← Back to blog

35 Obligations in the EU AI Act That Don't Tell You How to Comply

By Provectio | April 2026 | ~8 minute read

The EU AI Act is 113 articles of regulatory text. We measured every article using a traceable process that compares obligation strength, operational specificity, and relationships across the Act. Then we looked for a specific pattern.

Here's what we found.

The pattern: binding without direction

Across the Act, 205 articles carry high binding force — a demand score of 0.90 or above on a 0-1 scale. These are the provisions backed by the Act's enforcement apparatus. Fail to comply and you face administrative fines up to EUR 35 million or 7% of your worldwide annual turnover, whichever is higher.

Most of those 205 articles also carry high operational force. They tell you what to do. "Shall establish a risk management system." "Shall ensure that training data meets quality criteria." "Shall implement human oversight measures." The obligation is clear. The expected action is clear. You know what compliance looks like.

But 35 of them don't. They carry full binding force — demand scores of 0.85 to 0.95 — but their operational force is at or below 0.50. The law tells you that you must. It doesn't tell you what you must do.

This is the instruction gap.

What an instruction gap looks like in practice

Consider an article that says, in effect: "Providers shall ensure that high-risk AI systems are designed to achieve appropriate accuracy, robustness and cybersecurity throughout their lifecycle."

The binding force is unambiguous. "Shall ensure" is one of the Act's strongest imperatives. There's no "may" or "should" or "where appropriate." The obligation is absolute.

The action force is minimal. "Appropriate accuracy" is not a specification. Neither is "robustness." Neither is "cybersecurity throughout the lifecycle." These are objectives. They tell you where you need to arrive but not how to get there.

Now read that provision as a compliance officer. You must ensure appropriate accuracy. What is appropriate? The Act doesn't define it for your specific system. What robustness threshold must you meet? The Act doesn't specify. What cybersecurity controls satisfy the requirement? The Act references harmonised standards — but many of those standards haven't been published yet.

You are strictly bound. You have no clear roadmap.

Why this matters

Three consequences flow from instruction gaps.

Over-compliance. When the demand is mandatory but the action is undefined, rational compliance teams build extensive documentation, run redundant testing, and adopt every control they can think of. The cost of over-compliance is wasted engineering effort and delayed product launches. The cost of under-compliance is EUR 35 million. The asymmetry pushes every team toward the expensive side.

Implementation inconsistency. Two companies with identical AI systems will interpret the same instruction gap differently. One builds an elaborate accuracy validation framework. The other relies on industry best practice with minimal documentation. Both can claim compliance. A regulator auditing after an incident will pick which interpretation they prefer. Neither company knows in advance whether their approach will hold.

Legal drift. The first CJEU ruling on any instruction-gap provision will set the interpretation for every company in the market. Until that ruling happens, every compliance decision is a guess. Companies that guess wrong discover their exposure only when a regulator or plaintiff tells them.

The specific gap pattern

Across the 35 identified instruction gaps, the shape is consistent:

  • A strong performative verb ("shall ensure", "shall implement", "shall establish") creating binding force
  • A general objective ("accuracy", "appropriateness", "robustness", "transparency") without measurable criteria
  • A cross-reference to harmonised standards or delegated acts that haven't been adopted yet
  • A penalty article (typically Article 99(3) or 99(4)) linking non-compliance to significant fines

The cross-reference is the key. The Act doesn't specify the action directly because it expects the Commission to fill in the details through delegated and implementing acts, or through the harmonised standards process under Article 40. For most of the identified gaps, those secondary instruments are still pending.

As of April 2026, none of the major harmonised standards for high-risk AI systems have been published. The delegated acts on conformity assessment procedures are in draft. The implementing acts on technical documentation format are not yet adopted. Companies are told what they must achieve by August 2, 2026. They are not told what achieving it looks like.

Asymmetric enforcement

There's a related pattern worth naming. The Act's addressee structure is lopsided.

Across our analysis of 631 articles and paragraphs, the distribution of obligations by addressee breaks down as follows:

  • Providers: 222 article references
  • Commission: 197 references
  • National competent authorities: 169 references
  • Member States: 144 references
  • Notified bodies: 92 references
  • Deployers: 85 references
  • AI Office: 72 references
  • Authorised representatives: 35 references
  • Importers: 23 references
  • Distributors: 19 references

Providers bear the largest share of direct obligations — 222 references, roughly 2.6 times the number that apply to deployers. This is consistent with the Act's design intent: providers build the systems, deployers use them, and most of the compliance burden flows upstream to the builders.

But the provider burden compounds the instruction gap problem. If you build an AI system and put it on the market, you are responsible for satisfying every applicable obligation across its lifecycle. The instruction gaps are your problem. The implementation inconsistency risk is yours. The over-compliance cost is yours.

Deployers have an easier position. Fewer obligations, most of them procedural. They rely on the provider's declaration of conformity and the provider's technical documentation. If something goes wrong, the provider is the first line of regulatory fire.

What this analysis reveals

Conventional commentary on the AI Act describes its structure, explains its categories, and summarises the obligations. Our analysis does something different. It measures.

Every article has a force profile — a four-dimensional vector quantifying how much action it demands, how binding its requirements are, how much it asserts as fact, and how much scope context it sets. Every article has a dominant speech act class. Every article has cross-references to other articles and links to penalty provisions. Every article has an addressee.

Read one article at a time, you see a rule. Read them all together, you see a system. The instruction gaps aren't visible in individual reading. They only appear when you measure binding force against action force across the whole Act.

The 35 instruction gaps we identified aren't drafting errors. They're deliberate choices — the Act leaves room for the Commission's delegated acts and the standards process to fill in specifics later. But "later" was two years ago when the Act was adopted. The specifics are still not complete. The enforcement date is not waiting.

What to do about it

Three moves make sense if your AI system touches the EU market.

First, identify which of your obligations fall into the instruction gap pattern. You need to know where the law binds you absolutely and specifies nothing concrete. These are the provisions where legal interpretation will diverge and where the cost of getting it wrong is highest.

Second, document your interpretation. For each instruction gap, write down how you decided to comply, what standard or practice you relied on, and what evidence you have that your interpretation is reasonable. This document is not compliance. It's audit defence.

Third, build a monitoring process for secondary instruments. As delegated acts, implementing acts, and harmonised standards are published, your interpretation may need to shift. Companies that track the secondary instruments will adapt incrementally. Companies that don't will discover the gap between their interpretation and the final form of the law all at once.

A concrete offer

If you want to know where your AI system stands against every applicable obligation of the EU AI Act — including every instruction gap, every cross-reference, every penalty link, and every addressee role — we built a tool that answers those questions in two weeks.

The analysis is backed by forensic receipts with cryptographic integrity hashes. Every finding can be independently verified. The methodology is governed, not subjective. The output is not legal advice. It is structural input for the legal and compliance decisions you still need to make.

For a classification assessment of a single system: 48 hours, EUR 2,500.

For verification of your Annex IV technical documentation: one week, EUR 15,000.

For a comprehensive compliance audit covering every applicable article: two weeks, EUR 35,000.

The deadline is August 2, 2026. That is 115 days from the date of this post. If you haven't started, you don't have much more time.

Provectio is a regulatory evidence company. We measure legal text, quantify how strongly and specifically it binds, and produce cryptographically receipted analyses for compliance, research, and regulatory work. Our analysis of the EU AI Act covers 631 articles and paragraphs with 2,203 detected tensions, 1,058 addressee mappings, and 799 cross-references. This post draws on that dataset.

The analysis in this document is not legal advice. Consult qualified legal counsel for specific compliance decisions.