The AI Act: a practical guide for people who build and use AI

Why this guide, and why now

For months, 2 August 2026 was flagged as the date the AI Act would "fully enter into force." That is not quite what happened. In mid-2026, EU institutions approved a simplification package (the so-called "Digital Omnibus") that postponed exactly the heaviest obligations (those on high-risk systems), while leaving other obligations that were already scheduled unchanged.

The practical result: some things are mandatory as of today, others will become mandatory in one or two years, and others still have been in force for a while and are often overlooked simply because they're less well known. This guide separates the three categories and turns each into concrete actions, for people who build AI systems and for people who use them inside an organization.

You won't find an article-by-article legal commentary here: the goal is to understand what to do, not to become an AI law expert.

The logic behind the AI Act, briefly

The regulation doesn't treat artificial intelligence as one single block. It classifies systems by risk level and applies proportionate obligations:

Risk level	Treatment	Example
Unacceptable	Banned outright	Government social scoring of citizens
High-risk	Allowed, with strict obligations (documentation, human oversight, risk management)	Software that screens CVs during hiring
Limited risk	Allowed, with transparency obligations	Customer service chatbot
Minimal risk	No specific obligations	Spam filter, spell checker

A second, equally important axis distinguishes two roles:

Provider: whoever develops an AI system or model and places it on the market, even for free.
Deployer: whoever uses an AI system already developed by someone else, within their own professional activity.

The obligations differ significantly depending on the role. An agency that uses a hiring-screening tool bought from an external vendor is a deployer; the vendor of that tool is the provider. Someone who builds an application on top of a language model's API is, for that application, a provider, even without having trained any model from scratch.

The real timeline, updated

This is the part almost no generic guide gets right, because it changed during 2026.

Date	What happened / what's happening
February 2025	Ban on unacceptable-risk practices (Art. 5) and a general AI-literacy obligation for staff
August 2025	Obligations for providers of general-purpose AI (GPAI) models, and the start of EU-level governance (AI Office)
2 August 2026	Transparency obligations (chatbots, AI-generated or manipulated content, deepfakes); full Commission oversight over GPAI providers; penalties fully applicable
2 December 2026	End of the 4-month technical grace period on watermarking for systems already on the market as of 2 August 2026; separately, end of the transitional period for the two new prohibitions introduced in 2026 (non-consensual intimate imagery generated by AI, child sexual abuse material)
2 December 2027 (postponed from August 2026)	Obligations for stand-alone high-risk systems (Annex III: hiring, credit, justice, education, essential services, etc.)
2 August 2028 (postponed from August 2027)	Obligations for AI embedded in already-regulated products (Annex I: medical devices, machinery, toys, etc.)

What this means in practice: if you build or use a chatbot, or a tool that generates text, images or video today, you are already inside binding obligations. If you build or use a system that decides who gets hired, who gets a loan, or how students are graded, you have more time: that's a postponement, not an exemption, and it's worth preparing regardless.

Prohibited practices (in force since 2025)

Some uses of AI aren't "high-risk to be managed": they are simply banned, regardless of who uses them. Among these:

Subliminal or manipulative techniques that alter a person's behavior in a way that causes them significant harm
Exploiting vulnerabilities related to age, disability, or economic situation
Social scoring that leads to detrimental treatment in contexts unrelated to where the data was originally collected
Emotion recognition in the workplace or in educational institutions (except for medical or safety reasons)
Biometric categorization used to infer racial or ethnic origin, political opinions, or sexual orientation
Since 2026: non-consensual generation of intimate images or videos, and generation of child sexual abuse material

Example for developers: a marketing agency pitching a client a tool that "reads employees' emotions" from webcam footage during meetings is proposing something banned, not merely risky. It needs to be rejected or redesigned, not "managed with a disclaimer."

Example for users: a company that buys HR software assigning candidates a reliability score by cross-referencing social-media data unrelated to the job application falls under the same ban, even though it didn't build the software itself.

Transparency obligations (in force since 2 August 2026)

This is the section most relevant to anyone working with websites, apps, or digital content, because it covers tools already in widespread use.

Three main obligations:

Disclosing interaction with AI: if a user is talking to a chatbot or an automated system, they must be able to tell, unless it's already obvious from context.
Marking AI-generated or manipulated content: images, audio, video, and text must be recognizable as artificial through technical means (marking, metadata) where technically feasible.
Disclosing deepfakes: content that realistically depicts people, places, or events as authentic when they are not must be labeled as such, with exceptions for clearly satirical or artistic works.

Example for developers: if you're building a virtual assistant for a client's e-commerce site based on a language model, you need to make sure the interface clearly communicates that the user isn't talking to a person: a message in the first exchange or a persistent label is enough; silence isn't.

Example for users: a communications agency generating images with AI for an ad campaign must keep the technical marking on the generated content, even after further editing. For images this obligation has no exception tied to editorial review. The "reviewed content with identified editorial responsibility" exception specifically covers AI-generated text on matters of public interest (for example, an AI-assisted editorial piece that's been reviewed and signed off by an accountable person), not images or video.

General-purpose AI models: GPAI (in force since 2025)

This concerns whoever develops "general-purpose" models (the models underlying ChatGPT, Claude, Gemini, and similar), not whoever uses them via an API or interface.

Providers of these models must, among other things: maintain up-to-date technical documentation, provide information to downstream providers who integrate the model, respect copyright with regard to training data, and publish a sufficiently detailed summary of the content used for training. Models classified as carrying "systemic risk" (based on the computing power used for training) face additional obligations: model evaluations, adversarial testing, monitoring of serious incidents, and cybersecurity measures.

Example for developers: if you're fine-tuning an open-weight model and distributing it as a product of your own, you may become a "provider" of that derivative model yourself, with documentation obligations toward whoever uses it.

Example for users: a company integrating a third-party model's API into its own application doesn't carry GPAI provider obligations, but remains responsible, as a deployer, for how that model is used within its product (for instance, regarding the transparency obligations covered in section 5).

High-risk systems (postponed, not eliminated)

These concern AI used for decisions that significantly affect people: hiring, credit assessment, school admissions, critical infrastructure management, justice, border control, and essential insurance or social services.

The main obligations include: a documented risk management system, training data checked for quality and possible bias, effective (not merely formal) human oversight, traceability and logging of decisions, and a conformity assessment before the system is placed on the market.

The postponement to 2 December 2027 (stand-alone systems) and 2 August 2028 (AI embedded in regulated products) buys time, but anyone operating in these sectors would do well not to wait until the last month: the required risk-management systems and documentation aren't built in a few weeks.

Example for developers: a vendor of automated CV-screening software has, right now, the right window to design decision traceability and human-intervention capability in from the start, instead of having to bolt them on in a hurry in 2027.

Example for users: a company evaluating whether to adopt an AI-based credit-scoring system can use this time to contractually require the vendor to provide the compliance documentation it will still have to produce as a deployer.

AI literacy (softened obligation, not gone)

The obligation, in force since 2025, initially required ensuring an adequate level of AI competence among staff who use or develop it. With the 2026 revision, the text was softened: it now requires adopting "supporting measures" for developing this competence: a less stringent commitment, but still present.

Practical example: no formal certification course is required, but the organization needs to be able to show, if asked, that it has done something concrete: an internal session, training material, written guidelines on what staff can and can't do with the AI tools available to them.

Penalties

Penalties have been fully applicable since 2 August 2026 and are proportionate to the severity of the violation:

Up to €35 million or 7% of global annual turnover (whichever is higher) for prohibited practices
Up to €15 million or 3% of turnover for violating other obligations (including those on high-risk systems and GPAI models, where applicable)
Up to €7.5 million or 1% of turnover for providing false or misleading information to authorities

For small and medium enterprises and startups, the penalty applied is the lower of the two thresholds (percentage or fixed amount), not the higher one: a protection specifically designed not to crush smaller businesses.

Operational checklist

If you build AI systems (even just an app using an external API):

Check whether what you're building falls under a prohibited practice: when in doubt, redesign before launch
If the user is interacting with an automated system, make that explicit in the interface
If you generate content (text, images, audio, video), check what technical marking is available and keep it throughout the content's distribution chain
If you operate in a high-risk sector (HR, credit, education, healthcare, justice), start documenting training data, decision criteria, and human-intervention capability now
If you distribute or modify a third party's model, check whether this makes you a "provider" yourself, with documentation obligations
Keep at least minimal internal documentation of AI-related design decisions: in the event of a review, that's what gets asked for first

If you use AI systems in your business (even just a tool bought from a vendor):

Do a real inventory of the AI tools used in the business: not just the ones chosen by IT, but also the ones adopted individually by departments or people
For each tool, check whether the vendor has made the required transparency documentation available
If you use chatbots or automated assistants facing customers, check that they disclose their artificial nature
If your sector falls under high-risk, start asking vendors for compliance documentation contractually, even ahead of the 2027-2028 deadlines
Put in writing even just a few lines of internal guidelines on what staff can do with the AI tools available to them
Update contracts with AI vendors on data access, output ownership, and audit rights