Reviewed against the European Commission guidance published 20 July 2026.
The countdown to 2 August 2026 has created a dangerous misunderstanding. The EU delayed important rules for high-risk AI systems, so some management teams assumed the whole AI Act had moved. It did not.
The EU AI Act transparency requirements in Article 50 still start applying on 2 August 2026. They affect chatbots and AI agents that speak to people, systems that generate synthetic content, emotion-recognition and biometric-categorisation tools, deepfakes, and some AI-generated text on matters of public interest.
This is not only a European-company issue. A business based in the Middle East or North Africa may fall within scope when it places an AI system on the EU market or when the output of its system is used in the EU. For exporters, SaaS providers, agencies, shared-service centres and international groups, the practical question is therefore not “Are we incorporated in Europe?” It is “Where do our AI systems and their outputs reach people?”
This guide turns the Commission’s new guidance into an operational checklist for CEOs, legal teams, technology leaders, marketers and process owners. It is a management guide, not legal advice; organisations should confirm their precise obligations with qualified counsel.
What changed in July 2026?
On 20 July, the European Commission published guidelines and a detailed FAQ explaining who must comply with Article 50 and how the EU AI Act transparency requirements work in practice. The timing matters for three reasons:
- 27 July 2026, 18:00 CEST: the deadline to join the initial list of signatories to the voluntary Code of Practice on Transparency of AI-Generated Content.
- 2 August 2026: Article 50 becomes applicable to providers and deployers in scope.
- 2 December 2026: a limited transition date applies to the machine-readable marking obligation for certain generative AI systems placed on the market before 2 August 2026.
The high-risk timetable is separate. Following the AI Omnibus agreement, rules for stand-alone systems in specified high-risk areas are scheduled for 2 December 2027, while rules for high-risk systems embedded in regulated products are scheduled for 2 August 2028. Those delays do not cancel the August 2026 transparency duties.
Who needs to pay attention?
The EU AI Act transparency requirements distinguish between a provider and a deployer. The same company can be one or both, depending on what it does.
| Role | Plain-English meaning | Typical MENA example |
|---|---|---|
| Provider | Develops an AI system—or has one developed—and offers it or operates it under its own name or trademark. | A regional software company selling a branded customer-service agent to EU clients. |
| Deployer | Uses an AI system under its authority for professional activity. | An exporter using a third-party AI avatar in an EU-facing sales journey, or an agency publishing AI-generated campaign content for a client. |
Location alone does not settle the question. The Commission says providers outside the EU can be subject to the Act when their AI system’s output is used in the EU. Employees and contractors acting under a company’s authority are generally not treated as separate deployers; responsibility remains with the legal entity controlling the use.
EU AI Act transparency requirements at a glance
1. Tell people when they are interacting with AI
Providers of AI systems that directly interact with people must design them so that users know they are dealing with AI, unless that fact is obvious. The guidance covers genuine two-way, direct exchanges with natural persons—such as chatbots, AI agents and avatars—not background analytics or machine-to-machine processes.
The notice should appear from the start of the first interaction, be clear and distinguishable, and meet accessibility requirements. A vague disclosure buried in terms and conditions is unlikely to serve the same purpose as an immediate, visible notice.
Practical example: an EU visitor opens a support widget on a MENA exporter’s website. A concise opening line such as “You are chatting with our AI assistant” is clearer than relying on a robot icon or a product name that may not be obvious to every user.
2. Make synthetic content machine-detectable
Providers of systems that generate or manipulate synthetic audio, images, video or text must support machine-readable marking that is effective, reliable, robust and interoperable. This is a system-design obligation: it is not satisfied merely by adding a visual caption after publication.
The Commission identifies exclusions and limits, including source code, short strings, machine-to-machine outputs not exposed to people, certain closed-loop production uses and standard assistive editing. It also describes a narrow exemption for qualifying business-to-business or industrial contexts. “B2B” should not be treated as a blanket escape; the conditions must actually be tested and documented.
3. Inform people exposed to emotion recognition or biometric categorisation
Deployers of these systems must inform the people exposed to their operation, whether the processing happens in real time or later. This can affect workplace technology, visitor analytics, event systems, research tools and customer-experience applications.
For HR and operations leaders, the immediate task is discovery. A tool may include an emotion or biometric feature inside a wider recruitment, training, video, security or analytics platform. If procurement records describe only the main product, the relevant AI function can remain invisible.
4. Clearly label deepfakes and certain public-interest text
Deployers must disclose deepfake image, audio or video content by the time a person is first exposed to it. The disclosure must be perceptible without requiring a special tool, so a hidden machine-readable mark alone is not enough.
AI-generated or manipulated text published to inform the public about matters of public interest must also be labelled. The scope can include economic, financial, scientific, political, cultural, public-health, environmental and consumer-safety developments that may be debated publicly.
There is an important exemption when the text has undergone substantive human review or editorial control and a person or editorial entity holds responsibility for publication. A spelling pass or a quick formatting check does not qualify. The reviewer needs relevant knowledge and real authority to approve, change or reject the substance, including fact-checking and source assessment.
A practical EU AI Act compliance checklist for 2026
A legal memo alone will not make an organisation ready. The EU AI Act transparency requirements reach product design, content operations, vendor management, HR, customer service and governance. Use the following sequence to turn the rules into operating controls.
Step 1: Build an AI inventory around use cases, not product names
List every AI-enabled interaction and content flow that can reach an EU user, employee, customer or audience. Include sanctioned software, browser extensions, embedded features, agency tools and “shadow AI” adopted directly by teams.
For each use case, record the business owner, vendor, model or service, intended audience, countries reached, content types, whether people interact directly with the system, and whether the company presents the system under its own brand.
Need a reliable AI use-case map before the EU AI Act transparency requirements take effect?
Business Wheel can facilitate a focused readiness session with your technology, legal, operations and customer teams—then turn the findings into a prioritised action register.
Step 2: Decide your role for each use case
Do not assign one role to the whole company. A business may deploy a third-party writing assistant, provide a white-labelled chatbot, and publish synthetic video—all under different Article 50 duties. Map provider/deployer status use case by use case.
Step 3: Classify the transparency trigger
- Direct human interaction with a chatbot, agent or avatar
- Generation or manipulation of text, audio, images or video
- Emotion recognition or biometric categorisation
- Deepfake content
- Public-interest text without substantive human editorial control
If a team claims an exception—“obvious AI,” standard editing, closed-loop use, B2B context or editorial review—record the reasoning and evidence. Exceptions should be decisions, not assumptions.
Step 4: Design the user-facing disclosure
Specify exactly where the notice appears, when it appears, what it says and how it works for users with disabilities. Test it in the languages and channels actually used. For deepfakes, make sure the disclosure remains attached when content is embedded, clipped, downloaded or reposted where reasonably controllable.
Step 5: Validate machine-readable marking with providers
Ask vendors to explain their marking method, supported content formats, robustness after common transformations and interoperability. Contract language should address documentation, changes to the marking method, incident support and the evidence needed for regulatory enquiries.
Step 6: Create a real editorial-control path
For public-interest material, name a competent human reviewer, define the sources that must be checked, give the reviewer authority to reject publication, and retain a simple approval record. The objective is not bureaucracy; it is accountable judgement.
Step 7: Train the people who can create exposure
Prioritise marketing, communications, customer service, product, IT, HR, procurement and agency managers. Training should use the company’s real tools and scenarios. Teams need to know when a disclosure is required, how to escalate uncertainty and why removing a label can create risk. A role-based competency framework can turn that knowledge into clear expectations, assessment and reinforcement.
Step 8: Keep an evidence pack
Maintain the inventory, role assessment, exception decisions, screenshots of notices, accessibility checks, vendor statements, editorial approvals, training records and change logs. A control that exists but cannot be demonstrated will be harder to defend.
Should a company sign the voluntary Code of Practice?
The Code is voluntary, but the choice is consequential. According to the Commission, signatories can rely on an approved route for demonstrating compliance with the marking and labelling duties within the EU AI Act transparency requirements. Non-signatories may use other adequate measures, but should be prepared to explain them and may face more information requests.
A company should not sign simply for a badge. Senior leadership should first confirm that the commitments match the systems the organisation provides or deploys and that owners, controls and evidence are in place. The initial-signatory deadline is 27 July, but the Commission says organisations can sign later.
Common mistakes to avoid
Most failures are not caused by a lack of policy language. They happen when the EU AI Act transparency requirements are reduced to one department’s task and never translated into product, content and procurement decisions.
- “The AI Act was delayed.” Some high-risk rules were delayed; Article 50 was not.
- “We are outside Europe.” Non-EU providers can be in scope when their system’s output is used in the EU.
- “Our vendor handles compliance.” Providers and deployers have different duties. A vendor feature does not remove the deployer’s labelling or notification responsibilities.
- “Everyone knows it is AI.” The “obvious” exception is intended to be narrow. Test the experience from the perspective of an average user.
- “A human glanced at the article.” Substantive review, fact-checking, authority and responsibility matter; a cosmetic check is not editorial control.
- “A watermark solves everything.” Machine-readable marking and human-perceivable disclosure are separate controls in several scenarios.
A 10-day action plan for management teams
Use this EU AI Act transparency requirements action plan to move from legal interpretation to owned controls, tested user journeys and auditable evidence.
| Timing | Management action | Output |
|---|---|---|
| Days 1–2 | Appoint an executive sponsor and a cross-functional owner; freeze unreviewed launches affecting EU audiences. | Named accountability and scope |
| Days 2–4 | Inventory AI interactions and synthetic-content flows; map provider/deployer roles. | Prioritised AI register |
| Days 4–6 | Implement chatbot notices, deepfake labels and public-interest editorial review; contact critical vendors. | Minimum viable controls |
| Days 6–8 | Test notices, accessibility, content hand-offs and evidence capture. | Test results and remediations |
| Days 8–10 | Approve residual risks, train priority teams and decide whether to sign the Code of Practice. | Executive decision and evidence pack |
Already have policies, but execution is fragmented?
Connect governance to roles, workflows, vendor decisions and adoption metrics. Business Wheel’s transformation approach is designed for companies that need practical change without disrupting day-to-day operations.
Why this is a business-transformation issue
The companies that handle Article 50 well will not treat it as a label-writing exercise. They will use it to clarify who owns AI, which use cases create value, which vendors can be trusted and where human judgement must remain in the process.
Meeting the EU AI Act transparency requirements belongs in the operating model. It connects governance with process design, procurement, customer experience, workforce capability and performance measurement—the same foundations required to scale AI responsibly. Teams planning a wider rollout may also benefit from this practical guide to change management consulting and the Business Wheel framework for integrated business development.
Business Wheel helps organisations turn digital ambition into executable operating models, practical controls and capable teams. If your company serves European customers or operates cross-border AI workflows, explore our digital transformation solutions, see how our business development approach supports market readiness, or request a focused readiness discussion.
Frequently asked questions
Do the EU AI Act transparency requirements apply to companies outside the EU?
They can. The Commission states that a provider established outside the EU can be subject to the Act when the output of its AI system is used in the EU. Each organisation should assess its systems, role and value chain rather than relying on incorporation location alone when mapping the EU AI Act transparency requirements.
When does Article 50 apply?
Article 50 applies from 2 August 2026. A limited transition to 2 December 2026 covers the machine-readable marking obligation for certain generative AI systems placed on the market before 2 August; it does not postpone every transparency duty.
Must every piece of AI-generated text be labelled?
No. The deployer labelling duty focuses on published AI-generated or manipulated text intended to inform the public about matters of public interest. Substantively human-reviewed text with genuine editorial control and responsibility can qualify for an exemption.
Is a chatbot icon enough disclosure?
Not necessarily. People should be informed clearly from the start of the first direct interaction unless it is obvious that they are interacting with AI. A short, accessible written notice is usually easier to understand and evidence.
What is the first step for an executive team?
Name an accountable owner and build a use-case inventory covering customer, employee and content workflows that reach the EU. Without that map, the organisation cannot reliably identify its role, duties, exceptions or vendors.
Primary sources and further reading
- European Commission: Article 50 guidelines announcement, 20 July 2026
- European Commission: Article 50 transparency FAQ
- European Commission: signing the Code of Practice
- European Commission: AI Act overview and implementation timeline
- Council of the EU: updated high-risk AI timetable
Editorial note: This article’s explanation of the EU AI Act transparency requirements was prepared from official European Commission guidance available on 23 July 2026. The regulatory position can change; confirm decisions against current official guidance and legal advice.
