What B2B companies owe, who will enforce it, and what it misses

As of 2 August 2026, any company putting generative AI in front of people in the European market has to say so. Chatbots must identify themselves as machines. Synthetic images, audio, video and text must carry machine-readable marks. Deepfakes and AI-written content published to inform the public on matters of public interest must carry a visible label. Fines reach €15 million or 3 percent of worldwide annual turnover, whichever is higher. Similar to GDPR, the obligation attaches to where the output lands, so a Toronto firm running a chatbot for European customers is bound by it.

This is governed by Article 50 of Regulation (EU) 2024/1689, and it is the first tranche of the AI Act that lands on ordinary marketing and communications teams rather than on model developers and safety engineers.

The practical effect, as WIRED reported ahead of the deadline, is that Europeans are about to find out how much of their day already runs on AI. Advertising, music recommendations, calendar and scheduling apps, and complaint hotlines all carry disclosure obligations once AI is involved. Frederiek Fernhout, a technology lawyer at Stibbe, told the magazine that genuine compliance makes the extent of AI use visible, and that marketing is where it shows most. Her colleague Thibau Duquin pointed out that the model-developer obligations reach well past the names everyone expects, catching Spotify for recommendation systems and Adobe for generative editing in Photoshop.

The four disclosures

Article 50 sets out four duties, and they attach to different parties.

Systems that interact with people must be designed so the person knows they are dealing with AI, and the disclosure has to arrive at first contact rather than in a terms-of-service page. The exception is where the AI nature is obvious to a reasonably observant person.

Generative systems must mark their outputs in a machine-readable format so that synthetic audio, image, video and text is detectable as artificially generated. This is a provenance obligation aimed at detection tooling, not at the human eye. It carves out systems performing an assistive function for standard editing.

Emotion recognition and biometric categorisation systems require notice to the people exposed to them.

Deepfakes and public-interest text require a visible label applied by whoever publishes them. For text, three conditions have to stack up: the content is AI-generated or AI-manipulated, it is published, and its purpose is to inform the public on a matter of public interest. Text that has gone through genuine human review, with a person or legal entity holding editorial responsibility, falls outside the labelling duty.

The Commission adopted its guidelines on Article 50 on 20 July 2026, and they read the editorial exemption narrowly. Human review means deliberate examination of the substance by someone with relevant competence and professional judgement. Spell-checking, grammar correction, and a cursory sign-off do not qualify. The accompanying Code of Practice goes further and asks for a documented editorial workflow with named responsible people, rather than an assertion after the fact that a human looked at it.

Free and open-source licensing provides no exemption from any of this.

Who carries the obligation

The Act splits duties between providers and deployers, and B2B companies routinely misclassify themselves.

A provider develops an AI system, or has one developed, and places it on the EU market under its own name or trademark. That definition catches more B2B companies than most realise. A firm that builds a customer-facing assistant on top of somebody else’s model and ships it under its own brand is the provider of that AI system, even though it did not train anything. The obligations under 50(1) and 50(2) follow accordingly.

A deployer uses a system somebody else built. Marketing teams generating campaign assets in a commercial tool, agencies producing client creative, and communications departments drafting public statements are deployers. Their obligations are visible labelling under 50(3) and 50(4), which is a disclosure and workflow problem rather than an engineering one.

Both categories reach outside the EU. Article 2 applies the regulation to providers and deployers established anywhere, where the output of the system is used in the Union.

What changes for B2B companies

The immediate operational exposure sits in four places.

Website and product chat. Any assistant, agent or avatar that talks to a person needs to identify itself on first contact. Voice agents are the sharper case, because an unlabelled synthetic voice on an inbound support line is also a deepfake question. Contact centres running sentiment or frustration detection on live calls sit under 50(3) and owe a declaration at the start of the call, which is a script change and a telephony change at the same time.

Business-to-business operations. Practitioner readings of 50(1) run wider than most compliance memos assume. Stibbe’s lawyers take the view that commercial AI use between businesses is caught too, including systems that schedule appointments, handle correspondence, and negotiate contract terms. That reading turns disclosure into an operations question rather than a marketing one, because the systems involved are usually procured by sales ops and never reviewed by legal.

Content pipelines. Anything published to inform the public on matters of public interest is in scope, and the Commission reads that category broadly enough to pull in employment, health, financial, and legal communications. Thought leadership, market commentary, benefits explainers, security advisories, and regulatory updates are closer to the line than most content calendars assume. The route out is the editorial exemption, and the price of that route is a documented review process with named accountability.

Creative and advertising assets. Photorealistic synthetic people, AI-altered spokespeople, and generated product environments trigger the deepfake labelling duty. Metadata survival becomes an operational question, because CMS platforms, image optimisers and social re-encoding routinely strip provenance metadata from files that arrived carrying it.

Contracts. Agency agreements, freelance terms and vendor MSAs need clauses covering who marks, who labels, and who holds editorial responsibility. Procurement questionnaires from European customers will start asking, and a defensible answer is worth more than a compliant one that cannot be evidenced.

Article 4, the AI literacy duty, has applied since February 2025 and did not move in the Omnibus. Companies with staff using AI systems owe them a sufficient level of AI literacy, and that obligation predates everything discussed here.

What moved, and what did not

The Digital Omnibus on AI, proposed by the Commission on 19 November 2025, reached provisional political agreement on 7 May 2026, was endorsed by Parliament on 16 June, cleared the Council on 29 June, and appeared in the Official Journal on 24 July 2026. It is law.

It did however defer the heavy machinery. Obligations for stand-alone high-risk systems listed in Annex III, covering recruitment tools, credit scoring, education, law enforcement, border control and critical infrastructure, moved from 2 August 2026 to 2 December 2027. High-risk AI embedded in products already regulated under Annex I product safety law moved to 2 August 2028. The final text dropped the conditional trigger tied to the availability of harmonised standards, which makes both dates fixed.

Article 50(2) machine-readable marking got a narrow extension. Generative systems already on the EU market before 2 August 2026 have until 2 December 2026 to embed marks. Systems entering the market on or after 2 August get no runway at all. Every other transparency duty applied on schedule.

The reason for the deferral is administrative rather than philosophical. Member states were late designating enforcement authorities, CEN and CENELEC missed the 2025 deadline for the technical standards that high-risk conformity assessment depends on, and the Commission was demanding conformity against benchmarks that did not exist.

The Omnibus also added a prohibition on AI-generated non-consensual intimate imagery and child sexual abuse material into Article 5.

Who enforces this, and how hard

Three bodies share the work. The AI Office enforces the rules for AI systems offered by the same provider as the underlying general-purpose model. National competent authorities in each of the 27 member states enforce for everyone else. The European Data Protection Supervisor covers systems used by EU institutions.

The AI Office arrived at 2 August with real powers for the first time. It can compel technical documentation, run model evaluations, demand risk mitigation, restrict or withdraw a model from the EU market, and fine up to 3 percent of global turnover or €15 million. It has appointed Alessandro Abate as lead scientific adviser. General-purpose model obligations have applied since August 2025, so a year of voluntary preparation ended. Models placed on the market before 2 August 2025 have until 2 August 2027.

The national side is less robust. Member states were required to designate market surveillance and notifying authorities by 2 August 2025. As of a March 2026 European Parliament research report, eight of 27 had designated single points of contact. A tracker maintained by the Future of Life Institute’s AI Act observatory put nine member states in the clear as of mid-2026, twelve partially resolved, and six with nothing designated. Germany runs a hybrid model coordinated by the Bundesnetzagentur, Spain built a dedicated agency in AESIA, Ireland spread the work across fifteen authorities under a national AI office.

Rosie Nance, a data and AI regulatory lawyer at a major global law firm, told WIRED that the uneven state of national supervisory frameworks means enforcement will start slower and less uniformly than the headline date implies, and that consistency is unlikely early on. Which supervisor a company answers to becomes a material variable, and one that a company with EU customers in several member states does not get to choose.

For a mid-market B2B company, the realistic near-term risk is not a sweep. The EU launched a public complaint tool, a confidential whistleblower channel for people working with AI providers, and a separate channel for downstream providers to report the model providers they build on. Enforcement will be complaint-driven first, and employees and competitors are now formal channels into it.

The faster lever is adjacent. The Commission’s guidelines note that an unlabelled deepfake can itself constitute illegal content under the Digital Services Act. Platform notice-and-action moves in days. An AI Act fine moves in years.

DSA precedent suggests the sequence: model providers and very large platforms first, mid-market later, with the first year spent on formal proceedings that generate guidance rather than penalties.

What platforms have signed and shipped

Roughly 190 organisations signed the Code of Practice on Transparency of AI-generated Content before the 27 July signature deadline. The Commission concluded on 8 July that the code adequately covers the obligations in Article 50(2), (4) and (5). Signatories on the provider side include Aleph Alpha, Anthropic, Black Forest Labs, Cohere, Google, Meta, Microsoft, Mistral, OpenAI and Synthesia. The deployer side drew Bulgari, Fastweb, Getty Images, Iberdrola, Lenovo and Lufthansa. Two task forces launch in September 2026.

Signing matters commercially. For signatories, enforcement focuses on monitoring adherence to the code rather than opening from scratch, and the AI Office has signalled a degree of presumption of conformity. It also travels: a signatory gets the same treatment regardless of which national authority supervises it. Deployers can sign Section 2 on its own.

On tooling, the infrastructure is further along than the regulation. Google signed the code and is pushing SynthID watermarking through partnerships with Apple, ElevenLabs, Kakao, NVIDIA and OpenAI, alongside the C2PA provenance standard developed with Adobe, Microsoft and the BBC. The two technologies are complementary in a useful way. C2PA carries rich provenance and dies on a screenshot. SynthID survives compression and re-encoding and carries almost no information. Layering them means a file stripped of its credentials can still be flagged.

Platform-side labelling predates the deadline. Meta runs an AI Info label across Facebook and Instagram and moved ad disclosure from voluntary to automatic enforcement in July 2026. TikTok has required creator labelling for years and reports billions of tagged assets. LinkedIn surfaces C2PA content credentials. YouTube has a disclosure checkbox in Creator Studio. For a marketing team, these are the practical compliance surface, and using the native toggle is usually the cheapest defensible answer.

Enthusiasm is uneven. Google publicly signed while stating that added regulatory complexity cuts against Europe’s competitiveness and simplification goals, and that constant labelling risks users tuning the labels out.

Where the regime stops short

The marking standard does not exist yet. Article 50(2) requires marks that are effective, interoperable, robust and reliable, and leaves the technical definition to codes of practice and standardisation work that is still running. CEN and CENELEC are targeting the end of 2026. Companies are being asked to comply with a specification that has not been finalised.

Text marking is the weak point. Watermarking language robustly is unsolved in a way that image watermarking is not, and the Act treats both identically. Research on Article 50(2) implementation has flagged the deeper paradox: marks durable enough to survive human inspection risk being learned as spurious features during training, while marks suited to machine verification break under ordinary data processing.

Marks are not required to survive distribution. Downstream distributors and platforms are encouraged to preserve upstream provenance, not obliged to. A label that dies on re-upload is the ordinary case, and nothing in Article 50 fixes that.

The editorial exemption is wide. Any organisation willing to build a documented review workflow with a named responsible person exits the public-interest text labelling duty entirely. The rule therefore bites hardest on unattended generation pipelines, which is defensible as policy and leaves the majority of professional AI-assisted publishing unlabelled.

Public interest has no hard boundary. Guidance reads it broadly, which creates a compliance category that a marketing team cannot classify with confidence and a regulator cannot police consistently.

Agentic systems sit awkwardly. Practitioners read 50(1) to cover AI handling correspondence and contract negotiation on the commercial side, and that reading is defensible where a natural person is on the other end. It thins out where no person is. An agent transacting with another company’s agent, filling forms and executing purchases on a user’s behalf, is not squarely inside any of the four duties, and the volume of that traffic grows every quarter.

The definitions are not settled. Fernhout’s assessment is that many terms remain unclear and will need authority guidance, and that the requirements have to be engineered into underlying systems rather than bolted on. Both are true, and both mean the first year produces interpretation rather than penalties.

Disclosure fatigue has a precedent. GDPR gave people real rights over their personal data and gave the web the cookie consent banner, which most users now dismiss without reading. Google’s public position on signing the transparency code makes the same argument about labels, and it is not only self-serving. A disclosure that appears on everything communicates nothing, and Article 50 contains no mechanism for calibrating salience to actual risk.

No private right of action. The AI Liability Directive was formally withdrawn in October 2025 and the Commission has no plans to revive it. The revised Product Liability Directive extends to software and AI systems, and excludes B2B transactions and pure economic loss. Someone harmed by an unlabelled synthetic asset in a commercial context falls back on national law, which varies.

Decisions are not covered. The transparency regime governs content. AI-mediated decisions about hiring, credit and access sit in the high-risk regime, now deferred to December 2027 and August 2028. The obligations that would matter most to a person on the receiving end are the ones that moved.

What a compliance lead should do this quarter

Inventory the AI systems already generating or manipulating content, including the ones embedded in CRM, support desk, analytics and design tools that nobody classified as AI when they were bought. Determine, for each, whether the company is a provider or a deployer, because the answer changes what is owed and the branded-assistant case may catch people out and eventually be decided in court.

Check what your inventory of hosted platforms have already deployed. In addition to those already mentioned,WordPress for example, one of the largest content hosting platforms in the world, has already made detecting and labelling adjustments to its images prior to the Act coming live. Substack and LinkedIn have introduced the ability to vet content as AI produced (Substack) or allow users to submit any notification that they believe content has been AI produced (LinkedIn).

Fix first-contact disclosure on every conversational surface. Decide which content categories fall inside public interest and route those through a documented editorial process with named accountability, or label them. Test whether provenance metadata survives the asset pipeline end to end. Update agency and freelancer terms. Evaluate signing Section 2 of the Code of Practice, which is cheaper than the alternative and portable across all 27 supervisors.

The deferral to December 2027 bought time on the hardest work, which is classifying systems and determining which are high-risk. The classification exercise is the same one Article 50 compliance requires. Doing it once, properly, covers both.

The more interesting question sits underneath the compliance work: companies have to decide whether they are comfortable declaring it. Some will label everything and keep going. Some will find that the label on a customer service line or a piece of corporate correspondence costs more in trust than the automation saved, given growing global movements against the technology, and will face the choice toomove work back to systems that need no disclosure at all, or back to people.