“EU AI Act from August 2026: What labeling obligations really affect SMEs” (SMEs = Small and Medium

8 min. read

On 2 August 2026, AI labelling moves from optional to mandatory. Any mid-sized business running chatbots, publishing AI-generated text, or using synthetic images will fall under Article 50 of the AI Act from that date. The obligations are manageable – the lead time is not.

Related:AI Trust Under Pressure: Anthropic Makes Covert Interventions Visible  /  Desertec Reloaded: What Desert-Generated Power Means for Energy Costs

Why August 2026 Matters for Mid-Sized Businesses

Most discussions about the AI Act revolve around the major model providers. That is understandable – but it misses the reality facing many mid-sized companies. Nobody in the marketing department of an 80-person business is training a language model. But a great many are drafting product copy with an AI assistant, building a service chatbot into their website, or generating campaign visuals synthetically. That is precisely the practice covered by Article 50.

The deadline is 2 August 2026 – 24 months after the Regulation entered into force. On that date, the transparency obligations become applicable. They affect two roles: providers, who make an AI system available, and deployers, who use one. Most mid-sized businesses are deployers. That may sound like a supporting role, but it comes with its own set of obligations.

What is the deployer role? Under the Regulation, a deployer is anyone who uses an AI system under their own responsibility without having developed it themselves. A retail company that licences a ready-made chatbot and runs it on its own website is a deployer. It must ensure disclosure – even when the underlying technology comes from a third party.

The Three Obligations Under Article 50

Article 50 bundles several transparency rules that, in practice, boil down to three core obligations for mid-sized businesses. They are less abstract than the legislative text might suggest.

First, chatbot disclosure. Any organisation deploying an AI system that interacts directly with people must inform those individuals that they are communicating with a machine. This does not apply where it is self-evident. In practice, a notice at the start of a chat session is usually sufficient; a buried footnote is not.

Second, labelling of AI-generated content. Providers of generative systems must mark their outputs – audio, images, video and text – in machine-readable form as artificially created or edited. For operators, the downstream question is the relevant one: anyone who distributes AI-generated texts on matters of public interest must disclose their artificial origin.

Third, the deepfake obligation. Synthetic images, sounds or videos that resemble a real person or object must be labelled as artificially created or manipulated. For any company using AI avatars in advertising or synthesising voices, this is the sharpest of the three rules.

One Date, Multiple Deadlines

This is where things get practically important, because media coverage tends to create the impression of a single, clean cut-off date. The core obligations under Article 50 become applicable on 2 August 2026 – that date is fixed in the regulation. Alongside it, however, come further staggered deadlines: requirements for certain high-risk systems and for models already on the market before August 2025 will only take effect later under the regulation’s transitional provisions, in some cases as late as 2027. Which deadline applies to your specific system depends on its classification.

In addition, the EU is negotiating adjustments to individual deadlines under the so-called Digital Omnibus package. A provisional political agreement was reached in early May 2026, though formal adoption and publication in the Official Journal had not yet taken place at the time of writing. For mid-sized operators, the sober conclusion is this: prepare for the transparency deadline in August, keep a close eye on the detailed deadlines for existing and high-risk systems, and wait for the final text in the Official Journal.

Anyone already producing content today should not treat labelling as something to retrofit later, but as an integral part of the workflow. A requirement you build in from the start costs almost nothing. One you have to apply retroactively across hundreds of assets in July 2026 will be expensive.

What Operators Should Actually Do

From a founder and operator perspective, the appeal of this regulation is that it breaks down into small, immediately measurable steps. No slide marathon – just an inventory and four concrete actions.

The first step is an AI inventory. List every place in your organisation where generative AI produces output that goes external: website chat, newsletter copy, social visuals, product descriptions, promotional videos. This list is the foundation for everything that follows, and in most mid-sized businesses it can be compiled in a single afternoon.

The second step is disclosure at the point of contact. Chatbots get a clear notice that a user is interacting with an AI. The third step is a labelling rule for published content – for example, a standardised notice on AI-assisted posts. The fourth step concerns suppliers: check whether the AI tools you license already include machine-readable marking in their outputs, because that obligation sits with the provider. If it is not in the contract, it needs to be.

These four actions cover the bulk of what mid-sized businesses actually face. They are no substitute for individual legal advice, but they ensure that August 2026 does not arrive as a last-minute scramble.

Frequently Asked Questions

Editor’s Reading Picks

• A regulator shuts down an AI model: the lesson for SMEs
• What interchangeable AI copy really costs
• When the agent books the incoming invoice itself

More from the MBF Media Network

Cover image source: Pexels / Paras Katwal (px:4218883)