EU AI Act Chatbot Disclosure: A Practical Guide for Online Stores

If an AI bot answers shoppers on your Shopify or WooCommerce store, start with a question you can test: can a new visitor tell they are interacting with AI before the bot replies? Article 50 of the EU AI Act raises a separate legal question about when that information is required. Your chat design determines whether shoppers actually see it.

What Article 50 says about AI chatbot disclosure

Article 50 includes a transparency requirement concerning AI systems that interact directly with people. The provision addresses informing people that they are interacting with an AI system, subject to its exact wording and exceptions. For a store, the first question is whether AI communicates with a shopper directly. The next is what the shopper sees before and during that interaction.

Do not decide from the feature name alone. A button labelled “Chat with us” might open human live chat, an AI conversation or a combination. Map the path a visitor takes: what appears on the closed launcher, what appears when the panel opens, who sends the first reply and what changes when an agent joins. Give that map to your legal adviser rather than asking them to assess “our chatbot” in the abstract.

An AI identity notice and a GDPR privacy notice do different jobs. “You’re chatting with an AI assistant” identifies the participant. Privacy information explains the handling of personal data collected through chat. Whether a particular use of chat data calls for consent is another question for your privacy review. A visible AI label does not settle it, and a link to your privacy notice does not necessarily tell a shopper who is answering.

Does your shop chat need an AI disclosure?

List every way a message can reach a shopper, including greetings, order answers and handoff messages. Then use this table to decide what to inspect. It is a triage tool, not a ruling that a setup is compliant or exempt.

Chat setup	What reaches the shopper	Practical approach	Check with your adviser
Customer-facing AI bot	AI sends answers without an agent approving each message.	Identify the AI in the entry experience and before its first substantive reply. Label its messages clearly.	How does Article 50 apply to this direct interaction, including its exact wording and exceptions?
Agent-only live chat	A person writes and sends the replies.	Describe it as team chat rather than AI chat.	Are any greetings or replies actually generated and sent by AI?
AI drafts reviewed by an agent	An agent sees a private suggestion, checks or edits it, and chooses to send the reply.	Describe the customer-facing service accurately. Inspect other automated messages separately.	Does any suggestion bypass agent review and reach the shopper automatically?
AI first, human later	AI sends earlier replies; a person takes over after handoff.	Identify the AI phase, announce the transition and preserve the origin of earlier messages.	Can the shopper tell who sent each message, including after reopening the chat?

The sending action matters. If a system places a draft in an agent’s reply field and the agent decides whether to send it, that differs from a bot publishing the same text automatically. Likewise, a fixed menu, an automated greeting under an agent’s name or a conversation that alternates AI and human replies may need closer review. Do not infer an exemption from a familiar-looking interface.

In a mixed conversation, avoid a handoff banner that makes the whole transcript look human-written. If the shopper reopens the chat, earlier AI messages should still be distinguishable from the agent’s reply. Record a screenshot of the reopened transcript: it is an easy state to miss during a review focused only on the first visit.

Where should the notice appear in a Shopify or WooCommerce widget?

Follow the shopper’s first uninterrupted journey: they notice the launcher, open the panel, type a question and receive an answer. Aim to show the AI identity before they start that exchange. A launcher label can provide early notice where space allows; an immediately visible welcome message can explain it in full. If your widget cannot show a notice until after the shopper sends a message, flag that timing for legal and design review rather than assuming a later label fixes it.

The following are illustrative widget mockups, not askVela screenshots or examples of a verified product configuration.

The wording is visible before the first AI answer. It still needs testing: on a phone, the keyboard might cover the welcome message, and a returning visitor might land directly in an existing transcript.

Here the shopper would have to look for a settings menu to discover that AI is replying. A privacy-policy link elsewhere in the panel presents the same practical problem if it is the only place that identifies the bot.

Do not show the agent’s name on the waiting state before the agent actually joins. Equally, do not remove the AI label from earlier messages when the conversation changes hands.

Disclosure wording you can adapt

Choose wording that matches the controls the shopper actually has. “AI assistant” is plainer than a product nickname. If the bot cannot check orders, do not say it can. If a human is available only later, avoid “Talk to us now.”

Replace the brackets only after checking the published widget. For example, write “Select ‘Speak to a person’” only if that button exists in the state where the message appears. If requesting a person instead starts an email follow-up, describe that route and when the shopper should expect to hear back according to your actual process. Review the privacy wording for any email address or order details you collect separately.

Here is a sample order conversation. It deliberately avoids claiming that every store bot has order lookup:

If your bot looks up orders, replace the conditional answer with the exact request it makes, such as an order number and email address. If it cannot look up orders, remove the lookup sentence altogether. The same rule applies to handoff: do not promise that the bot will stop answering or that an agent will appear immediately unless you have tested those behaviours.

How to test the complete chat journey

Test on the published storefront, not only in a widget editor. Start without an existing conversation so you see what a new visitor sees. Ask someone who did not configure the widget to run the test; give them the steps, but do not point out where the notice is supposed to be.

1. Record the starting state. Note the storefront page, device, browser language and selected storefront or widget language. Capture the closed launcher. Write down whether it identifies AI.
2. Open without typing. Capture the panel immediately. Can the tester identify the AI without scrolling, opening settings or following a link? On mobile, repeat with the keyboard open and note if it covers the notice.
3. Send an order or delivery question. Capture the first AI response and its speaker label. Was the AI identity visible before the question was sent and before the reply appeared? If not, record precisely when it first appeared.
4. Navigate and reopen. Move from a product page to another page where the widget appears, then reopen it. Check the notice, transcript labels and whether an earlier AI answer now appears to have come from a person.
5. Request a person. Use the exact button or phrase advertised in the welcome message. Capture the acknowledgement, the waiting state and the agent’s first reply if an agent joins. If no agent joins during the test, verify that the waiting text does not promise an immediate response.
6. Repeat in each enabled language. Read the live launcher, welcome message, first reply and handoff text. A translated heading does not establish that all chat states use the same language or preserve the meaning of the AI notice.
7. Save a review record. Keep the exact wording and dated screenshots. Log each issue as “page and state; device; language; expected wording; observed wording” so someone else can reproduce it.

On Shopify, test the published theme on pages where chat is actually shown. On WooCommerce, test the published WordPress storefront, including product and checkout-adjacent pages if the widget appears there. Repeat affected tests after a theme change, translation update, new bot flow or widget redesign. A screenshot from an earlier configuration does not show what visitors see today.

What to check with your provider and legal adviser

Ask your provider what you can configure and what customers actually see. Ask your legal adviser whether that experience meets the applicable requirements. Neither a feature list nor a screenshot alone answers both questions.

If you use askVela, its AI bot and askVela’s live-chat widget give you a customer-facing journey to inspect. Its bot rules and test console let you preview bot behaviour without a real chat, and its bot can hand a conversation to the team. You write the widget’s greeting and welcome message yourself, so the AI notice can be the first thing a shopper reads. Then check the wording and visible states on your own storefront.

Keep the privacy review separate. askVela describes encrypted transmission and security and privacy controls; those do not replace your assessment of the information you give shoppers about personal data or whether consent is relevant to your setup.

Your useful output is a reviewable record: the reply paths, the exact notice, screenshots from a new visitor’s journey and a handoff, language checks, and your adviser’s assessment of the applicable law. That gives you something specific to retest when the storefront or chat behaviour changes.