With 3,000 SKUs across four markets, AI isn't a nice-to-have: it's the only way to keep listings alive without a team of ten. The trouble is that most people start in the wrong place and end up with 3,000 badly translated listings and a stack of catalogue quality alerts.
This comes down to separating three things: what you can automate outright, what needs a human before it goes live, and what you shouldn't touch at all.
Where it genuinely works
These are the cases where the return shows up fast and the risk is low.
First drafts of listings. AI doesn't write your final listing, it writes the draft your team corrects. Going from a blank page to a decent draft saves about 70% of the time. The trick is feeding it real product data rather than letting it fill gaps: if you don't supply the material, it invents it.
Translation and adaptation for other markets. AI is clearly better here than old-school machine translation, because it can adapt register, units and local search language. Translating your Spanish title literally into German is the surest way to appear in no German searches at all. Properly instructed, AI adapts instead of translating.
Attribute normalisation. You've got "50ml", "50 ML", "0.05 l" and "50 millilitres" in the same field. This is the perfect automation job: clear rules, verifiable output, high volume. Same with colours, materials, sizes and units of measure.
Product data enrichment. Filling in fields the supplier left blank, starting from the spec sheet, the manual or the manufacturer's PDF. Extracting data from a document you hold is safe. Deducing it from thin air is not.
Classifying reviews and tickets. Huge value, almost no risk, because nothing gets published. You classify 4,000 reviews by complaint type and discover 30% of returns come from one photo that misleads on size. That's money found in half an hour.
Price and stock anomaly detection. A daily process that flags a price that has doubled, an active product that has been at zero for three weeks, or a SKU whose margin has dropped below your threshold. It doesn't decide, it alerts. Perfect for automation.
Where AI alone is dangerous
There's a simple rule: if a mistake can end in a complaint, a fine or a returned product, it doesn't get published without human review.
Regulated claims. Anything that sounds like a health, cosmetic or food property is regulated. "Boosts your immune system", "reduces wrinkles", "suitable for coeliacs". AI writes these sentences with total confidence because it has seen them a thousand times online. Having seen them doesn't mean you're allowed to use them.
Technical data. Power, voltage, capacity, composition, dimensions. If the AI doesn't have the figure, there's a real chance it fills in something plausible. A plausible, false spec on a listing is a returned product.
Compatibility lists. "Compatible with models X, Y and Z" is the most dangerous case of all. A compatibility list is exactly the kind of content an AI completes by pattern. Every invented compatibility is a guaranteed return and a one-star review.
Safety, warnings and regulatory text. Minimum age, pictograms, usage warnings, certifications, CE marking. This gets copied from the official document. It isn't generated, isn't rewritten "to read better" and isn't loosely translated.
Brand names and third parties. Mentioning someone else's brand in your listing can trigger an intellectual property complaint. AI has no legal judgement.
The decision table
| Task | Level | Why |
|---|---|---|
| Normalising units and attributes | Fully automate | Clear rules, verifiable output, no content risk |
| Classifying reviews and tickets | Fully automate | Nothing customer-facing gets published |
| Price and stock anomaly detection | Fully automate | It only raises alerts, the decision stays human |
| Draft titles and bullets | With review | Drives ranking and conversion |
| Translation for other markets | With review | Register errors and wrong local search terms |
| Extracting attributes from a spec sheet | With review | Reliable when the document exists, validate by sampling |
| Long description and brand content | With review | Brand voice and risk of unverified claims |
| Health, cosmetic or food claims | Don't automate | Regulated, with real legal exposure |
| Compatibility and cross-reference lists | Don't automate | High invention rate, straight to returns |
| Safety warnings and regulatory text | Don't automate | Must be copied verbatim from the official document |
| Technical data with no source document | Don't automate | No figure means no figure: leave it blank |
Human review is part of the process, not an extra
The most common build mistake is treating review as something you'll do "if there's time". There's never time.
Put it in the workflow as a mandatory state. A generated listing isn't in "ready" state, it's in "pending validation". And nobody can publish from that state without passing through validation.
That doesn't mean reviewing 100% of everything. It means setting the level by risk:
- Full review: regulated categories, new products, high-rotation or high-price items.
- Sample review: low-risk content at volume. Check 10-20% and, if the batch error rate exceeds your threshold, send the whole batch back.
- Automatic validation only: normalisation and formatting, with rules that fail visibly.
One practical note: reviewing is far faster than writing, but only if the person knows what to look for. Give your team a short list of five checks, not a twenty-page manual.
How to build the pipeline
The architecture that works always has the same pieces:
- A source of truth. One place where the good product data lives: a PIM, a master spreadsheet, whatever. If it doesn't exist, build that first. AI doesn't fix a catalogue with no source.
- Generation with limited context. The AI can only use the data you pass from that source, plus your brand rules. Explicit instruction: if a value is missing, leave it blank and flag it. An empty field is a small problem; an invented field is a big one.
- Automatic validation before human validation. This catches an enormous number of errors for free: title length per channel, mandatory fields, banned words, out-of-range units, prices outside a band, attributes not allowed in the category. Whatever fails the validator never reaches a human.
- A review queue. With the diff visible: what was there, what the AI proposes. Reviewing by comparison is five times faster than reading from scratch.
- Publishing in small batches. Never 3,000 listings at once. Batches of 50 to 200, with a check after each one.
- A change log. What changed, when, with which prompt version, approved by whom. Without it you can't roll back and you can't learn.
If you can't undo it in ten minutes, don't publish the batch.
Mass publishing without review: what happens next
Marketplaces run catalogue quality systems. They don't detect "an AI wrote this", they detect the consequences: listings with inconsistent data, attributes that don't match the category, duplicate content across variants, copy that doesn't match the product.
What tends to follow, in order of severity: listing quality alerts, suppression from search results, publishing blocked until you fix it, and in the ugly cases, an account review. Add the direct commercial damage: returns for inaccurate description, one-star reviews and dented seller performance metrics.
And there's a cost nobody budgets for: rolling back 3,000 bad listings costs more than doing them properly. If you didn't keep the previous version, nobody is rolling anything back.
Recovering the ranking of a listing that sank after a badly executed bulk change takes weeks. The specific policies on generated content and catalogue quality are updated often, so confirm the current wording in Seller Central before launching a big project.
Measuring whether any of it worked
Automating without measuring is swapping work for noise. Define your metrics before you start and keep them few:
- Catalogue coverage: percentage of listings with every mandatory field complete. Easiest to move and usually the most valuable.
- Batch error rate: percentage of listings rejected at review. If it falls over time, your instructions are improving. If it doesn't, the problem is in your source data.
- Hours per listing: before and after, review included. If you don't count the review, you're kidding yourself.
- Conversion and traffic on edited listings, against a control group you leave untouched. Without a control group you can't tell your improvement from the season's.
- Returns for inaccurate description: the metric that tells you something bad is going live.
Always keep a control group. It's the only way to defend the project with data instead of impressions.
What to do this week
- List the catalogue tasks that eat the most hours and sort them with the table above: automate, review, or leave alone.
- Find your source of truth for product data. If you don't have one, start there rather than with AI.
- Write the automatic validator before the generator: lengths, mandatory fields, banned words and ranges.
- Pick a small batch of 50 low-risk listings and run it through the whole pipeline, review included, timing it properly.
- Define the control group and the two or three metrics you'll look at in a month's time.