This is a worked example of EU AI Act compliance on a realistic small business website. The business is fictional, but the setup is typical: a WordPress site for a garden furniture shop in Germany that sells across the EU, with a chatbot, AI-written content and AI images. We go through it the way you would on your own site, step by step, and show what changed and why.
The site
GardenLine (fictional) has six employees and runs a WooCommerce shop with about 400 products and a blog. Over the past year the team has started using AI in several places:
- A support chatbot widget called “Sophie” with a smiling stock photo.
- Product descriptions drafted with an AI writing plugin.
- Lifestyle images for product pages created with an image generator.
- A blog with care guides, plus a few articles on “new EU rules for outdoor heaters” and “tax deductions for home office furniture”, written with ChatGPT and published without review.
- “Customers also bought” recommendations from a WooCommerce extension.
- An AI SEO plugin that writes meta descriptions.
Step 1: the audit
The marketing manager is named owner and spends an hour on an inventory, following our website AI audit guide. She checks the plugin list, the tag manager and asks colleagues which tools they use. Result: the six items above, plus one she didn’t expect: the translation plugin uses machine translation for the French and Italian versions.
Step 2: classification
| Item | Role | Rule | Decision |
|---|---|---|---|
| Chatbot “Sophie” | Deployer | Art. 50(1) | Must say it’s an AI |
| Product descriptions | Deployer | None under AI Act (not public interest) | No label; fact-check claims |
| Lifestyle images | Deployer | Art. 50(4) only if realistic and misleading | Review each image |
| Care guides (how to oil teak) | Deployer | Not public interest | No label required |
| Articles on EU rules and tax | Deployer | Art. 50(4) public-interest text | Review or label |
| Recommendations | Deployer | None under AI Act | No action |
| AI meta descriptions | Deployer | None | No action |
| Machine translation | Deployer | Treated like editing support; public-interest articles follow the original’s decision | Translate reviewed versions only |
Step 3: fixes
Chatbot
The greeting changes from “Hi, I’m Sophie! How can I help?” to “Hi, I’m Sophie, GardenLine’s AI assistant. I can help with orders, delivery and care tips. Type ‘team’ to reach a person.” The widget header now says “AI assistant”. The team also restricts the bot to the shop’s real return and delivery policies after it was caught promising free returns on sale items, which the shop doesn’t offer. See chatbot compliance.
Images
Most lifestyle images put real product photos into generated garden scenes. That’s fine. Three images are different: one shows a lounge set in a colour that isn’t sold, and two show a “customer’s terrace” that never existed. The first is replaced, because it misleads about the product. The other two get a visible “AI-generated image” badge and caption text.
Public-interest articles
The heater article had an error about which models are affected. The shop’s product safety contact corrects it and signs off with name and date, and the shop takes editorial responsibility, so no AI label is needed. The tax article is labelled instead, because nobody in-house is qualified to review it: “This article was written with AI and has not been reviewed by a tax adviser.” The shop plans to have an accountant review it later.
Product descriptions
No label, but a spot check finds two descriptions claiming “FSC certified” for products that aren’t. The team adds a rule: certifications and dimensions are always copied from supplier data, never from AI.
Privacy notice
The chatbot vendor wasn’t listed. It’s added, with purpose, retention period and a note that conversations aren’t used for training (switched off in the vendor settings).
Step 4: policy and training
A one-page AI guide for staff: approved tools, no customer data in public AI tools, check every fact in AI drafts, and the review-or-label rule for public-interest topics. The team gets a 30-minute walkthrough. That covers Article 4 AI literacy in a way proportionate to a six-person shop.
Step 5: documentation
All decisions go into one table with dates and evidence: screenshots of the chatbot, the reviewer’s sign-off note, the list of labelled images. The owner sets a calendar reminder for a quarterly review.
What it took
- About one working day in total, spread over two weeks.
- Three changes visitors can see: the chatbot greeting, two image labels, one article label.
- Several fixes that had nothing to do with labels but would have caused trouble anyway: the chatbot’s invented returns policy, the wrong certification claims, the outdated heater rules.
That last point is typical. A compliance pass mostly surfaces quality problems in AI output. The labels themselves are the easy part.
How it maps to the plugin
On WordPress, the free EU AI Act Ready plugin covers most of GardenLine’s visible steps: the tax article gets the disclosure level “AI-generated” and shows a notice; the reviewed heater article is set to “AI-generated, human reviewed” and shows none; the two images get AI badges; the chat widget gets an AI notice; and the export gives the documentation table. Everything runs on the shop’s own server.
Apply it to your own site
- Name an owner and run the audit.
- Classify each item using the transparency rules for websites.
- Fix chatbots first, then images, then public-interest text.
- Write a short policy and keep the record. Our compliance strategy guide has a 90-day plan.
Frequently asked questions
Would GardenLine face a fine before the fixes?
It was in breach of Article 50 for the chatbot and possibly the images and tax article. For an SME the cap is the lower of 15 million euros and 3% of turnover, and fines must be proportionate. A quick fix and a record make enforcement far less likely. See fines and penalties.
Did the care guides need a label?
No. How to oil teak furniture isn’t a matter of public interest in the Article 50 sense. The shop could label them voluntarily.
Is the recommendation engine high-risk?
No. Product recommendations aren’t in the high-risk list. See high-risk AI explained.