# AI Act: how artificial intelligence affects startups in Europe

> The AI Act now requires European startups to comply with rules by risk level. Deadlines, fines and support for SMEs, explained without jargon.
> By José Maldonado · Artificial Intelligence · 2026-09-25

A six-person startup in Lisbon could be breaking a law that almost no one on its team has read. The AI Act, the European regulation on artificial intelligence, entered into force on 1 August 2024 and applies in phases: the obligations depend on what each system does, not on the size of the company using it. That leaves many founders in an uncomfortable zone, because they believe the rule is a matter for giant laboratories and, at the same time, they are not sure whether their chatbot falls within its scope.

The timetable shifted just before summer 2026. The AI Digital Omnibus, [Regulation (EU) 2026/1744](https://eur-lex.europa.eu/eli/reg/2026/1744/oj/eng), was published in the Official Journal on 24 July 2026 and entered into force on 27 July, six days before the date that had half the sector worried. Its most talked-about change was to delay the obligations for high-risk systems in Annex III until 2 December 2027, and those for regulated products in Annex I until 2 August 2028.

What was not delayed weighs just as much. The transparency rules of Article 50 have been in force since 2 August 2026, and the European AI Office has already made its move: on 29 August it confirmed the first requests for information to model providers, including OpenAI, Anthropic, Google and Mistral. Responding badly, incompletely or misleadingly can cost 15 million euros or 3% of global turnover, whichever is higher, a genuine outrage for small companies.

## What is the AI Act in plain words?

The regulation classifies each system by the risk it generates, not by the technology it uses. There are four tiers. At the lowest level sit everyday uses, such as a text corrector or a product recommender, which barely carry obligations. Above that comes limited risk, where chatbots and content generators sit, and there the rule demands transparency: informing the user and labelling what has been created with AI.

Then comes high risk, which is what complicates life for lawyers. Article 6 defines it by intended use, and Annex III lists the fields where it applies:

*   Staff selection and evaluation.
*   Credit scoring and eligibility for essential services.
*   Biometric identification and categorization of persons.
*   Educational assessment and exam monitoring.
*   Management of critical infrastructure and certain police uses.

## The paradox of models

The same model can be limited risk if it summarizes emails and high risk if it decides who gets hired. That dual face explains why classification is the first task of any team, before spending a single euro on consultancy.

Level

Examples

What the AI Act requires

Unacceptable

Harmful manipulation, social scoring

Directly prohibited

High

Staff selection, credit scoring

Risk management, documentation, human oversight

Limited

Chatbots, generated content

Transparency, inform and label

Minimal

Spam filters, recommenders

No specific obligations

Most of a startup's tools do not fall into the toughest tier. The European Commission estimated at the time that between 5% and 15% of systems would be high risk, and industry surveys have disputed that figure. For a small company, the sensible thing is to draft a one-page document setting out what it does with AI and why it believes its use is not high risk. That document, with a date, is worth more than any purchased template.

The logic is similar to a driving licence: it does not ban the car, it requires that whoever drives it knows the rules and answers if they cause harm. No one asks a bicycle to pass the technical inspection, and the same applies to a minimal-risk system.

## What did the AI Digital Omnibus change in 2026?

However, this is only part of the story. The amendment signed in July delayed the bulk of the heavy work, and is known as the Digital Omnibus. The text was published in the Official Journal on 24 July 2026 and entered into force three days later, according to the analysis by [White & Case](https://www.whitecase.com/insight-alert/eu-ai-omnibus-enters-force-amending-ai-act). Its main effect was to move the obligations for high-risk systems in Annex III from 2 August 2026 to 2 December 2027, and those for AI embedded in regulated products to 2 August 2028.

Date

What happens

1 Aug 2024

The AI Act (Regulation 2024/1689) enters into force

2 Feb 2025

Prohibitions and basic AI training

2 Aug 2025

Obligations for general-purpose models

2 Aug 2026

Article 50 transparency, not delayed

2 Dec 2026

Labelling of generated content

2 Dec 2027

Annex III high risk, previously 2 Aug 2026

2 Aug 2028

High risk embedded in products (Annex I)

The delay does not amount to a softer law. Brussels kept the Article 50 transparency rules intact and added a new prohibition against applications that generate intimate images without consent, according to [Usercentrics](https://usercentrics.com/knowledge-hub/eu-ai-act-high-risk-delay-article-50-transparency-consent/). The postponement responds to industry pressure and the lack of harmonized technical standards, not to a change of philosophy.

One point that comes up repeatedly in industry conversations deserves qualification. The Omnibus does not reduce the number of systems covered; it only shifts the date by which compliance must be demonstrated. For a startup already selling to hospitals or banks, the relief is real, because it gains almost a year and a half to document. For anyone who hoped the law would disappear, the breathing space is deceptive.

## What obligations are already active for your startup?

Since February 2025, basic AI training is mandatory for any team using these tools. No master's degree is required: a brief internal policy and a training session are enough to comply with Article 4. The second live obligation is transparency. If your product has a chatbot, it must state that it is an AI on first contact; if it generates images, audio or text, it must mark them as synthetic. Detailed labelling arrives on 2 December 2026.

Prohibited uses are also in force. The rule bans harmful manipulation, exploitation of vulnerabilities and emotion inference in the workplace. A startup selling sentiment analysis to a human resources department should review that case before scaling it, because the boundary between legitimate use and prohibited practice depends not only on the model, but on what it is sold for.

Let us take a concrete example. A company offering automated support with a conversational assistant almost always falls into limited risk. It is enough to inform the user and mark the generated responses. If that same assistant starts evaluating job candidates or deciding whether to grant a loan, it moves up a tier and takes on the full package of high-risk obligations.

## Are you a provider or just a user of an AI?

The question decides almost everything else. If you use someone else's AI as sold to you, you are a user or deployer and your duties are light. If you put your name to an AI feature, or modify it substantially, you become a provider and inherit the full burden. That leap is in Article 25 and is a common trap for anyone reselling a third-party model under their own brand.

The difference shows in money and in the timetable. A provider of a high-risk system takes on a considerably heavier package:

*   Risk management and data governance.
*   Technical documentation and activity logs.
*   Human oversight and conformity assessment.

A user only has to monitor the system, keep records and apply the human oversight the provider has planned. Most European startups are users of AI tools, not manufacturers.

The nuance matters because many companies believe that building on an API from a general-purpose model makes them providers of the model. It does not. The obligations of the underlying model fall on whoever trains and publishes it; the startup is responsible for the system it builds on top, classified by its use. The confusion between the two layers generates more legal queries than the law itself.

## How much does compliance cost and what support is there for SMEs?

The figures vary so widely that they are best treated as ranges rather than budgets. One market analysis puts the first year for a five-person startup with a chatbot subject to transparency at between 15,000 and 30,000 euros, and a high-risk system at between 40,000 and 90,000 euros or more, according to [Wavect](https://wavect.io/blog/eu-ai-act-compliance-cost-startup/). Another estimate lowers the floor to 5,000 euros for those who only deploy third-party tools without high-risk cases, according to [Praxikon](https://www.praxikon.com/en/posts/what-does-eu-ai-act-compliance-cost-realistic-numbers).

Profile

Indicative first-year cost

User without high-risk cases

5,000 to 25,000 euros

Chatbot or assistant with transparency

15,000 to 30,000 euros

Provider with a high-risk system

40,000 to 90,000 euros or more

The law includes concrete relief for small companies. Each Member State must offer a controlled testing environment with priority and free access for SMEs and startups, reduce conformity assessment fees and accept simplified technical documentation, according to the summary by [Confir](https://confir.eu/ai-governance/for-startups). In addition, the fine cap for SMEs and startups is the lower of the percentage and the fixed amount, not the higher, as detailed in Article 99.

That detail of Article 99 is rarely cited and it changes the conversation with investors. A large company may feel 7% of global turnover more than 35 million; a startup feels the fixed figure more. The fact that the law applies the lower amount reduces the theoretical risk, although it does not exempt anyone from complying. Documentation remains mandatory if the use is high risk.

**[Europe's technological suicide with the AI Act: it debuts sanctions against an industry it does not have](https://observatorioblockchain.com/en/ai-agents/suicidio-europa-ai-act-multas-industria/)**

### Does the AI Act really hold back European startups?

There are signals to suit every taste. In August 2026, European tech investment fell 63% compared with July, to 3.2 billion euros across 165 deals, although artificial intelligence took a fifth of the total, according to [AI in Europe](https://aiineurope.co/startups/european-startup-funding-falls-august-ipronics-2026-09-05). In the first half of the year, the continent raised some 44.1 billion, more than a year earlier, but spread across the lowest number of deals in six years.

Founders' perception is not optimistic. A DutchBasecamp survey of 155 founders published in June 2026 found that 52% had avoided entering the European Union market, 58% had delayed their market entry and 24% were considering withdrawing permanently. A report by the Joint Research Centre puts the share of European venture-backed startups relocating abroad at between 3.3% and 4.3%, compared with 0.3% to 0.5% of those not venture-backed.

### A weak counterargument

The counterargument has its own evidence. An academic study of 1,365 European startups found no significant effect of the law on capital allocation between August 2022 and July 2026. Some investors see the European requirement as a quality seal for selling outside the bloc, and several international fund managers use it as an upfront filter.

Yet reality contrasts with these numbers. Mistral, one of Europe's largest AI laboratories, has decided that betting on AI training is too costly, and has decided to change its business model to become a provider of Chinese models, with guarantees and guardrails such as those the EU requires. Curious, that Mistral decides that "training its own AI" in the EU is very risky for the sustainability of its business, and instead decides to use models, trained without the guarantees the EU requires, something that is possible because of the "laxity" of the rules for service providers, compared with the laboratories that train their own foundation models.

**[The European AI law harms SMEs](https://observatorioblockchain.com/inteligencia-artificial/la-ley-europea-de-ia-perjudica-a-las-pymes/)**

### What you should have ready before December 2027

In any case, the extra time is not an amnesty, it is time to document. What is cheap now is the register of uses and team training; what will be expensive later is hastily reconstructing data governance that no one recorded. If your product touches employment, credit or education, start with that piece, because it is the one that takes longest to mature and the one that draws the most questions in due diligence.

It is also worth reading the rule as an exercise in traceability that rewards those who can explain their system. Startups that turn that work into a commercial advantage will have something their competitors outside the bloc cannot yet show, especially when other markets approve similar rules. The useful question is no longer whether compliance will be required, but who will have prepared the answer first.
