Guides / GUIDE-014

EU AI Act for deployers: what SMBs actually have to do (risk classes, transparency, high-risk duties, dates after the Digital Omnibus)

Sort your AI systems into the AI Act risk classes, meet the transparency and literacy duties that already apply, and prepare for the high-risk deployer duties that start on 2 December 2027.

Most German SMBs do not build AI. They use it: Copilot, DeepL, a chatbot on the website, an applicant tracking system with ranking. In the AI Act's terms that makes them deployers (Betreiber). This guide explains which deployer duties apply today, which apply from 2 December 2027 under the Digital Omnibus (Regulation (EU) 2026/1744), and how to get an inventory, a classification and a short action list in a few days of work.

Why it matters

  • The dates are now fixed. The Digital Omnibus has been in force since 27 July 2026 and amends the AI Act (Regulation (EU) 2024/1689). It moved the duties for high-risk AI listed in Annex III (e.g. HR, credit, education) to 2 December 2027 and for high-risk AI in Annex I products (machinery, medical devices and similar) to 2 August 2028. The Art. 50 transparency duties have applied since 2 August 2026 [1][2][5].
  • Some duties already apply. The bans in Art. 5 have applied since February 2025. The duty to support AI literacy (Art. 4) has applied since 2025 and in its amended form since 27 July 2026: you must take measures to support your staff's AI literacy, not guarantee a level (see GUIDE-006) [5].
  • There is an authority. Since the KI-MIG came into force on 29 July 2026, the Bundesnetzagentur is Germany's central market surveillance authority, contact point and complaints office for the AI Act. It supervises HR, education and critical infrastructure, as well as transparency duties and prohibited practices in its remit. BaFin covers the financial sector and the state media authorities cover media [3].
  • Fines are real but capped for SMBs. Breaching deployer duties (Art. 26) or transparency duties (Art. 50) can cost up to EUR 15 million or 3% of worldwide turnover. Prohibited practices can cost up to EUR 35 million or 7%. For SMEs, the lower of the two amounts applies [9].

When you need it

  • Now: check for prohibited practices, label deepfakes and published AI text, and document your literacy measures.
  • Before buying or building any AI tool (GUIDE-002, GUIDE-010): classify it first. A "high" result changes the contract, the effort and the timeline.
  • By mid-2027: if you use or plan Annex III high-risk AI, start preparing; this takes months.
  • Whenever you change the purpose of a tool. Using a general chatbot to rank applicants turns it into a different system under the AI Act (see Step 6).

Who is involved

Role Part
Management Owns the duties, approves the inventory and the use or non-use of high-risk AI.
AI lead / IT manager Maintains the AI inventory, classifies systems and collects provider documentation.
Legal / compliance (often external) Checks borderline classifications and contracts.
DPO Links AI Act duties to the GDPR: DPIA (Art. 35), AVV (Art. 28), information duties (GUIDE-005).
Works council Must be informed before high-risk AI is used at the workplace (Art. 26(7)) and usually has co-determination rights under §87(1) no. 6 BetrVG (GUIDE-004).
HR and department leads Own the HR and customer processes where most high-risk and transparency cases sit.
Marketing / communications Responsible for labelling AI-generated images, video, audio and published text.

Step by step

1. Build an AI inventory (1-2 days). List every AI system in use: standalone tools, AI features in standard software (M365, CRM, ATS, DMS, accounting) and shadow AI found via a staff survey (GUIDE-007). Record purpose, department, provider, data used, people affected, and whether you bought, configured or built it.

2. Check your role per system. You are a deployer if you use a system under your authority for professional purposes. You become a provider, with much heavier duties, if you put your name or trademark on a high-risk system, make a substantial modification to one, or change the purpose of a non-high-risk system (including a general-purpose model) so that it becomes high-risk [7]. Building your own chatbot on an LLM API usually makes you its provider for Art. 50(1).

3. Classify each system into one of four classes. The Bundesnetzagentur's KI-Service Desk offers a free online tool for a first risk assessment [3].

Class Typical SMB examples What you must do
Prohibited (Art. 5) Emotion recognition of employees or applicants (except for medical or safety reasons), biometric categorisation by religion, union membership or similar, manipulative techniques that cause significant harm [10] Do not use. Check vendor features, e.g. "sentiment" or "engagement" scoring of staff in video calls or call centres.
High-risk (Annex III) CV screening and candidate ranking, AI interviews, performance evaluation, decisions on promotion or termination, task allocation based on behaviour, credit scoring of private persons, life and health insurance pricing Art. 26 duties from 2 Dec 2027 (Step 5).
Transparency (Art. 50) Website chatbot, voicebot, AI-generated marketing images or video, voice cloning, AI avatars, AI text published on matters of public interest Disclosure and labelling (Step 4), applying now.
Minimal Translation, drafting internal e-mails, coding assistants, invoice capture, forecasting, quality inspection Art. 4 literacy only, plus GDPR and the BetrVG as usual.

Most SMB inventories end up "minimal" or "transparency". High-risk cases cluster in HR and, for financial firms, credit and insurance.

4. Meet the transparency duties (now). - Chatbots and voicebots: people must know they are dealing with AI unless this is obvious. This is a provider duty (Art. 50(1)). Check that your vendor's product shows the notice. If you built the bot yourself, add it yourself [6]. - Deepfakes: if you publish AI-generated or manipulated images, audio or video that look like real people, objects or places, disclose it. For evidently artistic, satirical or fictional work, a discreet notice is enough [6]. - Published text: AI-generated text published to inform the public on matters of public interest must be labelled. This does not apply if a human reviewed it and someone holds editorial responsibility. A press release you edit and approve is covered by that exception. An unreviewed, automated news feed is not [6]. - Emotion recognition or biometric categorisation (where lawful at all): inform the people exposed to it [6]. - How and when: clearly and at the latest at the first interaction or exposure [6]. The Commission's final Code of Practice of 10 June 2026 describes practical labels, including a common EU icon and audio disclaimers [4]. Machine-readable watermarking (Art. 50(2)) is your vendor's job. Systems placed on the market before 2 August 2026 have until 2 December 2026 [2].

5. Prepare the high-risk deployer duties (due 2 Dec 2027). Art. 26 requires you to [8]: 1. Use the system according to the provider's instructions for use, through technical and organisational measures. 2. Assign human oversight to named people with the necessary competence, training, authority and support. 3. Ensure that input data under your control is relevant and sufficiently representative. 4. Monitor operation, inform the provider of problems, and suspend use and report to the provider and authority if the system presents a risk. Report serious incidents immediately. 5. Keep automatically generated logs under your control for at least six months, unless data protection law requires otherwise. 6. Inform the workers' representatives and affected employees before putting the system into service at the workplace. 7. Inform people who are subject to decisions made or assisted by the system, e.g. applicants. 8. Use the provider's information to carry out the DPIA.

A fundamental rights impact assessment (Art. 27) is required only from public bodies, private entities providing public services, and deployers of credit scoring of individuals or of life and health insurance pricing. It can reuse your DPIA [11]. Public authorities must also check that the system is registered in the EU database [8].

Budget one to three days of documentation per high-risk system, plus works council negotiation.

6. Govern changes of purpose. Write into your AI policy (GUIDE-007) that staff may not use general tools such as ChatGPT or Copilot to rank applicants, score employees or decide on credit. Such use would turn you into the provider of a high-risk system [7].

7. Fix it in contracts. For high-risk systems, require from the vendor: instructions for use, a statement on the Art. 6 classification, log access and export, a support contact for incident reporting, and a written commitment to AI Act conformity by 2 Dec 2027 (GUIDE-010).

8. Review twice a year, and whenever vendors ship new AI features (GUIDE-011).

Checklist

  • [ ] AI inventory exists, including AI features in standard software and shadow AI
  • [ ] Role (deployer or provider) recorded per system
  • [ ] Risk class recorded per system, with a one-line justification
  • [ ] No prohibited practices in use (emotion recognition of staff checked)
  • [ ] Chatbots and voicebots show an AI notice
  • [ ] Process for labelling deepfakes and published AI text, including the EU icon
  • [ ] AI literacy measures documented (GUIDE-006)
  • [ ] For high-risk systems: oversight persons named, log retention of 6 months or more, works council and staff informed, affected persons informed, DPIA done
  • [ ] FRIA done if you are a public body or do credit or insurance scoring
  • [ ] Policy forbids repurposing general AI for HR, credit or similar decisions
  • [ ] Vendor contracts cover AI Act documentation and incident support

Templates & tools

  • Bundesnetzagentur KI-Service Desk: free risk-class check, guidance and AI literacy resources, aimed at SMEs [3].
  • EU AI Act Service Desk: article texts including Omnibus amendments [6][7][8].
  • Code of Practice on marking and labelling AI-generated content (Commission, June 2026) [4].
  • IHK resources: IHK Ostthüringen/Gera collects IHK help on the AI Act, including a model AI policy (IHK Schwaben) and guidance on transparency duties (IHK Regensburg) [12].
  • GDPR templates: DSK lists for mandatory DPIAs and the AVV checklist (GUIDE-005).

Common pitfalls

  • Panic or complacency. "Everything is high-risk" paralyses. "Only for big tech" ignores Art. 4, 5 and 50.
  • Treating 2027 as far away. Works council agreement, DPIA and vendor changes for an HR tool take months.
  • Hidden HR AI. Ranking features in ATS platforms or "productivity insights" in collaboration suites can be high-risk or even prohibited. Ask vendors directly.
  • Accidental provider status. Your own label on a vendor's HR tool, or a self-built screening bot, puts provider duties on you [7].
  • Forgetting the GDPR and the BetrVG. The AI Act adds to them and does not replace them. §87(1) no. 6 BetrVG applies to most AI that logs user activity, whatever its risk class (GUIDE-004).

Sources

  1. Regulation (EU) 2026/1744 (Digital Omnibus on AI), Official Journal. https://eur-lex.europa.eu/eli/reg/2026/1744/oj. Accessed 2026-10.
  2. Gibson Dunn: EU AI Act Omnibus Agreement, Postponed High-Risk Deadlines and Other Key Changes. https://www.gibsondunn.com/eu-ai-act-omnibus-agreement. Accessed 2026-10.
  3. Bundesnetzagentur: Bundesnetzagentur übernimmt zentrale Rolle bei der Umsetzung der KI-Verordnung (press release, 29.07.2026). https://bundesnetzagentur.de/SharedDocs/Pressemitteilungen/DE/2026/20260729_KI_VO.html. Accessed 2026-10.
  4. Jones Day: European Commission Publishes Final Code of Practice on Marking and Labelling AI-Generated Content (June 2026). https://www.jonesday.com/en/insights/2026/06/european-commission-publishes-final-code-of-practice-on-marking-and-labelling-aigenerated-content. Accessed 2026-10.
  5. Regulation (EU) 2024/1689 (AI Act), consolidated text. https://eur-lex.europa.eu/eli/reg/2024/1689/oj. Accessed 2026-10.
  6. EU AI Act Service Desk: Article 50, Transparency obligations. https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-50. Accessed 2026-10.
  7. EU AI Act Service Desk: Article 25, Responsibilities along the AI value chain. https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-25. Accessed 2026-10.
  8. EU AI Act Service Desk: Article 26, Obligations of deployers of high-risk AI systems. https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-26. Accessed 2026-10.
  9. EU AI Act Service Desk: Article 99, Penalties. https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-99. Accessed 2026-10.
  10. EU AI Act Service Desk: Article 5, Prohibited AI practices. https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-5. Accessed 2026-10.
  11. EU AI Act Service Desk: Article 27, Fundamental rights impact assessment. https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-27. Accessed 2026-10.
  12. IHK Ostthüringen zu Gera: Fokus KI-Verordnung, weitere Informationen. https://www.ihk.de/gera/recht-und-steuern/fokus-ki-verordnung/weitere-informationen-7135530. Accessed 2026-10.