three descriptions you like produces far better output than a paragraph describing what you want. This is the highest-return technique available.

Say what it must not claim. Certifications, compliance, compatibility, materials not stated in the file. An explicit instruction works, and it prevents the most damaging category of error.

Create as draft

Not live. Not published. Not sellable yet.

Why: somebody has to check the descriptions before customers see them, and the volume that makes this workflow worthwhile is the same volume that makes one systematic error expensive.

A workable arrangement:

The workflow creates products unpublished, in a state that means “generated, not reviewed”.

Somebody reviews a batch — reading for truth, not for grammar.

Products are published in groups as they are checked.

The saving is still large. Reviewing 400 descriptions is a day. Writing them is weeks.

Review for truth, not style

Worth stating plainly, because it changes who should do the reviewing.

The grammar will be fine. Generated text is fluent.

The question is whether the claims are correct — and answering that needs somebody who knows the products, not somebody who can read.

What to check specifically:

• Dimensions and materials against the source file
• Any certification or compliance claim — these should not appear at all
• Compatibility statements
• Anything numeric

Generate ten before you generate four hundred

The habit that prevents most expensive mistakes.

Run the workflow on ten rows. Review those ten properly. Find the systematic error — and there usually is one, in the first attempt.

Adjust the instructions. Run ten more.

Then scale.

Generating everything at once means discovering the systematic error after it is everywhere, in four hundred products, some of which are already published.

Duplicates

Same rule as every import.

Check before creating, on a stable identifier — the supplier’s product code, a SKU, a manufacturer reference.

Never on the product name. Supplier names get reformatted between files, and matching on them creates duplicates every time the format changes.

And decide: if the product exists, update it or skip it? Updating keeps prices current. Skipping avoids overwriting descriptions somebody has already reviewed.

Usually: update the price, leave the description alone.

Units of measure

Worth its own warning.

Odoo blocks changing a product’s unit of measure once it has stock moves or order lines.

Which means a wrong unit at creation is effectively permanent. Fixing it means creating a new product and archiving the old one.

Never let a model guess this. Take it from the file, or set a default and have it reviewed. If the file does not specify, that is a question for the supplier, not for a model.

The short version

AI turns a supplier spreadsheet into a usable catalogue by writing descriptions and suggesting categories — the parts that would otherwise take weeks.

It must not touch prices, units of measure, or specifications the file does not contain.

Give it real attributes, your category list, and examples. Create as draft. Review for truth, with somebody who knows the products.

And generate ten before four hundred. That one habit catches the systematic error while it is still cheap.