Guide · 13 min read
EU AI Act for European Businesses: the 2026 operational guide
The 4 risk levels of the AI Act, what changes for HR, finance, marketing, support, and the 4 steps every European SME should take. Updated November 2026.
The EU AI Act came into force on 1 August 2024. Since then, European SMEs have been receiving emails from AI consultants raising the spectre of fines up to 7% of turnover. The operational reality is far calmer: for most SMEs, direct obligations are few and manageable.
This is the honest map: what actually changes for your business, the 4 concrete steps you should take before 2026, and how specific sectors (HR, finance, marketing, support) are affected.
The 4 AI Act risk categories
The regulation classifies AI systems into 4 risk levels. The level determines the obligations. Understanding where each of your systems sits is the starting point.
Unacceptable risk (prohibited)
Systems banned throughout the EU from February 2025 (Article 5):
- Social scoring in the style of China (evaluating people based on social behaviour).
- Subliminal manipulation causing harm.
- Exploitation of vulnerabilities of specific groups (minors, people with disabilities).
- Real-time remote biometric identification in public spaces for law enforcement, with narrow exceptions.
- Biometric categorisation based on race/religion/orientation.
- Emotion recognition in the workplace and in education, except for medical or safety reasons.
Does this affect your SME? Probably not, with one exception worth remembering: an HR tool that infers employees’ emotions is prohibited.
High risk (heavy obligations)
Systems that can have a significant impact on people’s rights, safety, and opportunities:
- Recruitment and HR (CV screening, candidate assessment, promotion decisions).
- Consumer credit scoring (banks, fintech).
- Diagnostic healthcare systems.
- Critical infrastructure (energy, transport, traffic).
- Education (access, student assessment).
- Law enforcement (with specific exceptions).
- Migration/asylum management.
- Emotion recognition where it isn’t prohibited, with the duty to inform the people exposed.
Does this affect your SME? Yes, if you operate in recruitment, HR, finance, or healthcare. For other sectors: only if you build a system that falls specifically into this category.
Obligations for these systems apply from 2 December 2027, or from 2 August 2028 for those built into regulated products such as medical devices, after the postponement agreed in the Digital Omnibus.
Limited risk (transparency)
Systems that interact with people or produce content that must be disclosed:
- Customer-facing chatbots.
- Deepfakes and texts generated to inform the public on matters of public interest.
Main obligation: transparency (Article 50), from 2 August 2026. Users must know they are interacting with an AI, and AI-generated deepfakes and public-interest texts must be disclosed (texts are exempt if a person reviews them and takes editorial responsibility). Emails to customers written with AI and reviewed by a person carry no labelling obligation.
Does this affect your SME? Yes, if you have chatbots or agents that talk to customers, or if you publish deepfakes. For marketing images, see how to label AI-generated content.
Minimal risk (residual obligations)
Everything else: spam filters, recommendation systems, internal process optimisation, use of ChatGPT/Copilot for office productivity.
Obligations: none specific to the system. As for everyone, training the people who use it still applies (Article 4). No heavy compliance required.
Does this affect your SME? Yes, for the majority of everyday use cases. Good news: zero bureaucracy required.
Summary table
| Level | Examples | Obligations |
|---|---|---|
| Unacceptable | Social scoring, manipulation | Prohibited |
| High | Recruitment, credit, healthcare | For users: provider’s instructions, human oversight, logs for at least 6 months, informing people, DPIA (GDPR) |
| Limited | Chatbots, deepfakes | Transparency |
| Minimal | ChatGPT in the office, spam filters | None specific (staff training still applies, Article 4) |
What changes by sector
Recruitment & HR (high risk)
If you use AI to select, filter or evaluate candidates, or to decide on promotions and performance reviews, from 2 December 2027 you must (Article 26):
- Use the system according to the provider’s instructions; the provider must have assessed and documented it.
- Assign human oversight to competent people.
- Keep the logs generated by the system for at least six months.
- Inform workers and their representatives before use, and candidates that they are subject to the system.
- Carry out the DPIA required by the GDPR, using the provider’s information.
In practice in Soraia recruitment projects:
- The AI agent filters and ranks candidates, but the human recruiter always signs off on the final decision.
- We maintain an immutable audit log for every decision, reconstructable over time.
- DPIA document generated as a sprint deliverable.
- The agent does not produce legally binding outputs without human review.
Even with the recruiter signing off, filtering and ranking candidates stays high risk: it is profiling, which the AI Act always treats as high risk (Article 6). That’s why we operate at the high-risk level.
Finance & Accounting (minimal risk if done right)
Agents that read invoices, perform OCR, validate VAT, and generate reports → minimal risk. Internal automation, no autonomous decisions impacting individuals.
Main obligation: internal transparency (the finance team must know the agent is AI-driven).
Exception: if the agent assesses the creditworthiness of natural persons → high risk (Annex III). If it automatically affects customer invoicing, it needs a case-by-case assessment.
Sales & Marketing (minimal or limited risk)
Agents performing personalised outreach, lead scoring, email drafting → usually minimal risk; it becomes limited risk if the agent converses with customers on its own.
If the agent writes and replies to customers on its own (e.g. automated follow-ups that carry the conversation forward):
- You must make clear that there’s an AI on the other side, at the latest at the first interaction (Article 50).
- It’s worth letting the recipient ask for a person.
If AI writes the email but a person reviews and sends it, there is no labelling obligation.
If the output stays internal (e.g. lead score for the sales team, not communicated to the customer):
- Minimal risk. Internal transparency only.
Customer Support (depends)
Frontline customer-facing chatbot → limited risk. Obligation: inform the user they are talking to AI, clearly and at the latest at the first interaction (Article 50).
Internal routing + response drafting (human approves and sends) → minimal risk. The human acts as the final check; the system is “assisted”.
Education & Training
If your SME provides external training that uses AI for student assessment → high risk. If AI is used only to generate learning materials → minimal risk. If AI is purely a productivity tool for internal trainers → minimal risk.
The 4 concrete steps for SMEs
Regardless of risk level, here are the 4 practical things to do before 2026. This is the minimum. Full stop.
Step 1, Internal AI policy (1 page)
A 1-page document stating:
- What is permitted when using AI in the business (using ChatGPT for standard tasks, etc.).
- What is prohibited (inputting sensitive personal data without a DPA, using personal accounts for company tasks, producing legally binding AI output without review).
- How to report issues (dedicated internal channel).
- Who is responsible for AI governance (designated manager).
Supports internal governance, is an implicit best-practice requirement, and satisfies the accountability principle.
Step 2, Audit log for agent decisions
Every agent in production must log every decision:
- Timestamp.
- Input received (anonymised where necessary).
- Rules/prompt applied.
- Output produced.
- Human escalation flag (yes/no/when).
For high-risk systems this is a regulatory requirement (Articles 12 and 26). For others it is smart regardless: debugging, accountability, continuous improvement.
Standard 2026 tooling: Langfuse, OpenLLMetry, custom systems on S3 + metadata. Cost: ~€100-300/month for an average SME.
Step 3, Human oversight on critical cases
Define upfront: when the agent decides autonomously, and when it escalates to a human.
Examples:
- Recruitment: agent filters and ranks; human always signs the final offer.
- Accounting: agent processes invoices below €X; human reviews anomalies above €10k.
- Support: agent handles standard requests; human manages negative or complex escalations.
Documenting the “automatic cases / escalation cases” matrix is the most direct and most powerful compliance deliverable.
Step 4, DPA with the vendor
If you use an external provider (Soraia, OpenAI direct, Anthropic, Microsoft, Google), you need a Data Processing Agreement (Article 28 GDPR).
- OpenAI has a standard DPA available for download.
- Anthropic has a standard DPA.
- Microsoft (Copilot) has a DPA included in the Business tenant.
- Google has a standard DPA.
- Soraia (and any serious agency) includes it in the sprint contract.
Without a DPA, handing personal data to a vendor breaches the GDPR. The most direct step, and the one most often skipped in a rush.
Timeline 2025-2028: what expires when
Key dates to plan your compliance:
- February 2025 ✓ already in force, prohibition of unacceptable AI practices and staff AI literacy (Article 4).
- August 2025 ✓ already in force, obligations for general-purpose AI models (mainly affects vendors like OpenAI/Anthropic, not end users).
- August 2026 ✓ already in force, most of the regulation applies, including transparency: chatbots and agents that talk to customers must disclose they are AI.
- 2 December 2026, new prohibitions added by the Digital Omnibus (non-consensual intimate content and child sexual abuse material generated with AI); the grace period for providers to mark content generated by systems already on the market also ends.
- 2 December 2027, obligations for Annex III high-risk systems. If you have recruitment AI or credit scoring, this is the operational deadline for compliance.
- 2 August 2028, obligations for high-risk systems built into regulated products, such as medical devices.
For most SMEs the date that mattered was August 2026, for transparency, and it has passed. Whoever uses high-risk systems has until 2 December 2027.
Fines: how they actually work
Scaremongering emails throw around figures like “up to 7% of turnover”. True, but misleading. The maximum fine is for prohibited practices; the lowest is for false information to authorities.
Operational summary:
| Type of violation | Maximum fine |
|---|---|
| Prohibited AI practices | €35M or 7% of turnover |
| Other obligations (high risk, transparency) | €15M or 3% of turnover |
| False or misleading information | €7.5M or 1% of turnover |
For large companies the higher of the two amounts applies; for SMEs, including start-ups, the lower (Article 99).
For an average European SME (turnover €5-20M), the real risk of a fine is low if you:
- Do not use prohibited systems.
- Have the 4 basic steps above implemented.
- Cooperate with authorities if approached.
Minor sanctions (warnings, requests for remediation) are far more likely and manageable.
What NOT to do (scaremongering red flags)
When an AI consultant tells you these things, it is a red flag:
“You are legally required to appoint an AI Officer.” False for SMEs. The AI Act does not mandate the appointment of a dedicated role.
“You need €50,000 worth of compliance work to be covered under the AI Act.” For most SMEs, the 4 basic steps cost €2-5k in one-off consulting, not €50k.
“The AI Act bans the use of ChatGPT in business.” False. Business use of ChatGPT is minimal risk; an internal policy and training for the people who use it are enough (Article 4).
“Automatic fines from 2025 with no warning.” False. Fines aren’t automatic: authorities decide case by case, taking into account gravity, duration, company size and cooperation (Article 99).
“Only certified vendors are acceptable.” False. Only high-risk systems need a conformity assessment and CE marking, which for Annex III systems, except biometrics, the provider carries out itself without external bodies (Article 43). For everything else there is no mandatory certification.
The Soraia advantage for compliance
Full disclosure: this guide is written by Soraia, an AI agency. Our sprints include by default:
- Immutable audit log for every agent decision.
- DPIA template already drafted for recruitment / finance / people-related processes.
- AI policy 1-pager customised for the client’s team.
- EU hosting for personal data (no unauthorised extra-EU data transfers).
- Art. 28 GDPR DPA pre-included in the contract.
Not because we are heroic, but because without these things we cannot do the work in regulated sectors at all.
Want an assessment of your specific situation under the AI Act? Do the 3-minute check-up to understand where you stand, or talk to Daniel for 20 minutes to walk through your current AI systems and real obligations together.
Frequently asked questions
What people usually ask us.
Does the AI Act apply to European SMEs?
When did the AI Act come into force?
What are the 4 risk levels of the AI Act?
If I use ChatGPT at work, do I need to do anything specific for the AI Act?
Recruitment with AI: what obligations apply?
What is a DPIA and when is it required?
What are the AI Act fines?
Do I need to appoint an AI Officer?
Keep reading
Guide · 14 min
AI Consulting in Italy: How to Choose the Right Partner
Operational map of AI consulting in Italy 2026: the 3 types of providers, selection criteria, red flags, and how to evaluate a quote. A guide for SME CEOs.
Read the guideGuide · 12 min
AI Agents for Business: what they are and how to choose
What an AI agent in business is, when it makes sense, how it's built and what it costs, with 4 real agents in production at European SMEs. Updated 2026.
Read the guideGuide · 12 min
AI consulting costs: price ranges, models and drivers
What an AI agent actually costs, the 4 commercial models (fixed price, retainer, % savings, mixed), what drives the price, and how to measure ROI.
Read the guideWant an opinion on your case?
20 minutes with the CEO to figure out together whether it makes sense. No commitment, no pitch: just a practical conversation about your processes.