EU AI Act for Startups: Reliefs, Sandboxes, and Where the Rules Still Bite
EU AI Act startup provisions: sandboxes (Arts 57–59), reduced fees, Art 99(6) fine cap, simplified docs. What still applies. Deadline 2 Dec 2027.
The EU AI Act gives startups genuine structural advantages — fee reductions, fine caps, sandbox access, and simplified documentation. None of them cancel the substantive requirements. Understanding what the reliefs actually cover, and what they don't, is where startup compliance begins.
This article focuses on the startup-specific provisions in Regulation (EU) 2024/1689: the SME support instruments built into the Act itself. If you're looking for a step-by-step how-to on registering your system and drafting the Annex IV technical file, that workflow is covered in the EU AI Act SaaS startup compliance guide — start there if you're a software founder building an AI product.
The Act's SME Support Architecture: Four Real Instruments
The EU AI Act is unusual among EU regulations in that it explicitly encodes startup-friendly provisions into the text rather than leaving proportionality entirely to member-state discretion. Four instruments are worth understanding precisely.
1. AI Regulatory Sandboxes (Articles 57–59): Supervised Testing Before You Commit
Article 57 requires every EU member state to establish at least one national AI regulatory sandbox by 2 August 2026. Some member states — Spain, the Netherlands, Norway — already operate sandboxes; the deadline forces every remaining member state to follow.
What a sandbox actually gives you: the ability to develop and test an AI system under regulatory supervision, outside the full compliance framework that would otherwise apply to a live market deployment. The competent authority monitors the testing; you operate under a structured plan agreed in advance.
Article 59 adds important operational specifics. Where the sandbox involves processing personal data, a competent authority from the data-protection supervisory authority must be involved. Participants exit the sandbox with a written report — a document you will want to preserve, because it records the supervisory authority's guidance and your system's testing history. That is genuine regulatory capital for a seed-stage company that cannot yet afford a full Annex IV technical file.
Priority access for startups and SMEs. Article 58(1) explicitly requires that SMEs and startups be given priority access to national sandboxes, and that participation fees be proportionate to their size — in practice often free for small companies, though the exact form varies by member state. A 12-person team applying alongside a large corporate is not competing on equal footing — the Act says you go first.
Limited real-world testing (Article 60). Where a sandbox plan includes limited real-world testing — deploying a system to actual users under supervised conditions — Article 60 sets the guardrails: the testing must be for a fixed duration, involve only users who have given informed consent, and include monitoring and stop mechanisms. For a startup wanting to test, say, a hiring-assistance tool with a handful of partner companies before formal market placement, this is the cleaner route than an unsupervised soft launch.
One practical point: applying to a sandbox takes time. If your target national authority has not yet established one (the August 2026 deadline is close), check the EUAIA sandbox network and Article 58 cooperation mechanisms — cross-border sandbox access is permitted where no national option exists.
2. Proportionate Conformity-Assessment Fees (Article 62)
Article 62(4) requires that the fees charged by notified bodies for conformity assessments of high-risk systems be proportionate to the size of the company, with an explicit reduction for SMEs. The Commission is also empowered to set standard fee scales to prevent the notified-body market from ignoring this.
In practice, this matters most for Annex III point 1 (biometrics), which requires the Annex VII notified-body route rather than internal self-assessment. For all other Annex III categories — employment, education, creditworthiness, and the rest — the default is Annex VI internal self-assessment, which costs only internal time. The fee-reduction provision applies to the minority of startups who need a notified body; the majority can self-certify.
3. The SME/Startup Fine Cap (Article 99(6))
This is the most financially concrete relief. Article 99 establishes three penalty tiers — €35 million or 7% of worldwide annual turnover (Article 5 prohibitions); €15 million or 3% (most other obligations, including high-risk requirements); €7.5 million or 1% (incorrect information to authorities) — each expressed as "whichever is higher."
Article 99(6) reverses that formula for SMEs and startups. For companies that qualify, the fine is the lower of the percentage or the fixed sum, not the higher.
Worked example: a startup with €2 million in worldwide annual revenue breaches a high-risk requirement — say, it places a system on the market without completing conformity assessment under Article 43. The standard fine ceiling is €15 million or 3% of €2 million (= €60,000), whichever is higher — so €15 million. Under Article 99(6), the ceiling flips to the lower: €60,000. That is not immunity from a fine, but it is a structurally different exposure.
Three points to keep in mind. First, Article 99(6) applies to the ceiling, not to whether a fine is imposed. Regulators retain discretion. Second, the protection applies while you remain an SME — scale past the threshold and the standard formula reinstates. Third, the Article 5 prohibition tier (€35M/7%) is still the ceiling even for SMEs; the reversal formula applies there too, but if your turnover is, say, €10 million, 7% = €700,000 becomes the lower figure.
4. Simplified Technical Documentation for SMEs (Article 11(3))
High-risk system providers must maintain technical documentation under Article 11, with content specified in Annex IV. Annex IV lists nine areas: system description; intended purpose and versions; elements and development process; monitoring, maintenance, and functioning; risk management; data and data governance; accuracy, robustness, cybersecurity metrics; human oversight; and the conformity assessment.
Article 11(3) provides that the Commission may adopt implementing acts specifying a simplified Annex IV form for SMEs. As of mid-2026, the delegated act has not been published, but the provision is binding on the Commission — it must be prepared. Watch for it; it is expected to reduce the documentation burden substantially for companies below the SME threshold, without removing the substantive requirement to have a technical file.
Until the simplified form is available, the full Annex IV remains the standard. The practical implication: structure your technical file now in a modular format so that if a simplified form is published before your deadline, you can reorganise rather than start over.
AI Literacy: The Obligation That Is Already Live (Article 4)
Article 4 — AI literacy — has applied to all organisations since 2 February 2025. It is not a high-risk obligation or a future deadline. Every company that uses or deploys AI must take measures to ensure its staff have sufficient knowledge to use AI systems competently and responsibly.
For a 10-person startup, "sufficient measures" is proportionate to size and risk. You are not expected to run a certification programme. What you are expected to do is ensure the people operating or supervising AI tools understand what those tools do, where they can fail, and when to escalate. For most startups this means: documented awareness of the systems in use, a brief record of who has been briefed and on what, and some structure around oversight for any system interacting with customers or making consequential decisions.
The Act does not mandate a certification standard. What it does do is create an audit footprint — if an incident occurs, a regulator will ask whether staff knew what they were running. A 15-minute onboarding note is better than nothing; a brief internal policy document is better than that.
What the Reliefs Do Not Do: Classification Still Applies in Full
All four instruments above are procedural and financial adjustments. They do not alter how your system is classified or what substantive obligations attach to that classification. The four risk tiers apply to startups exactly as they apply to large companies.
Most startups are deployers or limited-risk providers. If your company uses a third-party AI tool — a customer-service chatbot, an analytics layer, a recruitment screening tool from a vendor — you are a deployer under Article 26. Deployer obligations are lighter than provider obligations: follow the provider's instructions, maintain human oversight, log system use, monitor performance, and flag serious incidents to the provider. The documentation burden is real but manageable for a small team.
If you build an AI product but it does not land in Annex III, you are likely a limited-risk provider under Article 50. If your product involves a customer-facing conversational interface, synthetic content generation, or emotion/biometric processing, you owe transparency disclosures (Article 50) from 2 August 2026. No technical file, no conformity assessment — just the disclosure obligations.
High-risk only applies if Annex III says so. The eight Annex III categories are: biometrics, critical infrastructure safety components, education and vocational training, employment and worker management, access to essential private and public services (including creditworthiness scoring and health/life insurance risk assessment), law enforcement, migration and border control, and administration of justice. If your system is not in one of those categories, it is not high-risk. A scheduling tool is not high-risk. A customer churn model is not high-risk. A code suggestion tool is not high-risk. Confirm the classification by working through Article 6 and Annex III systematically.
Article 6(3) filter. Even if your system falls inside an Annex III category, it is not high-risk if it does not pose a significant risk of harm to health, safety, or fundamental rights — for instance, if it performs a narrow procedural function, assists a human activity that has already been completed, or detects patterns without replacing or influencing human judgment. Any system that profiles natural persons is always high-risk regardless of the filter. If you claim the Article 6(3) exemption, document the assessment and register the system in the EU database under Article 49.
The deadline for stand-alone high-risk systems is 2 December 2027. Under the Digital Omnibus agreed in May 2026, the original August 2026 date has been deferred. High-risk AI embedded in regulated products under Annex I (medical devices, machinery) faces a further deadline of 2 August 2028. That is real breathing room — but assembling the Annex IV technical file, running conformity assessment under Article 43, and preparing the Article 47 Declaration of Conformity takes months. Starting in late 2027 is not a viable plan.
What a Startup Does First
For most early-stage companies, the priority order is:
Step 1: Build an AI inventory. List every AI system you build or use, internally and in your product. Assign a role to each — provider or deployer — using Article 3's definitions and the Article 25 role-shift rules if you have substantially modified or rebranded a third-party system.
Step 2: Classify each system. Run each through the Article 5 prohibited list (is this system doing something the Act bans outright?), then the Article 6 / Annex III high-risk test, then the Article 50 limited-risk transparency check. Most systems will land at minimal or limited risk. Document the reasoning.
Step 3: Check Article 4 literacy. Confirm that the staff using AI tools have been briefed. Record it briefly.
Step 4: Apply for a sandbox if you are developing a high-risk system. Do not wait for the deadline to force the issue. Sandbox participation generates regulatory guidance that is difficult to obtain any other way, and priority access for startups means the queue works in your favour.
Step 5: For any high-risk system, begin the Article 11 / Annex IV technical file now. Modular documentation — system description, intended purpose, data governance records, risk management log — is far easier to maintain incrementally than to reconstruct later.
How Confir Helps
Confir is a self-serve EU AI Act compliance tool for companies that do not have dedicated legal teams or compliance departments. Its classification engine is rule-based and deterministic — same intake answers produce the same finding, and the rule that fired is human-readable.
For startups: Confir guides you through the Article 6 / Annex III classification, assigns your role (provider, deployer, or both under Article 25), and generates the Article 11 / Annex IV technical documentation pack and Article 47 Declaration of Conformity. For deployers, it runs the structured Article 26 assessment. Pricing starts at €600/year for companies with two to three AI systems. No consultants, no procurement cycle, no six-month implementation.
If your system is minimal or limited-risk, the classification output itself — documented in an audit trail — is the primary artefact you need. If you are heading toward a sandbox application, the technical file Confir generates gives you the documentation basis for the application.
Frequently Asked Questions
Does the EU AI Act apply to my startup if we are not based in the EU?
Yes. Article 2 extends the Act to any AI system placed on the EU market or put into service in the EU, regardless of where the provider or deployer is established. If you serve EU users or customers, you are in scope. The extraterritorial reach mirrors GDPR's market-access logic.
What is the Article 99(6) fine cap and does my startup qualify?
Article 99(6) means that for SMEs and start-ups, fines are capped at the lower of the percentage or the fixed sum — the reverse of the standard formula. For example, a startup with €3 million turnover facing a high-risk breach has a ceiling of €90,000 (3% of €3M) rather than €15 million (the fixed sum). The SME definition follows the EU standard: fewer than 250 employees and either turnover below €50 million or balance-sheet total below €43 million (per Commission Recommendation 2003/361/EC). Once you exceed those thresholds, the standard formula applies.
When must each member state have a national AI regulatory sandbox in place?
Article 57(1) sets the deadline at 2 August 2026. Several member states (Spain, the Netherlands, Norway) already operate sandboxes. SMEs and startups have priority access under Article 58(1), and fees must be proportionate to size — in practice often free for small companies. Cross-border access is permitted where a national sandbox does not yet exist.
My startup uses a third-party AI tool for HR screening. Are we the provider or the deployer?
You are most likely a deployer under Article 26 — you are using a system under your authority in a professional context, not placing it on the market under your name. Your obligations: use it per the provider's instructions, implement human oversight, maintain logs for at least six months, and report serious incidents to the provider. If you have substantially modified the tool or integrated it under your own brand, Article 25 can shift you to provider status — run that assessment carefully.
Does the Article 11(3) simplified documentation form exist yet?
Not as of mid-2026. Article 11(3) requires the Commission to adopt implementing acts specifying a simplified Annex IV form for SMEs, but those acts have not yet been published. Until they are, the full Annex IV structure applies. Draft your technical file in modular sections so you can adapt quickly when the simplified form arrives.
What does "AI literacy" actually require from a small team?
Article 4, in force since 2 February 2025, requires all organisations to take proportionate measures to ensure staff using AI systems have sufficient knowledge to operate them responsibly. No certification is mandated. For a small startup, documented briefings — who was told what, and when — plus an internal note on the AI tools in use and their known limitations will satisfy the proportionality test for most systems. For a high-risk deployment, the standard is higher: the staff overseeing the system should understand its intended purpose, its failure modes, and when to override it.
Related guides
- SMB compliance requirements
- Articles 6-11 risk levels
- risk management tools comparison
- Articles 6–29 compliance checklist
Manage your EU AI Act compliance in one place
Confir automates risk classification, technical documentation, and audit trails for any company. No consultants. No 6-month projects. 7-day free trial.
Start free trial →