Skip to main content

The two gates: is it an AI system, and does the Act apply to your organization at all

The short answer

Applicability is two gates, run in order

Gate one asks whether the thing is an AI system under Article 3(1); gate two asks whether the Act reaches your organization for it under Article 2. Both must open before you owe the Act anything for that system. Collapsing the two questions is the root of most confusion about the law.

What you will be able to do

  • Distinguish an "AI system" as the EU AI Act defines it in Article 3(1) from ordinary software the law does not reach, using the seven elements of the definition and the European Commission's 2025 guidance on it.
  • Analyze whether the EU AI Act applies to your organization at all under Article 2 (scope), separating the two independent questions: is it an AI system, and does the Act reach this organization for this system.
  • Identify which role your organization plays for each system, provider, deployer, importer, or distributor, because the role decides which obligations you carry.
  • Apply the extraterritorial-reach test in Article 2(1)(c): the Act can reach a provider or deployer outside the EU whenever the output the system produces is used inside the EU.
  • Recognize the main exclusions in Article 2, including military, defense, and national-security use, scientific research and development, purely personal use, and free and open-source systems, and apply them without over-reading them.
  • Detect the trap in Article 25 where a deployer can quietly become a provider, and carry the heavier obligations, by rebranding a system, changing its purpose, or substantially modifying it.
  • Produce an EU AI Act applicability memo that runs both gates against your own AI systems inventory and records, per system, a defensible answer and your role.
  • Justify, to a skeptical executive who insists "we are a small company, this European law cannot possibly apply to us," why physical location is not the test, and what the test actually is.
  • Separate the applicability question (does the Act reach this system) from the risk question (how heavily does it regulate it), so you run them in the right order and never classify the risk of a system the Act does not reach (see Topic 5.3).
  • Treat applicability as a living determination, recording per system the triggers, a new market, a rebrand, a repurpose, a vendor's model update, that would change the answer and require the memo to be reopened.
  • Connect the answers you write to the artifacts around them: the systems inventory this memo consumes (see Topic 0.2), and the risk classification (see Topic 5.3) and conformity file (see Topic 5.6) it feeds.

The ideas, one by one

Gate one turns on "infer," not on the label

An AI system is machine-based, somewhat autonomous, and, above all, infers its outputs from data beyond fixed human rules (Article 3(1)). Marketing that says "AI" is not the trigger, and a contract that says "automation" is not a defense. Run the infer test on the substance.

Adaptiveness is not required

The definition says a system "may" adapt, and the Commission's 2025 guidelines confirm a frozen model that never learns after deployment can still be an AI system. Do not disqualify a system at gate one just because it does not learn on the job.

The Act is built on reach, not address

Article 2(1)(a) reaches providers whether in the EU or a third country, and Article 2(1)(c) reaches third-country providers and deployers whenever the output is used in the Union. "We have no European office" is not a defense; it is the DeepSeek mistake.

Scope is a set of roles, and the role decides your obligations

Article 2 lists providers, deployers, importers, and distributors; Article 3 defines each. The provider carries the heaviest build-side duties, the deployer real but lighter use-side ones. Name your role per system, because one organization holds different roles for different systems.

A deployer can become a provider without buying anything

Under Article 25, putting your own name on a high-risk system, substantially modifying it, or repurposing it into a high-risk use promotes you to provider and its full obligations. Set this tripwire before a product decision trips it for you.

The exclusions are narrow and bounded by a single word

Military ("exclusively"), research ("sole purpose"), personal ("purely"), and open source (carved back to high-risk and Articles 5 and 50) each collapse the moment the system does real work on real people. If you cannot justify an exclusion in one specific sentence, you do not have it.

Effect opens the gate, not intention

Article 2(1)(c) turns on the output being used in the EU, whether or not you aimed it there. If your system's output reaches EU users in fact, the gate opens. Intent is not the test.

Run both gates against your whole inventory, not just the labels

The false negative, missing an unbranded model that infers and affects people, is more dangerous than the false positive, and the only guard is to run gate one against every system you listed in Module 0, not only the ones marked "AI."

Over-scoping is a misallocation, not caution

Declaring every spreadsheet an AI system drains the capacity your genuine AI needs. Precision at gate one, excluding what does not infer, is what lets you spend real effort on the systems that actually matter.

Applicability is a living determination

A vendor adds a model, you enter the EU market, you rebrand or repurpose a system, and the answer changes. Record the triggers that would change each answer and revisit the memo on a schedule and after any of them. A stale memo defends nothing.

A deployer role is not a zero-obligation role

"The vendor is the provider" correctly locates the heaviest build-side duties, but deployers still must use the system per its instructions, ensure human oversight, monitor it, and inform affected people where required. A use-side failure on a system you deployed is your incident, whoever built it.

Applicability is per system, not per company

Role and reach differ system by system, so one organization can be a deployer of one, a provider of another, a distributor of a third, and out of scope for a fourth. A single company-wide label misstates the obligations for most of your systems; run the gates and record the answer one system at a time.

The two gates are an operational skill, not a lawyer's performance

They rest on facts you hold, whether a system infers, where its output lands, whose brand is on it, and how you use it. Legal confirms and refines, but you can run and defend the sequence yourself, which is the whole point of doing it before anyone asks.

This memo is the ground floor of the module

Your risk classification grades the systems this memo confirms are in scope, your literacy program is scoped to the roles it names, and your conformity file uses it as proof you knew your role from the start. Every later EU AI Act decision stands on the answers you write here.

Listen to it

The same lesson, talked through by two hosts as episode 34 of the podcast.

Real cases

These examples show the two gates operating across different regions, sectors, and years. Each is attributed. Where a case is owned in depth by a later topic, it is pointer-referenced, not retold here.

Example 1: DeepSeek and the Italian Garante, Italy and China, 2025 (the anchor). On 30 January 2025 the Garante ordered Hangzhou DeepSeek and Beijing DeepSeek to stop processing Italian users' personal data, with immediate effect, after the companies' answers to its 28 January information request were found "insufficient and inadequate" (Garante, 30 January 2025; Bird & Bird, 2025). The companies' central argument was a gate-two claim: they did not operate in Italy, had pulled the app, and were not subject to the GDPR. The Garante held that European law reached them anyway, because their processing reached people in Italy, and their own privacy policy conceded that data was stored in China (Euronews, 31 January 2025). The case was decided under the GDPR, not the EU AI Act, but it is the clearest recent demonstration of the reach principle that the AI Act's Article 2 encodes: a foreign AI provider cannot answer a European regulator with "we are not here." The teachable core: the most important question was not what DeepSeek did, but whether the law reached it, and the company got that gate wrong in public.

Example 2: The vendor tool nobody called AI, any organization, ongoing (established pattern). A common gate-one case has no headline. An organization buys a product to screen job applicants or score fraud risk. The contract calls it "automation" or "smart matching," never "AI." Under the hood is a machine-learning model that infers its outputs from data, which means it passes gate one whatever the paperwork says, and if it drives decisions about people it may be exactly the high-risk kind the Act watches most closely (see Topic 5.3) (see Topic 5.4). Organizations that scan only for systems labeled "AI" miss these entirely. This is the false negative at gate one, and it is the most frequent applicability error in practice. (Marked established as a widely observed pattern; specific products vary.)

Example 3: The GDPR reach precedent, EU and United States, 2018 onward (established). The reach principle did not begin with the AI Act. Since 2018 the GDPR's Article 3 has reached organizations with no EU establishment when they offer goods or services to, or monitor the behavior of, people in the Union. European authorities have applied it consistently to United States and other non-EU firms, establishing years before the AI Act a predictable scope rule: "we have no European office" is not a defense in EU digital law. The AI Act's Article 2(1)(c) deliberately carries the same logic into the AI domain. This example is named to show that the reach model is settled European practice, not a novelty of the AI Act.

Example 4: The deployer who became a provider, global, ongoing (established pattern). A frequent Article 25 case: an organization licenses a general-purpose model, wraps it in its own product, puts its own brand on the result, and points it at a regulated, high-stakes decision. It thinks of itself as a user, a deployer. Under Article 25 it may have become the provider of a high-risk system, inheriting the technical documentation, risk management, and conformity obligations it never planned for. The upstream relationship, what the foundation-model vendor owes and what the deployer must verify, is owned by a later topic (see Topic 5.5); it is named here only to show that role is a determination, not a given, and that gate two can flip a company's obligations without it noticing.

Example 5: The over-scoping organization, any sector, ongoing (established pattern). The opposite failure is quieter and rarely makes news, but it is real. An organization, frightened by the Act, declares that everything digital it runs is an "AI system" and tries to build conformity files for spreadsheets, rule-based workflow tools, and simple automations that never inferred anything. It exhausts its governance capacity on systems gate one should have excluded, and the genuine AI, the model that actually infers and affects people, gets the leftover attention. This is the false positive at gate one, and it is why the "infer" test matters as much for what it excludes as for what it catches. (Marked established as a widely reported compliance pattern.)

Example 6: The narrow exclusion stretched too far, global, ongoing (established pattern). A recurring error is to reach for an Article 2 exclusion and read it far wider than it goes: treating a dual-use system as exempt because it has a defense customer, or a deployed product as "research" because it began in a lab, or an open-source component as fully out of scope when the high-risk and transparency obligations still apply. Each of these collapses under the exclusion's own limiting word, "exclusively," "sole," or the open-source carve-back to Articles 5 and 50. The pattern to learn: an exclusion you cannot defend in one specific sentence is an exclusion you do not have.

Example 7: The frozen model wrongly waved through, any sector, ongoing (established pattern). A team runs a machine-learning model that was trained once and then frozen, a credit-scoring or hiring model that never updates on the job. Because "real AI learns" is a common intuition, the team decides the frozen model is "just software" and fails it at gate one. That is a misreading of Article 3(1), which says a system "may" exhibit adaptiveness, not that it must, and which the Commission's 2025 guidelines confirm. The frozen model infers its outputs from data, so it passes gate one, and it may be exactly the high-stakes kind the Act scrutinizes. Freezing a model does not remove it from scope; it only makes its behavior more stable, which is often why it was frozen. (Marked established as a common definition error.)

What the seven examples share. Across a Chinese chatbot, an unbranded vendor model, a settled data-protection precedent, a repurposed general model, an over-scoping compliance team, an over-stretched exclusion, and a wrongly excluded frozen model, the mechanism is the same: applicability is not obvious, and it is not about where you sit or what the marketing says. It is a two-gate determination, is it an AI system, and does the Act reach us and in what role, that has to be run on evidence, per system, and written down. Every failure above came from skipping a gate, collapsing the two, misreading the definition, or answering from address instead of reach.

Where people go wrong

  • "We have no office in Europe, so this European law cannot apply to us." This is the DeepSeek mistake, made in writing. The EU AI Act is built on reach, not address. Article 2(1)(a) reaches providers whether or not they are established in the EU, and Article 2(1)(c) reaches third-country providers and deployers whenever the output is used in the Union. The question is never where you sit; it is whether your system reaches into the EU. Answer from location and a regulator will correct you the way the Garante corrected DeepSeek.
  • "It has AI in the name, so it is an AI system under the Act." Marketing is not the test. Gate one turns on Article 3(1), and specifically on whether the system infers its outputs from data beyond fixed human rules. A rule-based tool branded as AI fails gate one; you owe the Act nothing for it, and treating it as in scope wastes governance capacity the real AI needs. Run the "infer" test, not the label test.
  • "It is just automation, not AI, so gate one is shut." The opposite error, and the more dangerous one. A vendor tool sold as "automation" or "smart matching" may be a machine-learning model that infers, which passes gate one whatever the contract calls it. The unbranded model buried in a hiring or fraud process is exactly the kind the Act cares about most. Look under the label, not at it.
  • "If the Act applies to us, we are just 'in scope' and everyone in scope has the same duties." Scope is not a single status; it is a set of roles. Article 2 lists providers, deployers, importers, and distributors, and Article 3 defines each. The provider carries the heaviest build-side obligations; the deployer carries real but lighter use-side ones. Failing to name your role per system means you cannot know which obligations are actually yours.
  • "We only use the tool, so we are a deployer and always will be." Article 25 can promote you from deployer to provider without you buying anything, if you put your own name on a high-risk system, substantially modify it, or repurpose it into a high-risk use. Organizations that wrap a general model in their own brand and point it at a regulated decision often become providers without noticing. Role is a live determination, not a permanent label.
  • "The system frozen and never learns after deployment, so it is not really AI." Article 3(1) says a system "may" exhibit adaptiveness, and the Commission's 2025 guidelines confirm adaptiveness is not required. A deliberately frozen model that infers its outputs is still an AI system. Do not disqualify a system at gate one just because it does not learn on the job; many high-stakes models are frozen on purpose.
  • "We do research with AI, so the research exclusion covers us." The Article 2(6) exclusion is for systems developed and used for the sole purpose of scientific research and development, and Article 2(8) covers pure development before market. Involving real people, a clinical trial, a study, does not by itself break the exclusion; what breaks it is turning the system to a real, non-research deployment the research was not for, or testing it in real-world conditions, which Article 2(8) keeps covered. "It started in a lab" is not a defense once the system is doing that other job.
  • "It is open source, so the Act does not apply." Article 2(12) excludes free and open-source systems, but carves back in anything placed on the market as high-risk, and anything under the prohibited-practice rules (Article 5) or transparency rules (Article 50). "Open source" is not a blanket exemption; the high-risk and transparency obligations can still bind you. Check the carve-back before relying on the exclusion.
  • "A defense customer buys it, so it is exempt as military use." Article 2(3) exempts systems used exclusively for military, defense, or national-security purposes. "Exclusively" is the constraint. A dual-use system you also sell or use commercially is not exempt because one customer is a ministry. An exclusion bounded by "exclusively" or "sole" collapses the moment the system does anything outside that sole purpose.
  • "We answered the applicability question last year, so it is settled." Both gates can change. A vendor adds a model to a rule-based tool; you expand into the EU and your output now lands there; you rebrand or repurpose a system and cross the Article 25 line. Applicability is a living determination that must be revisited on a schedule and after any market, branding, purpose, or modification change. A stale memo defends nothing.
  • "Deciding scope is legal work, so it is not my job as an operator." The two gates are operational judgments grounded in what your systems actually do and how you use them, and that is knowledge you hold, not your lawyers. Legal counsel confirms and refines; you are the one who can say whether a system infers, where its output lands, and what role you play. Handing the whole question to legal, with no operational input, is how organizations misstate their own reach.
  • "We should over-scope to be safe: call everything AI and comply with everything." Over-scoping is not caution; it is misallocation. Building conformity artifacts for spreadsheets and rule-based automations exhausts the capacity the genuine AI needs, and it trains the organization to treat the whole exercise as theater. Precision at gate one, excluding what does not infer and catching what does, is what lets you spend real effort on the systems that actually matter.
  • "The output is used in the EU, but we did not intend that, so it does not count." Article 2(1)(c) turns on the output being used in the Union, not on your intention. If your system's output reaches EU users in fact, the gate opens whether or not you aimed it there. Intent is not the test; effect is. This is the reach principle again, and it does not wait for you to mean it.
  • "The vendor is the provider, so as the deployer we carry nothing." Deployers carry real obligations under the Act, using the system per its instructions, ensuring human oversight, monitoring it, and informing affected people where required. "The vendor is the provider" correctly locates the heaviest build-side duties, but it does not empty your own. Treating a deployer role as zero-obligation is how a use-side failure, poor oversight of a system you deployed, becomes your incident regardless of who built it.
  • "One applicability answer for the whole company is enough." Applicability is per system, not per company, because role and reach differ system by system: you may be a deployer of one, a provider of another, and out of scope for a third. A single company-wide label misstates the obligations for most of your systems. Run the gates and record the answer one system at a time.

Questions people ask

What is EU AI Act (Regulation (EU) 2024/1689)?
The European Union's horizontal law on artificial intelligence, published in the Official Journal on 12 July 2024 and in force from 1 August 2024, with obligations applying in phases. It regulates AI by risk and by the role an organization plays. This topic covers the two threshold gates that decide whether it applies to you at all. More on EU AI Act (Regulation (EU) 2024/1689)
What is AI system (Article 3(1))?
A machine-based system that operates with some autonomy, may exhibit adaptiveness, and, for explicit or implicit objectives, infers from its inputs how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments. The capacity to infer, beyond fixed human rules, is the dividing line.
What is gate one?
The first threshold question of this topic: is the system an "AI system" under Article 3(1)? A question about the technology, answered with the seven-element definition and the "infer" test. If gate one is shut, the Act does not apply to the system.
What is gate two?
The second threshold question: given that it is an AI system, does the EU AI Act reach your organization for it under Article 2, and in what role? A question about reach, role, location, where the output lands, and any exclusion.
What is infer (the "infer" test)?
The element of the Article 3(1) definition, and the practical dividing line, that a system derives or computes its outputs from the inputs it receives in a way that goes beyond fixed, human-written rules. A model that learned a pattern from data infers; a hand-coded rule does not.

Keep going

Watch the lesson and prove it

Proving it happens inside the program: watch, explain it in your own words, and it goes on a record employers can check.

Watch the lesson and prove it →