Sign in Subscribe

Disclaimer: These are my research notes, not legal advice. I'm not a lawyer, and I can't guarantee that they are complete, correct or current. For a binding assessment, ask a qualified lawyer or the relevant supervisory authority.

Updated 26 Sep 2026: the full list of bans, the fines, the Commission's transparency guidelines, Germany's supervisory authority and the status of the Council of Europe convention. Deployer duties and the Swiss scope example are corrected.

First published in January 2026, updated to the state of 26 September 2026. It's written for developers, cloud architects and the people responsible for AI and compliance in Germany and Switzerland.

The short version: five dates and one rule for Swiss companies

The EU AI Act is partly in force. Its bans have applied since February 2025, the rules for general-purpose AI (GPAI) since August 2025 and the transparency rules since August 2026. The high-risk rules were postponed to December 2027 (Annex III) and August 2028 (products under Annex I).

Switzerland has no horizontal AI act and doesn't plan one right now. It relies on the Council of Europe's Framework Convention and on existing law: the data protection act and sector rules such as FINMA guidance for the financial sector. A consultation draft is due by the end of 2026.

For Swiss companies, the main point is that the EU AI Act applies to them as soon as their AI systems, or what those systems produce, are used in the EU.

What applies now, and what waits until 2027 and 2028

The AI Act (Regulation 2024/1689) works like a product safety law: it sorts AI systems by risk. The Digital Omnibus on AI (Regulation 2026/1744, in force since 27 July 2026) amended it: later high-risk deadlines, a lighter AI literacy duty, two new bans, simpler documentation and lower fine caps for SMEs and small mid-caps, and more supervision by the AI Office. The GPAI obligations stayed as they were.

Already in force

  • Prohibited practices (since 2 Feb 2025): "unacceptable risk" systems are illegal, from social scoring to emotion recognition at work. The full list is under Risk classes below.
  • AI literacy (since 2 Feb 2025): Article 4 originally required providers and deployers to ensure a sufficient level of AI literacy. Since the Omnibus they must take measures to support it, without guaranteeing a level for each person. Documented training is still the practical answer.
  • General-purpose AI (since 2 Aug 2025): providers of GPAI models (OpenAI, Anthropic, Google and others) must meet transparency and copyright obligations and publish summaries of their training data. Models with "systemic risk" also need adversarial testing and incident reporting. The voluntary GPAI Code of Practice (July 2025) shows one way to comply. Since 2 Aug 2026 the Commission can enforce these rules, with fines of up to 3% of global turnover. Since the Omnibus, the AI Office also supervises AI systems that a provider builds on its own GPAI model, and AI systems in very large online platforms and search engines.
  • Transparency (since 2 Aug 2026): the Article 50 duties apply, for example telling people they're talking to a chatbot. The Commission published guidelines on Article 50 on 20 July 2026 and assessed the Code of Practice on AI-generated content as a suitable way to comply. Systems already on the market before 2 Aug 2026 have until 2 Dec 2026 to mark generated content in a machine-readable way. Content generated before 2 Aug 2026 doesn't have to be marked after the fact.

Next deadlines

  • 2 Dec 2026, new bans: the Omnibus adds AI systems that generate child sexual abuse material or non-consensual intimate images to Article 5 (Freshfields summary). Providers need safeguards such as refusal training and output filters by then.
  • 2 Dec 2027, high-risk systems (Annex III): originally 2 Aug 2026. This is the deadline that matters for most enterprise AI. Systems used in HR (recruitment), critical infrastructure, credit scoring or education need a formal conformity assessment, a CE marking and an entry in the EU database. Post-market monitoring (Article 72) is part of the same package and applies with it.
  • 2 Aug 2028, high-risk products (Annex I): AI as a safety component of products such as medical devices or toys, originally 2 Aug 2027. AI in machinery is largely taken out of the AI Act's high-risk rules; the Commission can set AI-specific requirements under the Machinery Regulation instead.

In the January version I wrote that post-market monitoring applied from 2 Feb 2026. That was wrong: 2 Feb 2026 was the Commission's deadline for a monitoring template, not an obligation for providers.

If you use a prohibited practice, you're already in breach. If you're building a high-risk system, use the extra time to finish the technical documentation and the quality management system (QMS).

Fines

Set in Article 99, each time the higher of the two amounts:

  • Prohibited practices: up to €35 million or 7% of global annual turnover
  • Most other obligations (high-risk, transparency): up to €15 million or 3%
  • Wrong or misleading information to authorities: up to €7.5 million or 1%

For SMEs and start-ups the lower of the two amounts applies, and the Omnibus added lower caps for small mid-caps. GPAI providers are fined by the Commission under Article 101 (up to €15 million or 3%).

Your risk class decides how much work it is

The risk class follows from how you design and use the system, and decides whether you need a full QMS or a notice to users. For a first classification, use the EU AI Act Explorer. The definitions are in Article 3.

Unacceptable risk (prohibited), Article 5

Banned. Legal text: Article 5.

  • Manipulation: subliminal or deceptive techniques that distort behaviour and cause significant harm
  • Exploiting vulnerabilities: targeting age, disability or social and economic situation
  • Social scoring: rating people on social behaviour or personal traits, with unjustified or disproportionate consequences
  • Predictive policing: predicting that a person will commit a crime based solely on profiling or personality traits
  • Facial image scraping: building facial recognition databases by untargeted scraping from the internet or CCTV
  • Emotion recognition: inferring emotions at work or in education, except for medical or safety reasons
  • Biometric categorisation: using biometric data to infer race, political opinions, trade union membership, religious beliefs, or sex life and sexual orientation
  • Real-time remote biometric identification in public spaces by the police, with narrow exceptions (for example searching for victims or preventing an imminent threat), each needing prior authorisation
  • Sexual abuse and intimate imagery: generating CSAM or non-consensual intimate content (from 2 Dec 2026)

High risk, Annex III

Allowed, with strict obligations. The full list is in Annex III.

  • Biometrics: remote identification (where not banned), emotion recognition outside work and education
  • Critical infrastructure: safety components in water, gas, electricity and road traffic
  • Education: admission, grading, detecting cheating in exams
  • Employment: recruitment screening, task allocation, performance evaluation
  • Essential services: credit scoring, eligibility for benefits, life and health insurance pricing
  • Law enforcement: assessing the reliability of evidence, risk assessments that aren't based solely on profiling
  • Migration and border control: assessing visa and asylum applications
  • Justice and democracy: supporting judges, influencing elections

Limited risk, Article 50

People must know they're dealing with AI. See Article 50.

  • Chatbots must say that they are AI.
  • Synthetic content: providers mark generated images, audio, video and text in a machine-readable way.
  • Deepfakes and published text: deployers label deepfakes, and AI-generated text on matters of public interest unless a person reviewed it and takes editorial responsibility.

Minimal risk

Most AI systems, such as spam filters. The AI Act doesn't regulate them; the GDPR still applies.

Your role decides your obligations

Your legal role decides your obligations: following the instructions for use, or running a full compliance programme. The Small Businesses Guide to the AI Act explains the roles in detail.

Provider (Anbieter)

You develop an AI system, or have one developed, and put it on the market under your own name or trademark. Examples: OpenAI with its GPT models, or a software vendor selling its own HR analytics platform.

For high-risk systems, a provider must:

  • run a quality and risk management system (Art. 9, 17)
  • govern training, validation and test data to reduce bias (Art. 10)
  • keep full technical documentation and automatic logs (Art. 11–12)
  • complete the conformity assessment, register in the EU database and affix the CE marking

Deployer (Betreiber)

You use an AI system under your own authority in a professional context. Examples: a company using ChatGPT for internal knowledge management, or calling a vendor's credit scoring API.

Every deployer must:

  • take measures for the AI literacy of the staff who use the system (Art. 4)
  • meet the transparency duties of Art. 50, for example labelling deepfakes

For high-risk systems, a deployer must also (Art. 26):

  • use the system as the provider's instructions say
  • put human oversight in place, with people who have the competence and authority for it (Art. 14)
  • make sure input data under its control is relevant and representative
  • keep the logs and inform the provider about serious incidents
  • inform workers' representatives and the affected employees before using it at work
  • run a fundamental rights impact assessment first if it is a public body, provides public services, or uses AI for credit scoring or life and health insurance pricing (Art. 27)

In the September version I listed the high-risk duties as if they applied to every deployer. They don't.

Other roles

  • Authorised representative (Bevollmächtigter): represents a non-EU provider in the EU and is legally responsible for its compliance.
  • Importer (Importeur): brings an AI system from outside the EU (for example Switzerland) onto the EU market.
  • Product manufacturer (Produkthersteller): builds a third-party system such as an LLM into its own product and sells it under its own name. The law treats it as the provider.

The trap: becoming a provider without noticing

If you put your name or brand on a high-risk system, substantially modify one, or change the purpose of a system (a GPAI model, for example) so that it becomes high-risk, you move from deployer to provider (Art. 25). You then carry the full provider load. The IHK München guide goes into more detail.

Note for automation work: if you build a workflow for a client (n8n plus OpenAI, say), agree in the contract who the provider is. If you deliver it as your own product, the provider liability is yours.

Switzerland has no AI act, but it has rules

Switzerland has no single AI act, but it has rules. It regulates by sector and by principle, which makes compliance more fragmented.

What applies now

  1. Federal Act on Data Protection (FADP/DSG): revised version in force since 1 September 2023, technology-neutral and directly applicable to AI. The FDPIC (EDÖB) confirmed this on 8 May 2025: disclose the purpose, functioning and data sources of AI processing, tell people when they're talking to a machine, run a data protection impact assessment for high-risk processing, and let people have automated individual decisions reviewed by a human. The FDPIC considers real-time facial recognition and social scoring incompatible with the FADP.
  2. FINMA (financial sector): Guidance 08/2024 (18 December 2024) requires governance, risk management and data quality controls for AI, and makes clear that existing supervisory law applies fully to AI.
  3. Council of Europe Framework Convention: signed by Switzerland on 27 March 2025, not yet ratified. It sets binding principles on human rights, democracy and the rule of law in AI, without the EU's technical requirements. The EU ratified it on 15 May 2026, as the first party. The convention itself isn't in force yet: it needs five ratifications, three of them by Council of Europe member states. Until Switzerland ratifies, it isn't part of Swiss law.

What's coming

On 12 February 2025 the Federal Council decided to ratify the convention and adapt existing laws by sector instead of writing an AI act. The Federal Office of Justice is preparing the consultation draft for the end of 2026; as of September 2026 the consultation hasn't opened. A message to Parliament would follow in 2027 at the earliest. The draft is meant to cover:

  • transparency
  • data protection
  • non-discrimination
  • oversight

Alongside it, OFCOM (part of DETEC) is preparing non-binding measures such as industry solutions and voluntary commitments, also by the end of 2026.

EU and Switzerland compared

EU AI Act Switzerland (current)
Approach One horizontal law Sector rules plus existing laws (FADP)
Type Product safety law (CE marking) Principles and risk management
Bans Statutory list (Art. 5) No AI-specific list; the FDPIC considers real-time facial recognition and social scoring incompatible with the FADP
Enforcement National authorities and the AI Office Existing bodies (FDPIC/EDÖB, FINMA)
High-risk requirements Prescriptive (QMS, logging, accuracy metrics) Risk-based due diligence (governance)

When EU law reaches Swiss companies

The EU AI Act applies to providers that place an AI system on the EU market or put it into service in the EU, wherever they are based. It also applies to providers and deployers outside the EU when the output of their AI system is used in the EU (Art. 2).

  • A Swiss software vendor sells an AI-based HR tool to a German client. The vendor must comply with the AI Act in full, as a provider.
  • A Swiss company runs a credit scoring model in Switzerland and delivers the scores to lenders in the EU. The system runs outside the EU, but its output is used there, so the Act applies. Where the people being scored come from doesn't decide it.

The provider trap above applies to Swiss developers in the same way: sell an OpenAI-based system under your own brand in the EU, and you carry a provider's full obligations.

Five steps in the EU, three in Switzerland

Companies in the EU (Germany)

  1. Audit: make sure you use none of the prohibited practices, such as emotion recognition at work.
  2. AI literacy: document your training to meet Article 4.
  3. High risk: start the conformity assessments now, even with the December 2027 deadline. The Annex IV documentation is extensive and takes time.
  4. Transparency: check chatbots and generated content against Article 50 and the Commission's guidelines, which apply already.
  5. Know your authority: the Bundestag passed the implementing law on 11 June 2026. It makes the Bundesnetzagentur the market surveillance authority for most AI systems.

Companies in Switzerland

  1. Plan for both regimes: if you sell into the EU or your AI output is used there, build for the EU AI Act and the FADP. Treat the EU Act as the engineering baseline.
  2. FINMA: financial institutions should check their AI governance against Guidance 08/2024.
  3. Label AI by default: mark chatbots and generated content. That meets Article 50 and the FDPIC's transparency expectations.

Checklist

Based on the IHK München checklist:

  1. Inventory: list every AI system in use, with where and what for.
  2. Classify: place each system in a risk class (unacceptable, high, limited, minimal) to see which obligations apply.
  3. Know your role: decide per system whether you are provider or deployer, and check the provider trap.
  4. Assign owners: name who is responsible for running, monitoring and keeping each system compliant.
  5. Train people: take and document measures so the relevant staff can use AI professionally (required since 2 Feb 2025, softened by the Omnibus).
  6. Check transparency: add the required notices, such as "I am an AI" labels or machine-readable marks on synthetic content.
  7. Review regularly: audit your AI systems on a fixed cycle and follow changes in the law. This article needed two updates within eight months.

Sources

Find out if the EU AI Act applies to your Swiss company (2026)

Know whether the EU AI Act applies to your Swiss company and what to do first, with a checklist. Where the Act stands in September 2026, and how Switzerland regulates AI without an AI act. My research notes, not legal advice.