AIBusiness softwareContractsChoosing a vendor
The EU AI Act for companies building software
7 min read
The EU AI Act, Regulation (EU) 2024/1689, entered into force on 1 August 2024 and applies in stages. For a company building software with AI features, most of what matters already applies: prohibited practices since 2 February 2025 and transparency rules since 2 August 2026. The high-risk obligations, however, were pushed back to 2 December 2027, or 2 August 2028 for regulated products, by a simplification regulation published in July 2026.
This article reflects the position on 23 October 2026, in general terms. It is not legal advice: if your software touches a sensitive area (hiring, credit, health, education), have your analysis checked by a lawyer who specialises in EU technology law.
Who it applies to
The Act is not limited to European companies. Under Article 2, it covers providers placing AI systems on the EU market wherever they are established, deployers located in the EU, and providers and deployers in third countries when the output of the AI system is used in the EU. A UK, US or Canadian company selling software with AI features to EU customers should plan as if the Act applies.
The timeline, after the July 2026 delay
The original dates are set in Article 113 of the AI Act. On 29 June 2026 the Council adopted the "Digital Omnibus on AI", which became Regulation (EU) 2026/1744 of 8 July 2026. It was published in the Official Journal on 24 July 2026 and entered into force three days later, just before the general deadline of 2 August 2026.
| Date | What applies | Where |
|---|---|---|
| 1 August 2024 | The Act enters into force | Art. 113 |
| 2 February 2025 | Prohibited practices, AI literacy | Art. 5 and 4 |
| 2 August 2025 | General-purpose AI models, governance, penalties | Ch. V, VII, XII |
| 2 August 2026 | General application, including transparency | Art. 50 |
| 2 December 2026 | Two new prohibitions, end of the content-marking grace period | Art. 5 and 50(2) |
| 2 December 2027 | High risk: Annex III systems | Ch. III |
| 2 August 2028 | High risk: regulated products in Annex I | Ch. III |
Three details matter. The two new prohibitions target systems that generate intimate images of an identifiable person without consent, and child sexual abuse material. The grace period to 2 December 2026 for machine-readable marking of generated content only covers systems placed on the market before 2 August 2026. And the AI literacy duty in Article 4 was softened: providers and deployers must now "take measures to support" AI literacy among their staff rather than ensure it.
Provider or deployer: which one are you?
Obligations depend on your role, as defined in Article 3.
- A provider develops an AI system, or has it developed, and places it on the market or puts it into service under its own name or trademark. The provider carries most of the design obligations.
- A deployer uses an AI system under its authority in a professional context. Its duties are about use: follow the instructions, ensure human oversight, inform the people affected.
The point that surprises buyers: a company that has a software studio build an AI application, then releases it under its own name, can be the provider, even without writing a line of code. Article 25 also turns a deployer into a provider if it puts its name on a high-risk system or substantially modifies it. The company behind the underlying language model has separate obligations as a provider of a general-purpose AI model.
When the vendor remains the publisher and operator of the application, as in a subscription model, the split changes. Either way, it should be written into the contract: who is the provider, who supplies documentation, who keeps the logs.
The four risk levels
| Level | Examples | What to do |
|---|---|---|
| Prohibited | Emotion recognition at work, social scoring | Do not build it |
| High risk | CV screening, credit scoring, school admissions | Full compliance |
| Transparency | Chat assistants, generated content | Disclose and mark |
| Minimal risk | Invoice extraction, internal search, spam filters | Good practice |
Prohibited. Article 5 bans, among other things, manipulative subliminal techniques, exploiting vulnerabilities, social scoring, and inferring people's emotions in the workplace or in education, except for medical or safety reasons. A feature that "analyses employee mood" is not a grey area: it has been banned since February 2025.
High risk. Annex III lists the areas concerned: biometrics, critical infrastructure, education, employment and worker management (CV screening, performance evaluation, allocating tasks based on behaviour), access to essential services (creditworthiness, life and health insurance pricing), law enforcement, migration and justice. An Annex III system can fall outside the high-risk regime if it does not materially influence decisions, for example when it performs a narrow procedural task (Article 6(3)), but the provider must document that assessment. An Annex III system that profiles people is always high risk.
Transparency. Article 50 requires telling people they are interacting with an AI system unless it is obvious, marking generated content in a machine-readable format, and disclosing deepfakes.
Minimal risk. Most features in a business application, such as those described in our guide to adding AI features to a business application, carry no specific obligation beyond AI literacy.
What most companies should do now
- List every AI feature in your software and assign each one a risk level.
- Check that none is prohibited, especially anything that monitors staff.
- Apply transparency: a clear notice when users talk to an assistant, and marking of generated content when you are the provider.
- Train the people who use or configure these features.
- Document your reasoning when a feature touches an Annex III area without being high risk.
The AI Act does not replace the GDPR: when personal data flows through an AI feature, both apply. Our GDPR checklist for a business application covers the data protection side.
If your use case is high risk
The provider must run a risk management system, govern training data, keep technical documentation, log events automatically, design for effective human oversight, and complete a conformity assessment before release, with registration in the EU database. Under Article 26, the deployer must use the system according to its instructions, assign human oversight to competent and trained people, make sure the input data it controls is relevant, and monitor how the system performs.
The delay to December 2027 buys time, not an exemption. A CV screening tool designed today should be built with these requirements in mind, or it will need rework before the deadline.
Penalties
Article 99 sets maximum administrative fines:
| Breach | Cap | Or, if higher |
|---|---|---|
| Prohibited practice | 35 million euros | 7% of worldwide turnover |
| Other obligations (provider, deployer, transparency) | 15 million euros | 3% |
| Incorrect information to authorities | 7.5 million euros | 1% |
For small and medium-sized enterprises, including start-ups, the lower of the two amounts applies instead. Regulation 2026/1744 extends that rule to "small mid-cap" enterprises for the last two rows.
What to put in your contract with a software vendor
- each party's role under the AI Act, feature by feature;
- the list of AI models used and who provides them;
- the documentation and logs handed over to you;
- what happens if a feature moves into a different risk level.
These clauses come on top of a standard development contract, covered in the clauses to check in a software development contract. Commission guidelines and technical standards are still being published: check their status on the European Commission's AI Act page before locking in a decision.
Where to start
List the AI features you have in mind and ask the risk question for each one before writing the specification. At LPR Studio, the visual prototype is free and presented in a video call within 5 working days, wherever you are: it shows where the AI steps in, where the user confirms and what notice they see. That is where our approach always begins.