EU AI Act: What applies to your company now, and what can wait

Two questions decide what applies

Every duty in the AI Act hangs on two things: your role for a given AI system, and how risky its use is. You answer both for each system, so one company can have several answers.

First, check it is an AI system at all. The Act covers software that works out its output from the data it receives, such as a model that sorts support tickets by urgency. A fixed rule a person wrote, like "send every ticket that mentions an invoice to finance", is ordinary software and falls outside the Act.

Your role: provider or deployer

The provider is whoever develops an AI system, or has one developed, and puts it on the market under their own name. If your product has an AI feature, you are the provider of that feature, even when a model from OpenAI or Mistral runs underneath.

The deployer is whoever uses an AI system in their work. If your sales team drafts emails with ChatGPT, your company is a deployer of ChatGPT.

Most companies are both: provider for their own product, deployer for the tools their staff use. One thing can change your role. If you put your own name on a high-risk system someone else built, or change it substantially, you take over the provider's duties (Article 25 of the AI Act).

The risk level

The Act sorts uses into four levels:

• Banned. A short list of uses that are not allowed at all, such as scoring people on their social behaviour or reading employees' emotions from a webcam.
• High-risk. Uses where a mistake can seriously affect someone's life. Annex III of the Act lists them: recruitment and other human resources (HR) decisions, credit scoring, life and health insurance pricing, access to education, biometric identification, critical infrastructure, and public-sector uses such as police work, border control, courts and decisions on access to public benefits. AI built into regulated products such as medical devices or machinery is high-risk as well.
• Limited risk (transparency duties). AI that talks to people (chatbots) and AI that generates content, including deepfakes. The duty is to be open about them.
• Minimal risk (everything else). Spam filters, code assistants, a model that suggests the next reply to a support ticket. No duties beyond AI literacy.

The level follows the use, and the technology doesn't matter. The same language model is in "everything else" when it summarises meeting notes and high-risk when it ranks job applicants.

What applies today

Four parts of the AI Act already apply. Here is each one and who it concerns.

No banned uses (since February 2025)

The list of banned practices in Article 5 of the AI Act has applied since 2 February 2025. Most of it is far from a normal company's work, such as predicting crimes from a person's profile or building facial recognition databases by scraping photos from the internet. Two items are worth checking against your own tools: emotion recognition at work or in education, meaning AI that tries to read how people feel from their face or voice, and techniques that manipulate people or exploit their age or situation in harmful ways.

An HR tool that claims to read engagement from employees' faces in video calls counts as emotion recognition at work and is banned. Outside work and education, emotion recognition is allowed, but the people it is used on have to be told, and from December 2027 it counts as high-risk.

From 2 December 2026, the list also covers AI systems that produce sexualised images of real people without their consent, or of children.

AI literacy (since February 2025)

Article 4 of the AI Act asks providers and deployers to support AI literacy among the people who work with AI on their behalf. The July 2026 changes softened the wording: you have to take measures to support it, and no particular level is required.

In practice, the people using AI at work should know what the tools are good at, where they go wrong, and what your rules say about company data. Those rules usually sit in an AI usage policy, a page saying which AI tools staff may use and what data may go into them. A short session built on that policy is enough to meet the literacy requirement. Keep a list of who attended, so you can show it if anyone asks.

Transparency (since August 2026)

Article 50 of the AI Act has applied since 2 August 2026, and the July changes left it in place. It has four parts:

• Providers: if people interact with your AI, for example a support chatbot, tell them it is AI unless that is obvious.
• Providers: if your product generates text, images, audio or video, mark the output in a machine-readable way, such as metadata or a watermark, so software can detect it as AI-generated. For products already on the market before 2 August 2026, this applies from 2 December 2026.
• Deployers: if you publish deepfakes, meaning AI-made images, audio or video of real people, places or events that look authentic, say they are artificially generated. The same applies to AI-written text published to inform the public on matters of public interest. If a person has reviewed the text and someone takes editorial responsibility for it, no label is needed.
• Deployers: if you use emotion recognition or biometric categorisation, which sorts people into groups such as age or gender from their face or voice, tell the people you use it on.

In June 2026 the European Commission published a voluntary code of practice on marking and labelling AI-generated content, which shows what good marking looks like. An AI-generated image, for example, carries embedded metadata saying it was made with AI, plus an invisible watermark that detection software can read.

General-purpose models (since August 2025)

Rules for general-purpose AI models, the large models behind tools like ChatGPT, have applied since 2 August 2025. They fall on the companies that build and release those models. If you send requests to a model through its API (the interface developers use to call it from their own software), or train one further on your own data at a modest scale, these rules stay with the model's provider.

What can wait until 2027

The high-risk rules are the heavy part of the AI Act. They apply from 2 December 2027 for the uses listed in Annex III, such as hiring, credit scoring and education, and from 2 August 2028 for AI built into regulated products. These dates are fixed.

If you provide a high-risk system, you will need:

• a risk management process for the system
• rules for the data you train and test it on
• technical documentation, a written description of how the system works, what it was trained on and how it was tested
• logging, meaning records the system keeps automatically of what it did
• human oversight, a person who understands the output and can overrule or stop it
• a quality management system under Article 17 of the AI Act, the written processes that keep all of the above running as the system changes
• a conformity assessment before release, a documented check, for most uses done by the provider itself, that the system meets these requirements
• registration in the EU's database of high-risk AI systems

If you deploy one, for example a tool that screens CVs, your duties are lighter:

• use it as the provider's instructions say
• have a trained person oversee it
• watch how it performs, and report serious incidents to the provider and the authority
• keep the logs it produces
• tell your staff before it is used on them, and tell the people it makes decisions about, such as job applicants

Deployers that use AI for credit scoring or life and health insurance pricing, and public bodies, also have to assess the effect on people's fundamental rights before use.

There is one way out. A system in an Annex III area is not high-risk if it only does a narrow task, such as turning unstructured documents into structured data, and does not build a profile of people to predict their behaviour or performance. You still have to document that assessment and register the system.

December 2027 sounds far away, but for a provider, a quality management system plus technical documentation for each system takes most of that time. If nobody in the company owns this yet, an outside security lead can carry it until someone does.

A worked example

Take a 60-person software company that sells an HR platform.

• Its product has a chatbot that answers employees' questions about leave and payroll. The company is the provider and the use is limited risk: the chatbot has to say it is AI. This applies now.
• The product drafts job adverts with a language model. As provider of a generative feature, the company has to mark the text in a machine-readable way. This applies now for new features. For a feature already on the market before 2 August 2026, it applies from 2 December 2026.
• Its staff use ChatGPT and Copilot. The company is a deployer, at the minimal-risk level, so it needs AI literacy measures, and a usage policy is the simplest way to deliver them. This applies now.
• It plans a feature that ranks job applicants. The company is the provider of a high-risk system, with the full set of high-risk duties from 2 December 2027. The work should start now.

Going through your own AI tools and features in the same way is the first step of checking where you stand against the AI Act.

Who checks this in Germany

Since 29 July 2026 Germany has had its own implementing law, the KI-MIG (Gesetz zur Marktüberwachung und Innovationsförderung von künstlicher Intelligenz). It makes the Bundesnetzagentur, the federal network agency, the main market surveillance authority for the AI Act, meaning the authority that checks AI systems on the market and can order changes. BaFin, the financial supervisor, stays in charge for banks and insurers. The Bundesnetzagentur also runs an AI Service Desk (KI-Service Desk) with a free online check (in German) for a first estimate of your risk level, a useful way to test your own reading. General-purpose models are supervised by the EU AI Office.

Where to start

1. List where AI sits in your business: features in your product, AI inside tools you resell, and tools your staff use. One table is enough.
2. For each entry, note your role and its risk level. If anything lands in banned, stop using it.
3. If your product talks to people or generates content, check the chatbot disclosure and the output marking now. For products already on the market before 2 August 2026, the marking applies from 2 December 2026.
4. Run one AI literacy session on your usage policy and keep the attendance list.
5. If an entry is high-risk, start on the quality management system this quarter, and give AI governance a named owner.

What's next?

Some companies will want to show customers how they govern AI, and who actually needs an AI management system explains when ISO 42001 helps with that and when it is sold to companies that don't need it. And because step one of all this is a list of where AI sits, how to build an asset inventory that holds up under audit shows how to keep that kind of list current.