This sheet is past its review date. Deadlines or advice may have changed since 4 Nov 2026.
AI Act: what do you owe for each AI system, and by when?
This quarter, list every AI system you use, build or sell with your role and its risk class, and check that none is banned and that chatbots and AI content are disclosed as required since 2 August 2026. Then prepare the duties for any Annex III use, such as hiring or credit scoring, before they apply on 2 December 2027.
If you only do five things
Train staff by role and keep a record.
Since July 2026 you must support AI literacy, not guarantee a level2. Give everyone the basics and people who configure or oversee AI more, and keep attendance lists.
Classify one AI system
Is it an AI system at all?
Software that only runs rules people defined, plain spreadsheets and forecasts that just use averages are not6. Mark them out of scope.
Which role do you hold?
Provider if you develop it, or have it developed, and supply or run it under your name; deployer if you use it1. Rebranding or substantially modifying a high-risk system, or turning a general-purpose one to a high-risk use, makes you its provider1.
Is the use banned or high-risk?
Check Article 5 and Annex III. An Annex III system isn't high-risk if it only does a narrow procedural or preparatory task, improves finished human work or flags deviations for human review, unless it profiles people; the provider documents this and registers the system1.
Does a disclosure duty apply?
Chatbots, generated content, deepfakes and emotion recognition each carry one1. Record the result, the reasoning and a review date in the inventory.
What each role owes
Deployer dates are also on the agentic AI sheet.
Deployers (most organizations)
- High-risk, from 2 December 2027. Follow the instructions for use, assign trained oversight staff with authority, keep logs at least 6 months, report serious incidents to the provider and the authority, and inform workers and the people affected1.
- Impact assessment. Public bodies, providers of public services and users of credit scoring or life and health insurance pricing assess fundamental rights impacts before first use and notify the authority1.
Providers of AI systems
- Chatbots and generative AI. Say it is AI unless that is obvious. Mark generated content machine-readably; systems on the market before 2 August 2026 have until 2 December 202612.
- High-risk, from 2 December 2027. Risk management, data governance, documentation, logging, human oversight, accuracy and cybersecurity, a quality management system, conformity assessment, CE marking and registration1; from 2 August 2028 in products under EU safety law2.
- Systems already on the market. Covered only after a significant design change, or by 2 August 2030 if public authorities use them2.
- Incidents. Report serious ones within 15 days, 10 days after a death and 2 days for a widespread infringement or critical infrastructure disruption1.
Providers of general-purpose AI models
- Who. Indicatively, providers of models trained with over 10²³ FLOP that generate language, images or video, and anyone fine-tuning one with over a third of its original training compute7.
- Duties since 2 August 2025. Documentation, information for downstream providers, a copyright policy and a public training-content summary; open-source models skip the first two unless systemic. Models placed earlier comply by 2 August 20271. The Commission deems the Code of Practice an adequate voluntary way to show compliance8.
- Systemic risk. Presumed above 10²⁵ FLOP: evaluations with adversarial testing, risk mitigation, incident reporting and cybersecurity1.
What changed in 2026
- The Digital Omnibus on AI took effect on 27 July 2026. It moved Annex III high-risk duties from 2 August 2026 to 2 December 2027, softened AI literacy, added the intimate-deepfake ban and extended SME relief to small mid-caps21.
- Transparency guidance arrived. A code of practice of 10 June covers marking and labeling4; draft guidelines approved on 20 July give examples3.
- The first harmonized standard was published. Since October 2026, EN 18286 gives high-risk providers a presumption of conformity for quality management9.
Who enforces it, and in Germany
- Fines scale with the duty. Up to €35 million or 7% of worldwide turnover for banned practices, €15 million or 3% for most other duties and €7.5 million or 1% for misleading authorities1.
- The AI Office handles general-purpose AI. It can fine model providers up to 3% or €15 million since 2 August 2026, and supervises most systems built on their own models12. National authorities do the rest1.
- Germany: the Bundesnetzagentur. Since 29 July 2026 it is the default market surveillance authority, contact point and complaints office; BaFin covers AI tied to regulated financial activities10.
- Germany: the works council. Inform it in time about planned AI use; it co-determines tools able to monitor behavior or performance1112.
Where the AI Act meets the GDPR and the CRA
- Build on the DPIA. High-risk deployers use provider information for their DPIA1, and the fundamental rights impact assessment may cross-reference it2.
- Ask where the model came from. Models trained on personal data aren't always anonymous. Check that a model wasn't developed unlawfully, starting with its data sources and any regulator findings13.
- Bias testing has a legal basis. Special categories of personal data may be processed where strictly necessary to find and correct bias, under safeguards such as pseudonymization2.
- CRA compliance can cover AI cybersecurity. A high-risk system whose CRA declaration of conformity shows it meets the CRA's essential requirements is deemed to meet the AI Act's cybersecurity requirement2.
What you can skip
- ISO/IEC 42001 certification as AI Act proof. Only EN 18286, which can be layered on ISO/IEC 42001, gives a presumption of conformity for a quality management system9. Deployers need neither.
- Inventorying macros and fixed-rule workflows. They aren't AI systems6.
- AI literacy exams and certificates. The law asks for measures, not a guaranteed level2.
- Labeling internal drafts. Posts on internal networks and professional correspondence aren't published3.
Mistakes that cost the most
- Building CV screening on a general assistant. It can make you the provider of a high-risk system, with every provider duty1.
- Emotion analytics on staff. Inferring employees' or candidates' emotions from calls or meeting video is banned, even during probation5.
- Relying on the vendor's compliance. Oversight, logs and worker information are deployer duties1. Put instructions for use, log access and incident notice in the contract.
What this looks like at your size
Small team, no security staff
- Keep the inventory in a spreadsheet: tool, owner, purpose, data, role and risk class.
- Publish a one-page AI policy (see the AI inputs sheet) and run a short AI basics session each year.
- In Germany, use the guidance the Bundesnetzagentur provides for SMEs and start-ups10.
Growing, with IT or security staff
- Name one owner for AI governance and send new AI uses through a short intake form.
- Run the AI Act check, the DPIA and the vendor review as one workflow.
- Map Annex III uses and plan oversight training before December 2027.
Enterprise or regulated
- Run AI governance in an existing management system such as ISO/IEC 42001, and add EN 18286 for high-risk systems you provide9.
- Prepare fundamental rights impact assessments for credit scoring, insurance pricing and public services1.
- Track fine-tuning compute against the one-third threshold for general-purpose models7.