Filling product passports with AI: what it does well, and what it must never invent

The walkthrough in creating your passports in twenty minutes follows the manual path: store defaults, composition per category, and a full catalogue without ever leaving WooCommerce. It is free, uncapped, and enough for plenty of stores.

The other path has an AI read your product pages and propose the missing values. It saves real time on a varied catalogue. Above all it raises a question the tools carefully avoid asking: on data that ends up in a regulatory document, what do you let a machine write?

Here is DPTag's answer, and what follows from it.

The rule that governs everything: the AI proposes, it does not write

An AI proposal is never saved into the passport. It is stored separately, in a pending state, waiting for your review. The form shows you every proposed value, and nothing enters the passport until you save.

The direct consequence: an existing value is never overwritten. Not one you typed, and not one coming from your defaults. The AI only works on empty fields.

This is not a courtesy, it is the only defensible position. A passport commits the brand to facts. A machine that silently replaces a verified value with a plausible one manufactures a false document that nobody rereads.

What it sees, and what never leaves your store

This is the question we get most, and the answer is a short list. The text sent contains the product title, its short description, its long description, its categories and its attributes.

Nothing else. In particular no post metadata from your database, no customer data, no orders, no prices. It is exactly the text any visitor already reads on your public product page.

That limit has a consequence worth knowing before you try. If your composition lives in a custom field, put there by a custom-fields plugin, the AI cannot see it. It will find nothing to extract and tell you to fill in by hand, while the value sits in your database. An extension point exists to open those fields to it, but it takes a line of code: the one case in this article that needs a developer.

The call never fires on its own. You have to click.

The hierarchy: your values win

The order is the same as everywhere else in the plugin, and the AI comes last.

A product attribute beats a category value, which beats a store value, which beats an AI proposal. The message on the card says it plainly: the AI only fills empty fields, and anything from your defaults or your own typing is kept.

One exception exists, and it is off by default. A setting lets the AI revisit your store-wide defaults, the ones you set once for the whole catalogue and that age badly on unusual items. Even then, category and product values stay untouchable, and so does anything you typed.

In other words, you can let the AI argue with the general. Never with the particular.

An example, to see what it actually does

Take a description of the kind people really write, for a sweatshirt:

Our winter sweatshirt in brushed fleece, 85 percent certified organic cotton and 15 percent recycled polyester. Knitted and made up in Portugal, dyed in Italy. Wash at 30 degrees, dry flat.

Here is what the automatic read pulls out of it, field by field:

Two rows deserve attention, and they are the empty ones. "Certified" is not a certification: no scheme is named, so there is nothing to record. And the batch number exists only in your own records.

That is exactly the behaviour to want. A tool that had filled those two rows would have produced a more complete passport and a less true one.

Confidence and source: reviewing fast, and well

Every proposal arrives with two things most tools do not give you.

A confidence score, saying how sure the model is. Use it to order your review: start at the bottom, not at the top of the list.

A source, saying where the value was read from. That is what makes a proposal checkable at a glance: when the source says the product description, you know where to look to confirm it.

That pair is what turns a painful review into a few minutes of sorting. Without it, approving three hundred proposals means trusting with no way to verify, which is the worst of both worlds.

What it does well, and what it must never invent

The AI is good at one precise job: reading prose and pulling structure out of it. A description saying "this sweatshirt in organic cotton fleece, with 15 percent recycled polyester" becomes two composition lines with their percentages. Marketing copy mentioning dyeing in Italy becomes a field. It normalises, it files, it does not create.

Three kinds of data must never be handed to it, because they are written nowhere in your store.

The country of manufacture, when the text does not state it. A machine inferring an origin from a brand name or a product style manufactures a false regulatory statement. It is also the one that exposes you most.

Certifications. A label is a document, not an impression. If the certificate is not mentioned, the correct answer is an empty field.

Batch numbers and serial identifiers. They exist only in your own records.

The general rule fits in one sentence: an AI can extract a fact you have already written, it cannot know a fact you have never written. An empty field is an honest answer, and it can be filled later.

Reviewing a batch without losing a day

Three habits are enough.

Sort by confidence, from the bottom up. Confident proposals are confident because the value was written in plain sight. It is the hesitant ones that deserve your eyes.

Read the source before the value. A composition read from an attribute is confirmed at a glance. The same value read from marketing copy means going to look at the text.

Work through one category at a time. If the first ten sweatshirts are right, the rest probably are too, since they come from descriptions written by the same hand. Conversely, two errors in one category point at a formatting problem in your descriptions, not at a problem with the model.

And if the plugin tells you there is nothing to add because everything is already filled by you or your defaults, that is good news rather than a failure: you have just saved credits.

What it costs, and why no subscription is required

Every site gets free samples, with no account and no card, so you can see what it does on your real products before deciding anything.

Beyond that it runs on credits, which are bought on their own. No subscription is needed to use the AI, and credits you have bought do not evaporate if you later stop a subscription. That is deliberate: billing usage rather than access avoids selling a subscription to somebody who fills their catalogue once a year.

The difference between the free plugin and the Pro add-on is not the quality of the filling, it is how a run behaves. In the free plugin the batch runs in your browser, one passport after another, and the tab has to stay open. On a few dozen products that is fine. On a catalogue of several hundred, one network hiccup stops the run. The Pro add-on moves the work server-side, into a queue that retries by itself: you click once, close the laptop, and come back.

When the AI is not worth it

This is not the answer a vendor is supposed to give, but it is true.

If your catalogue is homogeneous, the AI will buy you almost nothing. Three hundred references across eight material categories are filled by the category composition, in about ten minutes, free, with no review to do. Paying credits to rediscover what you already know makes no sense.

The AI pays off in the opposite case: a varied catalogue where each product has its own material, and where the information exists but lives as prose in descriptions written over the years. That is where reading by hand costs hours and automatic extraction genuinely saves them.

The test is simple. Count your material categories. If that number is small against your number of references, take the manual path. Otherwise try the AI on the free samples, and look at the confidence scores before deciding.

For the rest of the path, see creating your passports in twenty minutes, and for which data actually matters, the textile passport data.