AI · AI at work
The EU AI Act in November 2026: What changes on 2 December
Who explains a chatbot, who labels a deepfake, and which older systems face the December deadline? A practical guide to assigning AI Act duties at work.

In this article
Reviewed on 4 November 2026. A retailer buying a customer-service chatbot and the company supplying that system have different duties. That distinction matters ahead of 2 December. The transition period concerns a technical requirement for certain providers. It does not postpone information that people should already receive when they open a chat.
Which date applies to your use case?
The dates come from the regulation as amended in July. The Commission has also published guidelines on the transparency obligations. Consolidated AI Act · Digital Omnibus on AI · Article 50 guidelines
Three questions before you sign off
What does the customer see at first contact? A provider must design a system that interacts directly with people so they are informed they are dealing with AI. An exception applies if that is already obvious under the conditions set out in the law. Open the actual support chat, including on a phone. A clear notice at the top reaches the customer before they describe their problem.
What survives in the file? Providers of systems generating synthetic content must implement applicable technical marking and enable detection. There are exceptions for standard editing assistance and changes that do not substantially alter the input or its meaning. Ask an image provider for an original file from the production system. Find out which markings survive the export you use. A disclosure on the product page and a machine-readable mark in the file serve different purposes. EU code on marking and labelling
What do you publish yourself? A deployer using a deepfake generally has to disclose its artificial origin. Artistic and satirical works have specific disclosure arrangements. For text on matters of public interest, human review and editorial responsibility also matter. Record who checked and approved the piece. The exception for reviewed text does not automatically extend to an image or a voice recording.
What AI literacy looks like in practice
Article 4 requires measures to support AI literacy. A support team reviewing suggested replies needs different skills from a recruitment team screening candidates. Give the support team a plausible but invented returns policy. Can they find the mistake in the approved policy and correct the reply before it goes out? Record who practised and who handles uncertain cases. Experience, people affected and the setting all shape what is appropriate. Commission questions and answers on AI literacy
Who can produce the evidence?
For each use, start with its purpose, the people affected, your role, a named owner and a piece of evidence someone else can inspect. For a chatbot that could be the first screen. For a synthetic video, keep the version actually published. If buying a potentially high-risk system, ask the supplier to explain its classification and provide the relevant documentation. The deployer needs to decide who oversees its operation and who can intervene. The provider ordinarily handles conformity assessment. A substantial modification or sale under your own name can change that division of responsibilities.
Germany’s Federal Network Agency provides a central complaints service. Assessing a specific product or deployment calls for its precise intended use. Federal Network Agency complaints service · A practical guide to AI adoption at work