EU AI Act Transparency Rules: How to Design AI Lead Forms Before 2 August

On 20 July 2026, the European Commission published final guidance on the EU AI Act’s Article 50 transparency obligations. The relevant obligations begin applying on 2 August 2026.

For marketing, sales, and RevOps teams, this is not mainly a footer-text exercise. If an AI assistant qualifies a prospect, asks follow-up questions, recommends a route, schedules a meeting, or triggers a follow-up journey, transparency needs to appear in the flow where the person experiences the AI.

The practical job is to make four things clear: that AI is involved, which organization it represents, what the interaction is for, and where a person takes responsibility for what happens next. This article is an operational framework, not legal advice. Have counsel assess your specific system architecture, organizational role, markets, data processing, and vendor arrangements.

What changed on 20 July, and why 2 August matters

The Commission’s final guidelines cover transparency obligations for providers and deployers of certain AI systems. The Commission describes the guidance as support for consistent, effective, proportionate, and uniform application. The guidelines are interpretive guidance. Article 50 itself is the binding legal provision.

Do not turn the date into a broader claim than it supports. The EU AI Act entered into force on 1 August 2024, and its requirements have different application timelines. The immediate 2 August 2026 milestone concerns the relevant Article 50 transparency obligations, not every AI Act requirement.

For direct interactions, Article 50(1) requires providers of AI systems intended to interact directly with natural persons to design and develop them so people are informed they are interacting with AI, unless that is obvious from the circumstances and context. The Commission’s announcement also addresses other categories, including generative-AI output marking, notices for emotion-recognition and biometric-categorization systems, deepfakes, and certain AI-generated or manipulated text published on matters of public interest.

That range matters because “AI lead form” can describe very different systems. A form drafted internally with AI is different from an AI agent that interviews a visitor. A video response field is different from a system that infers emotion from video. Treat the actual interaction and system behavior as the unit of review.

Provider versus deployer: map the role before choosing a control

Start with a role map, not a generic policy statement. In simplified operational terms, a provider is closer to the party placing an AI system on the market or putting it into service under its name. A deployer uses an AI system under its authority. Those labels are not a shortcut to a conclusion.

A company may be a provider, deployer, or both depending on its product architecture, branding, configuration, contractual allocation, and market placement. Bird & Bird’s analysis of the final guidance highlights why teams should not assume the model vendor bears every relevant obligation.

Create a one-page record for each prospect-facing AI workflow. Name the system, vendor, branded interface, purpose, inputs, outputs, automated actions, human owners, and markets where it runs. Then validate the role allocation with counsel and the vendor. This makes the next design decisions much easier.

The AI form and funnel use cases that deserve an audit

Begin with systems that a prospect can experience directly. An AI agent that asks “What are you trying to achieve?” and adapts its next question is a more obvious audit candidate than a conventional form that simply sends submissions to a shared inbox.

Also audit the invisible layer when it changes the prospect’s outcome. Examples include an AI-generated lead score that routes someone away from sales, a booking recommendation, an AI-written follow-up that sends automatically, or an internal summary that becomes the only record a salesperson reads.

Use a simple test: could an AI-generated statement, question, classification, or recommendation materially change what the visitor sees or what the team does next? If yes, make the workflow inspectable and assign an accountable owner. This is sound operational practice even where your legal assessment finds that a particular interaction is outside Article 50.

Independent legal analyses report that in-scope AI agents should disclose both their artificial nature and the party on whose behalf they act. They also report that a generic assistant reference, terms-and-conditions notice, platform-level disclosure, or machine-readable marker alone may not be enough for the direct-interaction disclosure. Treat these as legal interpretations to validate against the final guidance and your use case, not as a substitute for legal advice.

A disclosure pattern that fits a high-converting lead flow

A useful disclosure does not need to be a long legal notice. It needs to be understandable at the moment the person decides whether to continue. Put it on the start page or immediately before an AI conversation begins, in the same visual context as the primary action.

Example: B2B demo qualification flow

1. Start page: “You’ll chat with Acme’s AI qualification assistant. It will ask a few questions to help route your request. A member of our sales team reviews qualified requests.”
2. Intent: Ask what the visitor wants to accomplish. Use their declared answer to choose relevant follow-up questions.
3. Company context: Ask for the facts needed to route responsibly, such as company name, team size, use case, and timeline. Do not ask for data merely because it may be useful later.
4. Contact details and permissions: Ask for the details needed to respond. Keep any marketing-consent choice distinct from the request to contact the person about their inquiry.
5. Routing outcome: Say what will happen next: for example, “Based on your answers, we’ll recommend a conversation with our sales team” or “We’ll send the relevant implementation resources.” Avoid language that makes an opaque system outcome sound like a final human decision.
6. Ending page: Confirm the next step, response expectation where you can support it, and a human alternative such as a contact route for questions or correction.

This pattern gives the visitor context before disclosure-sensitive interaction, while keeping the form focused on their goal. It also avoids an easy mistake: treating AI transparency as a replacement for privacy notices, consent controls, or other applicable obligations. It is not.

For the underlying qualification design, see How to Build a Lead Capture Funnel That Qualifies Visitors Before They Book. The key addition here is visibility into the AI and decision boundary, not more questions.

What Article 50 covers, and what it does not

Article 50 is broader than direct AI interaction, but the categories should not be collapsed into one generic “AI label” requirement.

• Direct interaction: The Article 50(1) rule concerns AI systems intended to interact directly with natural persons, subject to the stated obviousness exception.
• Generative AI output marking: Article 50(2) concerns providers of generative AI systems that create synthetic audio, image, video, or text, including machine-readable marking and detectability, subject to technical feasibility and state-of-the-art limitations.
• Emotion recognition and biometric categorization: Article 50(3) concerns deployers of these systems. A respondent recording a video or voice answer is not, by itself, emotion recognition or biometric categorization.
• Deepfakes and certain public-interest text: Article 50(4) addresses deployers creating or manipulating deepfake image, audio, or video content and certain AI-generated or manipulated text published to inform the public on matters of public interest.

Three boundaries prevent overcorrection. First, do not label every internal AI-written draft question as public AI-generated content. Second, do not treat ordinary one-to-one sales outreach as public-interest text. Third, do not assume every online form is in scope simply because it is online. The Commission’s Article 50 overview is the right starting point for scope, followed by advice on the specific system and role.

Five funnel mistakes to avoid

1. Giving an AI a human-like identity without a clear disclosure

A friendly name and avatar can make a flow approachable, but should not obscure that the visitor is interacting with AI. Put the disclosure near the interaction, not in a separate policy page.

2. Relying on a footer-only notice

A footer is easy to miss and detached from the interaction. Use a concise, plain-language statement where the AI starts asking questions, and make it readable on mobile.

3. Letting an AI summary replace the original response

Summaries help sales teams move faster, but they can omit nuance or introduce interpretation. Preserve the prospect’s original answers and clearly distinguish declared data from generated internal analysis.

4. Sending people through unexplained rejection routes

If automation sends one visitor to a sales calendar and another to self-serve resources, make the route understandable. Avoid claiming a system has “approved” or “declined” someone when the actual workflow is a qualification recommendation.

5. Treating a vendor disclosure as the whole solution

Vendor documentation matters, but it does not design your customer-facing experience, define your routing ownership, or preserve your operational evidence. Ask vendors about role allocation, disclosure support, data use, retention, marking responsibilities, and available documentation.

A seven-point readiness checklist for form and RevOps teams

1. Inventory every prospect-facing AI touchpoint. Include chat, question generation, scoring, routing, booking recommendations, summaries, and follow-up.
2. Map the role and system boundary. Record what your team configures, brands, deploys, and controls. Review this with counsel and relevant vendors.
3. Write the interaction disclosure. State that the visitor is interacting with AI, identify the organization represented, and explain the purpose in plain language.
4. Review each question and route. Check that questions are necessary, branches are understandable, and automated outcomes do not overstate certainty or human review.
5. Separate source facts from generated interpretation. Retain original answers, campaign context, and submission timing. Label any generated summary as internal interpretation.
6. Set a human handoff and owner. Define when a person reviews a submission, who can override a route, and how visitors can ask a question or correct relevant information.
7. Test and document the workflow. Test disclosures, routes, automations, and handoffs. Keep versioned evidence of the live flow and vendor information available for review.

Measure after release. Track starts, completions, abandonment by step, route selection, human overrides, response quality, and downstream qualification. A transparent flow may require copy refinement, but speculation is a poor substitute for funnel data. For a practical approach to partial submissions and friction analysis, read How to Reduce Form Abandonment and Recover Partial Leads.

Build the operational layer in Stepform

Legal assessment belongs with your counsel. The operational challenge is making the approved design real, reviewable, and measurable. Stepform can support that work without determining whether your workflow is compliant.

Use AI-assisted drafting to create an initial intake flow or refine disclosure language, then have a human review every question, branch, field mapping, and follow-up action before publishing. Keep the final wording accountable to your team, not to the drafting tool.

Use visual conditional logic to show how declared answers lead to a sales-ready route, a self-serve resource route, or a human review queue. This is more reviewable than routing that exists only in an undocumented prompt. Dynamic content can repeat a prospect’s declared use case or timeline on the confirmation page without presenting a generated inference as fact.

On the response side, partial-response capture can distinguish a started interaction from a completed request, which helps prevent premature high-stakes actions. Structured Person, Company, and custom fields can retain declared context for routing and follow-up. Hidden fields and UTM capture preserve acquisition source and campaign provenance.

Finally, use pipeline stages, assignees, and notes to establish human ownership. Automations, Slack notifications, webhooks, test mode, and execution logs can help teams test what fires after a submission and investigate what happened later. Analytics can show whether the disclosure changes funnel behavior.

This is the managed-form layer: not just collecting an answer, but retaining the context, rules, ownership, and activity around it. It does not replace legal advice, privacy analysis, marketing-consent design, accessibility review, vendor due diligence, or any product-level content-marking requirement that may apply to a particular system.