Skip to main content

01 - The EU AI Act for Developers: Obligations and Dates

Which AI Act obligations apply to your code and from when after Regulation 2026/1744: roles, Annex III, credit scoring and technical requirements.

What do you think about this article?

01 - The EU AI Act for Developers: Obligations and Dates
AI Policy EUPhoto:Tara WinsteadPexels

At a Glance

Which AI Act obligations apply to your code and from when after Regulation 2026/1744: roles, Annex III, credit scoring and technical requirements.

Which AI Act obligations apply to your code, and from when

If you build a system that uses a machine-learning model and someone in the European Union uses its output, Regulation (EU) 2024/1689, the AI Act, concerns you too. The useful question is not "am I compliant?" but "which part of the regulation applies to me, and from which day?", because the dates are no longer the ones that circulate in many articles: Regulation (EU) 2026/1744, in force since 27 July 2026, moved the application of the obligations for high-risk systems to 2 December 2027 (Annex III) and 2 August 2028 (Annex I).

This first article of the series gives you the map: the roles, the regimes a system can fall under, a concrete case (a credit-scoring system) and the consequences for whoever writes the code. The following articles go into detail: civil liability, general-purpose AI models, content marking, the overlap with the GDPR and an audit kit. Nothing that follows is legal advice: it says what the rule provides, for whom and from when, with a pointer to the article.

The calendar, read on the consolidated text

The regulation entered into force on the twentieth day after its publication in the Official Journal of the European Union (12 July 2024), that is on 1 August 2024, but it applies in stages. Article 113 sets a general date, 2 August 2026, and then the exceptions. I transcribed the dates of the consolidated version of 27 July 2026 into a small program, so that the machine counts the days left and not a hand-copied table that goes stale.

from datetime import date

# Dates from Article 113 of Regulation (EU) 2024/1689, consolidated text of 27.7.2026.
SCADENZE = [
    (date(2025, 2, 2), "Chapters I and II: prohibitions (Art. 5) and AI literacy (Art. 4)"),
    (date(2025, 8, 2), "general-purpose AI models, governance, penalties (not Art. 101)"),
    (date(2026, 8, 2), "general application; Art. 50; Art. 101 (fines on GPAI providers)"),
    (date(2026, 12, 2), "new prohibitions Art. 5(1), points (ba) and (bb); Art. 50(2) for systems already on the market"),
    (date(2027, 12, 2), "Annex III high-risk (Art. 6(2))"),
    (date(2028, 8, 2), "Annex I high-risk (Art. 6(1))"),
]

def stato(oggi):
    for quando, cosa in SCADENZE:
        giorni = (quando - oggi).days
        etichetta = "in force" if giorni <= 0 else f"in {giorni} days"
        print(f"{quando.isoformat()}  {etichetta:>16}  {cosa}")

stato(date(2026, 11, 13))

# Output (python3 -I, Python 3.14.4):
# 2025-02-02          in force  Chapters I and II: prohibitions (Art. 5) and AI literacy (Art. 4)
# 2025-08-02          in force  general-purpose AI models, governance, penalties (not Art. 101)
# 2026-08-02          in force  general application; Art. 50; Art. 101 (fines on GPAI providers)
# 2026-12-02        in 19 days  new prohibitions Art. 5(1), points (ba) and (bb); Art. 50(2) for systems already on the market
# 2027-12-02       in 384 days  Annex III high-risk (Art. 6(2))
# 2028-08-02       in 628 days  Annex I high-risk (Art. 6(1))

The listing is a list of (date, description) pairs plus a function that prints how many days are left to each. The parameter oggi ("today") is passed explicitly instead of reading the clock, and this matters: the output above is the one of 13 November 2026 and the run is repeatable, while a date.today() would make it different at every reading. The "in force" column means that the rule already applies; it does not mean that you are compliant.

Two things surprise whoever has read only the 2024 popularisation. The first: "general application from 2 August 2026" does not mean "high-risk from 2 August 2026", because Article 113 takes Sections 1, 2 and 3 of Chapter III, that is the obligations for high-risk systems, out of the general date. The second: the new prohibitions (Article 5(1), first subparagraph, points (ba) and (bb), on generating intimate images or child sexual abuse material) apply from 2 December 2026, while the original prohibitions and Article 4 apply from 2 February 2025. Article 111(4) also gives generative systems already on the market before 2 August 2026 until 2 December 2026 to comply with Article 50(2) (marking).

Your role: provider, deployer, and when you change role

The obligations depend on the role, not on the technology. The provider is whoever 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, whether for payment or free of charge. The deployer is whoever uses the system under its own authority, except in the course of a personal non-professional activity. There are also importers, distributors and authorised representatives. The scope is extraterritorial: Article 2(1) includes providers in third countries that place systems on the Union market, and providers and deployers in third countries where the output of the system is used in the Union. A server in the United States does not take you out of scope.

The role is not fixed. Article 25(1) says that a distributor, importer, deployer or other third party is considered a provider of a high-risk AI system in three cases:

  • you put your name or trademark on a high-risk system already placed on the market;
  • you make a substantial modification to a high-risk system, which remains high-risk;
  • you modify the intended purpose of a system not classified as high-risk (including a general-purpose one) in such a way that it becomes high-risk.

The third case is the one that concerns developers most closely: you take a generic model or service, configure it to select candidates or to assess creditworthiness, and from that moment you are the provider of a high-risk system, with the obligations of Article 16. The initial provider is no longer considered the provider of that specific system, but must cooperate and make documentation and technical access available (Article 25(2)).

Not a four-level pyramid: four regimes in the text

You often read about "four risk levels": minimal, limited, high, unacceptable. It is a popularisation shorthand that helps to remember the idea, but the text does not use it as its structure: it spreads the obligations over different chapters, and a system can fall under more than one. To decide what to do, it is better to reason by regime.

  • Prohibited practices (Chapter II, Article 5): for example systems that exploit the vulnerabilities of a person to distort their behaviour, social scoring with disproportionate effects or in unrelated contexts, and the inference of emotions in the workplace and in education institutions (except for medical or safety reasons).
  • High-risk (Chapter III): the systems of Annex I (safety components of products already covered by harmonisation legislation) and those of Annex III (areas such as employment, education, access to essential services).
  • Transparency (Chapter IV, Article 50): saying that one is interacting with an AI, marking synthetic content, disclosing deepfakes.
  • General-purpose AI models (Chapter V): obligations for whoever places the model on the market, not for whoever uses it downstream.

Not everything disappears outside these regimes: Article 4 requires providers and deployers to take measures to support the AI literacy of their staff, and since 27 July 2026 the wording has changed, because the rule specifies that the obligation does not require anyone to guarantee any specific level of AI literacy of any individual. Your internal code assistant, for example, has no regime of its own, but Article 4 applies all the same.

Classifying a use case: Annex III and the Article 6 derogation

A system is high-risk under Annex III if its intended use falls under one of the eight areas listed: biometrics, critical infrastructure, education, employment, access to essential services (including creditworthiness), law enforcement, migration, administration of justice. Article 6(3) provides for a derogation: the system is not considered high-risk if it does not pose a significant risk of harm to health, safety or fundamental rights, and in particular when it performs a narrow procedural task, improves the result of a previously completed human activity, detects deviations from prior decision-making patterns without replacing the human assessment, or performs a preparatory task. But the paragraph closes the door: if the system performs profiling of natural persons, it is always considered high-risk.

# Teaching simplification: NOT a legal classification.
# Maps keywords of the use case to the points of Annex III and says
# which question stays open (Article 6(3)).
ALLEGATO_III = {
    "biometrics": "point 1 (biometric identification and categorisation)",
    "critical infrastructure": "point 2 (safety components of critical infrastructure)",
    "education": "point 3 (admission, evaluation, proctoring of exams)",
    "recruitment": "point 4 (employment and workers' management)",
    "creditworthiness": "point 5(b) (creditworthiness)",
    "life insurance pricing": "point 5(c) (life and health insurance)",
    "law enforcement": "point 6 (law enforcement)",
    "migration": "point 7 (migration, asylum, border control)",
    "justice": "point 8 (administration of justice)",
}

ESCLUSIONI = {"fraud": "point 5(b) excludes the detection of financial fraud"}

def esamina(descrizione, ambito, profila_persone):
    if ambito in ESCLUSIONI:
        return f"{descrizione}: outside Annex III ({ESCLUSIONI[ambito]}); Art. 5 and Art. 50 still apply"
    punto = ALLEGATO_III.get(ambito)
    if punto is None:
        return f"{descrizione}: area '{ambito}' not in Annex III -> check Articles 5 and 50"
    if profila_persone:
        return f"{descrizione}: Annex III, {punto}; profiles people -> high-risk, no derogation (Art. 6(3))"
    return (f"{descrizione}: Annex III, {punto}; "
            "high-risk unless the Art. 6(3) derogation applies, to be documented (Art. 6(4))")

casi = [
    ("Credit scoring for mortgages", "creditworthiness", True),
    ("Fraud detection on cards", "fraud", False),
    ("Ranking CVs for a job offer", "recruitment", True),
    ("Extracting personal details from CVs", "recruitment", False),
    ("Internal code assistant", "software development", False),
]
for c in casi:
    print(esamina(*c))

# Output (python3 -I, Python 3.14.4):
# Credit scoring for mortgages: Annex III, point 5(b) (creditworthiness); profiles people -> high-risk, no derogation (Art. 6(3))
# Fraud detection on cards: outside Annex III (point 5(b) excludes the detection of financial fraud); Art. 5 and Art. 50 still apply
# Ranking CVs for a job offer: Annex III, point 4 (employment and workers' management); profiles people -> high-risk, no derogation (Art. 6(3))
# Extracting personal details from CVs: Annex III, point 4 (employment and workers' management); high-risk unless the Art. 6(3) derogation applies, to be documented (Art. 6(4))
# Internal code assistant: area 'software development' not in Annex III -> check Articles 5 and 50

The ALLEGATO_III dictionary ("Annex III") links a keyword to the point of the annex; the esamina ("examine") function applies, in order, the exclusion for fraud (point 5(b) excludes systems used to detect financial fraud), the lookup of the area and the profiling rule. The output shows four outcomes: a system that profiles (high-risk with no derogation), one that performs a narrow procedural task (high-risk unless a derogation applies, to be documented), one expressly excluded by point 5 (fraud) and one outside the annex. It is a keyword-based teaching simplification: the real judgement requires the text of the annex and the guidelines on Article 6 that the Commission was due to publish by 2 February 2026 (Article 6(5)), which I have not verified here. Whoever invokes the derogation must document the assessment before placing the system on the market and must still register in the EU database (Article 6(4) and Article 49(2)).

A case: credit scoring, from the provider's and the deployer's side

A system that evaluates the creditworthiness of natural persons is in Annex III, point 5(b), and by definition it profiles people: so it is high-risk. The obligations are split. The provider (whoever develops and sells the scoring engine) must meet the requirements of Section 2 of Chapter III and those of Article 16; the deployer (the bank that uses it) those of Article 26.

  • Provider: risk management system (Article 9), data governance (10), technical documentation (11), automatic recording of events (12), instructions for use (13), human oversight (14), accuracy, robustness and cybersecurity (15), quality management system (17).
  • Provider, before the market: conformity assessment (for points 2 to 8 of Annex III, by internal control under Annex VI, without a notified body), EU declaration of conformity, CE marking, registration in the EU database (Articles 43, 47, 48, 49).
  • Deployer: use according to the instructions, oversight assigned to competent persons, relevant and sufficiently representative input data, monitoring, keeping the logs for at least six months, information to the persons subject to decisions (Article 26(1), (2), (4), (5), (6), (11)).
  • Deployer of points 5(b) and (c): fundamental rights impact assessment before first use (Article 27).

The common thread is that the provider builds the conditions of control and the deployer exercises them: if the engine does not produce logs, the bank cannot keep them; if the instructions for use do not declare the accuracy (Article 15(3)), the deployer does not know what to monitor. This is why, as a developer, the part that concerns you most is Articles 12 to 15: they are design requirements, not paperwork. For financial institutions, Article 19(2) and Article 26 also provide that logs and monitoring can be absorbed into the obligations that already exist under Union financial services law. In Italy, Law 132/2025 (Article 20) leaves the Bank of Italy, CONSOB and IVASS the role of market surveillance authorities for these cases, in line with Article 74(6) of the regulation (my summary of the Italian text).

What this means in practice for the code

The requirements of Articles 9 to 15 translate into precise design choices, and it is worth reading them as non-functional requirements. The risk management system is an iterative process running throughout the entire lifecycle (Article 9(2)), not a launch document; testing is done against prior defined metrics and thresholds and before placing on the market (paragraph 8). The automatic recording of events over the lifetime of the system (Article 12) and the keeping of logs for a period appropriate to the intended purpose, in any case of at least six months unless otherwise provided by Union or national law (Article 19(1) and Article 26(6)), call for an architecture in which the log is an output of the system, not an add-on.

Human oversight (Article 14) deserves particular attention. The text requires that whoever oversees can understand the capacities and limitations of the system, remain aware of "automation bias" (the tendency to rely on the output), interpret it, disregard or override it, and stop the system with a "stop" button or a similar procedure. An interface showing only "approved/rejected" and a "confirm" button does not meet it; an interface that shows the factors, the level of uncertainty and a simple way to disagree comes closer. As for the "maximum time of 24 hours" or thresholds like "5% drift" found in many templates: the regulation does not provide for them, they are your own design choices and must be justified in the risk management system, not attributed to the law.

Penalties, without exaggeration

The penalties are administrative fines, not criminal ones, and the detailed rules are set by the Member States (Article 99(1)). There are three ceilings. For the prohibited practices of Article 5: up to EUR 35 000 000 or 7% of total worldwide annual turnover, whichever is higher. For non-compliance with most obligations of providers, deployers, importers, distributors and with the transparency obligations of Article 50: up to EUR 15 000 000 or 3%. For incorrect, incomplete or misleading information supplied to the authorities: up to EUR 7 500 000 or 1%. For SMEs, including start-ups, each fine is capped at the lower of the two amounts or percentages (paragraph 6). The ceilings are not expected penalties: the authority must take into account the nature, gravity and duration of the infringement, the size of the operator, the degree of cooperation and the technical and organisational measures implemented (paragraph 7). That last item is the only one you can influence with code.

Frequently asked questions

Does my customer-service chatbot fall under the AI Act?

Yes, it is an AI system, but the regime depends on the use. If it interacts directly with people, Article 50(1) requires that they be informed they are interacting with an AI, unless that is obvious to a reasonably well-informed, observant and circumspect person. It is not high-risk until it decides on an area of Annex III, and Article 4 on AI literacy applies anyway.

I use an open-source pre-trained model: do I have to prepare technical documentation?

It depends on your role on the system, not on the model. If you configure it for an Annex III use you become the provider of a high-risk system (Article 25(1), point (c)) and Article 11 applies to you; if you only use it for tasks outside the annex, it does not. The obligations on the model itself fall on whoever places it on the market (third article of the series).

Can I hand compliance over to a consultant?

You can hand over technical work, but the addressees of the obligations are the operators in their role (provider, deployer) and the Article 99 penalties hit them. The regulation does not provide for a transfer of responsibility by contract; the contract governs the internal relationship.

Does the AI Act apply if my server is outside the Union?

Yes, if you place the system on the Union market or if the output is used in the Union (Article 2(1), points (a) and (c)). Providers of high-risk systems established in third countries must also appoint, by written mandate, an authorised representative established in the Union (Article 22).

If my system is high-risk, do I have to be compliant today?

The high-risk obligations of Annex III apply from 2 December 2027 and those of Annex I from 2 August 2028. What does apply from today is the prohibitions of Article 5 and Article 4, and for systems already on the market before that date Article 111(2) provides that the regulation applies only if they undergo significant changes in their designs (for public authorities compliance is due in any case by 2 August 2030).

What you can do now

You can say which role you are in and when you change role; you can read the Article 113 calendar without confusing general application with the application of the high-risk obligations; you can apply Annex III and the Article 6 derogation to a use case and state the limits of your reasoning; and you know which requirements of Articles 9 to 15 become design requirements. The next step is to understand what happens when a system gets it wrong and someone is hurt: you will find it in article 02. For a case applied to the financial sector you can also read AI in Finance: Fraud Detection, Credit Scoring and Risk, and for the process side ML Governance: Compliance, Explainability and Ethics.

Sources

The two listings were run with Python 3.14.4 and the output shown is the real one, identical in two consecutive runs. The classifier is a teaching simplification, not a legal assessment. The guidelines under Article 6(5), the detailed Member State rules on penalties and the notion of "significant changes" in Article 111(2) are not defined in the passage I read.

Try These Tools

Free developer tools related to this topic.

Commenti

Caricamento commenti...

Sign in to comment

New articles and tools on Telegram too

The Telegram channel flags the site's new articles and tools.

Did you like this article?

Explore other content on the blog or discover my projects