We use cookies on this page. View our Privacy Policy for more information.

The AI Act: what companies need to watch when deploying AI agents and apps

The AI Act sorts AI by risk. What applies to company chatbots and agents since August 2026, when high risk applies and why duties moved to 2027.

The AI Act: what companies need to watch when deploying AI agents and apps

Short answer: The AI Act (EU Regulation 2024/1689) divides AI systems into four risk categories and sets obligations accordingly. Most company chatbots and assistants fall under transparency obligations: a person should know they are communicating with AI. High risk typically concerns AI in recruitment and employee evaluation, credit assessment or critical infrastructure. The regulation has been generally applicable since 2 August 2026, but the obligations for high-risk systems were pushed back by the AI Omnibus package to 2 December 2027 and 2 August 2028 respectively. Check the specific impact on your company with a lawyer.

Why it matters

The regulation entered into force on 1 August 2024 and applies in stages. The bans on unacceptable practices and the AI literacy obligation have applied since 2 February 2025, the rules for general-purpose AI models (GPAI) since 2 August 2025, and since 2 August 2026 the regulation is generally applicable, including the transparency rules. That is how the European Commission describes it today.

For a company that deploys a chatbot, an agent or an app with AI, this means the question "does the AI Act apply to us?" is no longer a question about the future. For most ordinary deployments the obligations are quite manageable: let the user know they are talking to AI, train your people and know what runs in the company. It is harder where AI decides about people. This article is an overview of principles, not a legal analysis. We verified the situation on 4 November 2026, and the timetable may still change.

An inventory of AI systems, the rules, and the environment in which agents run are something we handle within AI Ops and AI Infrastructure. That is exactly what lets you work out what the AI Act means for you.

Deploying AI and not sure which category you fall into? In 30 minutes we will go through what your agents do and what follows from it. Book an online call

Four risk categories and where company AI falls

The AI Act regulates not the technology but its use. The same model can have no special obligations in one application and be high-risk in another. The European Commission distinguishes four categories.

CategoryWhat it meansTypical example in a companyWhat follows
Unacceptable riskProhibited practicesSocial scoring, manipulative techniques, some biometric usesMust not be used; the bans have applied since February 2025, and another ban (nudification apps) is added in December 2026
High riskPossible serious impact on health, safety or fundamental rightsAI in recruitment and employee evaluation, credit assessment, critical infrastructure, educationRisk management, data quality, logging, documentation, human oversight, cybersecurity
Transparency riskA person should know they are communicating with AI or that content was created by AICustomer chatbot, AI assistant, generated content, deepfakeInform about the AI, label generated or altered content
Minimal riskNo special rulesSpam filter, AI in games, text autocompleteNo special obligations under the AI Act (AI literacy applies generally, though)

The category is decided not by the model vendor but by the purpose for which you use the AI in your company. A chatbot that advises customers on a product typically raises transparency. The same chatbot that pre-screens job applicants can end up in high risk.

Provider or deployer: the role matters

The AI Act divides obligations by role. A provider develops an AI system or places it on the market under its own name. A deployer uses the system in its own activities, for example a company that puts a chatbot on its website. One company can have different roles in different projects, and if it significantly modifies a system or deploys it under its own brand, it can take on the provider's obligations too.

In practice: if you have an agent developed for you, clarify with the supplier and with a lawyer who has which role and who is responsible for what. The contract should say so.

Chatbots and assistants: transparency obligations

The transparency rules (Article 50) apply from 2 August 2026. Simplified, they say:

  • A system that communicates directly with people should be designed so that the person knows they are talking to AI, unless it is obvious. This is the provider's obligation.
  • Synthetic audio, images, video and text should be marked by the provider of the generative system in a machine-readable way, so that it is recognisable as AI-generated.
  • Whoever uses AI (the deployer) should disclose that a deepfake is artificially created or altered, and likewise AI text published to inform the public on matters of public interest, unless it has been through human editorial review.
  • The information should be clear and given at the latest at the first interaction.

Technically this is usually not demanding, but somebody has to decide it and implement it. How an assistant differs from an agent is described in the article What is an AI agent.

Our view: transparency is not security. It deals with the relationship to the user, not with what the agent is allowed to do. A bigger risk than the model itself is often architecture and permissions, for instance an agent with direct access to a production database or to payments. What to focus on is summarised in the article Security of AI apps according to OWASP. Injected instructions (prompt injection) are covered in the article Prompt injection and AI agents.

When a company agent becomes high risk

High risk does not depend on how clever the agent is, but on what it decides about. High-risk areas include, among others, biometrics, critical infrastructure, education, employment, access to essential services and migration. For an ordinary company the most common situations are three:

  • HR. An agent that sorts CVs, evaluates candidates or underpins decisions about promotion or dismissal.
  • Credit and access to services. AI that assesses whether a person gets a loan or another essential service.
  • Critical infrastructure. AI in the management of energy, transport or water supply, where an error endangers safety.

Stricter obligations apply to high-risk systems: risk management, data quality, logging, documentation, human oversight and robustness including cybersecurity. The AI Omnibus package (Commission proposal of 19 November 2025, political agreement of 7 May 2026, in effect since 27 July 2026) pushed back their start: for areas under Annex III to 2 December 2027, and for AI embedded in regulated products (Annex I) to 2 August 2028.

The delay does not mean you can forget about it. Logging, documentation and human oversight are harder to add to a finished system than when they are planned from the design stage.

What makes sense to address in every deployment

  • AI literacy. Article 4 requires providers and deployers to take measures so that people who work with AI have sufficient knowledge. In practice, training by role: different for a salesperson with a chatbot, different for a developer with an agent.
  • Inventory. Know what AI systems run in the company, for what, with what data and who is responsible for them. Without it you cannot determine the risk category. A typical problem is unapproved use of tools, which we write about in the article Shadow AI. NIS2 and GDPR obligations also build on the inventory, and we will cover them separately.
  • Human oversight proportionate to impact. For high risk it is an obligation, elsewhere good practice. Derive the level of control from what happens when it goes wrong and whether it can be undone. A draft reply to a customer, or a write you can restore from a backup, can take a lighter regime than a payment sent or a decision about a person that cannot be reversed.

Penalties and supervision in the Czech Republic

Fines under Article 99 reach up to 35 million euros or 7% of worldwide turnover for prohibited practices, and up to 15 million euros or 3% of turnover for most other obligations, including transparency. For small and medium-sized companies the lower of the two amounts applies.

In the Czech Republic the Czech Telecommunications Office is being prepared as the main supervisory authority, and since September 2026 it has been building a department for AI oversight. The regulation applies directly, but the Czech implementing act, which is meant to regulate supervision and penalty proceedings, had according to available reports not been adopted as of the verification date. Check the current status before making a decision.

When the AI Act is your topic and when it is not

You can handle it calmly if:

  • you use AI internally for texts, code or analyses without deciding about people,
  • your chatbot advises on a product and clearly introduces itself as AI,
  • you have no AI in recruitment, employee evaluation, credit or infrastructure management,
  • you have an overview of AI systems and the people who work with them are trained.

Deal with it now if:

  • AI at your company sorts applicants, evaluates employees or assesses applications for credit or a service,
  • an agent communicates with customers and nowhere says it is AI,
  • you publish deepfakes or AI texts on public topics without labelling,
  • nobody in the company knows how many AI systems run and for what,
  • you supply AI solutions to customers and they ask about AI Act compliance.
If you recognised yourself in the second list, send us a few sentences about what your agents do. We will get back to you with a proposal on how to proceed. Send an inquiry

How MeguMethod approaches it

We do not do the legal classification into a risk category; that belongs to a lawyer. But the technical things that compliance stands on, we handle from the start of a project. We use AI already in the product and technical specification, and every project has rules for AI: architecture, limits and restrictions, what the agent may and may not do.

The assistants we develop have a defined domain and traceable answers. Lennie for 2N is set up for technical support of 2N products, answers from 2N manuals and gives the source with each answer; AI in the 2N app works only online. FlexiBot for FlexiFin takes data through the app's internal API, and the conversation history is stored and logged on servers in the EU. More on the AI assistants page.

We check code created with the help of AI according to risk: first automated AI code review, then a person for the substantial risks. We recommend a penetration test where the impact and sensitivity of the system really require it. How to involve AI agents in development is described in the article AI agents for development in your company.

Frequently asked questions

Does the AI Act apply to a small company that uses ChatGPT?

Internal use for texts or analyses without deciding about people is typically minimal risk with no special obligations. AI literacy, however, applies to everyone who uses AI systems in their activities, so training people makes sense here too.

Does our chatbot have to introduce itself as AI?

Generally yes, unless it is obvious from the circumstances. The transparency rules have applied since 2 August 2026 and the information should be given at the latest at the first interaction. Who is responsible for it in a specific solution depends on the roles of provider and deployer.

When is an AI agent high risk?

When it decides or substantially influences decisions in sensitive areas, typically recruitment and employee evaluation, credit and access to services, critical infrastructure or education. What decides is the purpose, not the technology. Check the specific classification with a lawyer.

Since when do the obligations for high-risk systems apply?

After the AI Omnibus package, for areas under Annex III from 2 December 2027, and for AI in regulated products from 2 August 2028. Verified as of 4 November 2026.

Who supervises the AI Act in the Czech Republic?

The main supervisory authority is to be the Czech Telecommunications Office, and for general-purpose models the rules are enforced by the EU AI Office. The Czech implementing act was, as of the verification date, still in the legislative process.

Next step

Want an overview of which AI systems run in your company and what follows from them? See how AI Ops works, or book an online call right away and we will go through your agents one by one.

We will go through your AI deployments in 30 minutes. Book an online call

Sources

  • European Commission: AI Act, Shaping Europe's digital future (timetable, categories, AI Omnibus), updated 19 October 2026. digital-strategy.ec.europa.eu
  • AI Act, Article 50 (transparency), text on artificialintelligenceact.eu, opened 4 November 2026. artificialintelligenceact.eu
  • AI Act, Article 99 (penalties), text on artificialintelligenceact.eu, opened 4 November 2026. artificialintelligenceact.eu
  • Lupa.cz: Výtvory umělé inteligence nyní musíte označovat (you must now label AI creations), 3 August 2026. lupa.cz
  • Lupa.cz: ČTÚ rozjíždí dohled nad umělou inteligencí (the Czech Telecommunications Office launches AI oversight), 1 September 2026. lupa.cz

FAQ

Does the AI Act apply to a small company that uses ChatGPT?

Internal use for texts or analyses without deciding about people is typically minimal risk with no special obligations. AI literacy, however, applies to everyone who uses AI systems in their activities, so training people makes sense here too.

Does our chatbot have to introduce itself as AI?

Generally yes, unless it is obvious from the circumstances. The transparency rules have applied since 2 August 2026 and the information should be given at the latest at the first interaction. Who is responsible for it in a specific solution depends on the roles of provider and deployer.

When is an AI agent high risk?

When it decides or substantially influences decisions in sensitive areas, typically recruitment and employee evaluation, credit and access to services, critical infrastructure or education. What decides is the purpose, not the technology. Check the specific classification with a lawyer.

Since when do the obligations for high-risk systems apply?

After the AI Omnibus package, for areas under Annex III from 2 December 2027, and for AI in regulated products from 2 August 2028. Verified as of 4 November 2026.

Who supervises the AI Act in the Czech Republic?

The main supervisory authority is to be the Czech Telecommunications Office, and for general-purpose models the rules are enforced by the EU AI Office. The Czech implementing act was, as of the verification date, still in the legislative process.

AppCast

Czech apps conquering the world

Recipes for successful apps, in Czech. MeguMethod CEO Vratislav Zima talks business with the founders of Czech apps.

Gradient noisy shapeMegumethod CEO Vratislav ZimaMegumethod Scribble