The "Safety Component" Test Under the EU AI Act (Article 3(14))

If your AI is a "safety component", it can be high-risk regardless of whether it appears in Annex III — and it likely routes through a different deadline and a different conformity process than the Annex III systems people usually worry about. The trigger is a precise statutory definition in Article 3(14) of Regulation (EU) 2024/1689, and it bites in two places: the Article 6(1) product route for Annex I products, and point 2 of Annex III (critical infrastructure). This page gives you the test and a worked walk-through.

The one-line answer: a component is a "safety component" when it fulfils a safety function or its failure or malfunctioning endangers the health and safety of persons or property. Since the Digital Omnibus, AI used solely for non-safety purposes is excluded unless its failure would endanger health and safety. If your AI fits the test, you check Article 6(1) and Annex III point 2 before you conclude anything about risk tier.

The Article 3(14) definition

Article 3(14) was rewritten by the Digital Omnibus, Regulation (EU) 2026/1744 (in force since 27 July 2026). It now treats a component of a product or system as a safety component where it fulfils a safety function — its intended purpose is to prevent or mitigate risks to the health and safety of persons or property — or where its failure or malfunctioning endangers the health and safety of persons or property.

Two limbs, either of which is enough:

• It fulfils a safety function — the component's job is to keep the product or system safe (a braking-control module, a pressure cut-off, a collision-avoidance routine).
• Its failure endangers health, safety, or property — even if "keeping things safe" is not its primary purpose, if it breaks and people or property are put in danger, it qualifies.

The test is about function and consequence of failure, not about whether the word "safety" appears in the product spec. The Omnibus also added Article 6(1a): AI used solely for non-safety aspects — user assistance, performance optimisation, service efficiency, automation, convenience or quality control — is not a safety component. Article 6(1b) pulls it back in where its failure would endanger health and safety. So an AI module that optimises throughput on a production line is not a safety component merely because it sits in the machine; it becomes one if its malfunction can injure an operator.

Where the test triggers high-risk

Route 1 — Article 6(1): safety components of Annex I products

Article 6(1) makes an AI system high-risk when both conditions hold:

1. the AI system is itself a product, or is a safety component of a product, covered by the Union harmonisation legislation listed in Annex I; and
2. that product is required to undergo third-party conformity assessment under that Annex I legislation before being placed on the market. Under Article 6(1c), a third-party assessment required only for non-health-and-safety risks (radio spectrum or electromagnetic compatibility, for instance) does not meet this condition.

Annex I has two sections. Section A covers sectors like medical devices (MDR/IVDR), lifts, radio equipment, toys and pressure equipment; Section B covers vehicles, civil aviation, marine equipment, rail — and, since the Digital Omnibus, machinery (the Machinery Regulation (EU) 2023/1230, point 21). An AI safety component embedded in a CE-marked machine or medical device is high-risk via Article 6(1) — and it travels with that sector's deadline and conformity route, not the Annex III route.

For these product-embedded systems the deadline is 2 August 2028 (originally 2 August 2027; moved by Regulation (EU) 2026/1744). For Section A products, conformity is assessed through the product's own sectoral procedure, with the AI requirements integrated (Article 43(3)). For Section B products, including machinery, Article 2(2) applies only Article 6(1), Article 60a and Articles 102 to 112 directly; the AI requirements reach the product through the sectoral legislation — for machinery, delegated acts amending the Machinery Regulation, applying by 2 August 2028.

Route 2 — Annex III point 2: critical infrastructure

The safety-component concept also appears inside Annex III. Point 2 covers AI systems intended to be used as safety components in the management and operation of critical digital infrastructure, road traffic, or in the supply of water, gas, heating or electricity. Here the qualifier is the same Article 3(14) test: the AI must be a safety component of that infrastructure. A traffic-signal optimisation AI whose failure could cause collisions is in scope; an energy-billing analytics tool whose failure just produces wrong invoices is not — its malfunction does not endanger persons or property.

Annex III point 2 systems follow the Annex III stand-alone track: the deadline is 2 December 2027 (originally 2 August 2026; moved by Regulation (EU) 2026/1744), and conformity is the Annex VI internal-control route.

A decision walk-through

Work it in order for any AI you suspect might be a safety component:

1. Does failure endanger persons or property, or does the component fulfil a safety function? (AI used solely for optimisation, efficiency, convenience or quality control only counts if its failure endangers health and safety — Article 6(1a)–(1b).) If no to both — it is not a safety component; classify it the ordinary way (intended purpose, Article 5, Annex III use-cases). If yes, continue.
2. Is it embedded in, or itself, an Annex I product that needs third-party conformity assessment for health and safety reasons? If yes → high-risk via Article 6(1); sectoral conformity route (Section A) or requirements via the sectoral law (Section B, including machinery); 2 August 2028.
3. If not Annex I — is it a safety component of critical infrastructure (digital infra, road traffic, water/gas/heating/electricity)? If yes → high-risk via Annex III point 2; Annex VI internal control; 2 December 2027.
4. Neither? It is a safety component in the everyday sense but outside both high-risk routes; finish the ordinary classification — it may still be limited-risk under Article 50 or minimal.

Document the reasoning at each step; the record is your evidence of diligence if a market-surveillance authority asks.

Worked example: Adriatic Robotics, ~140 employees, ~€26M turnover

Adriatic Robotics, a Trieste manufacturer, builds an AI vision module that detects when a worker's hand enters the danger zone of an industrial press and halts the stroke. It sells the press, with the module embedded, as a CE-marked machine.

Walking the test:

1. Safety function / failure consequence? Yes — the module's whole job is to stop the press to protect the operator; its failure could crush a hand. It is squarely an Article 3(14) safety component.
2. Annex I product needing third-party conformity assessment? The press falls under the Machinery Regulation (EU) 2023/1230 (now Annex I Section B), and a machine-learning safety component of this kind requires third-party conformity assessment for safety reasons. Both Article 6(1) conditions are met.

Verdict: the vision module is high-risk under Article 6(1). Because machinery now sits in Section B, the AI Act's high-risk requirements do not apply to it directly: they reach Adriatic through delegated acts amending the Machinery Regulation, applying by 2 August 2028, and its conformity work runs through the Machinery Regulation's procedure. Adriatic, building and branding the machine, carries those duties as the manufacturer and AI provider; a factory that buys and operates the press is the deployer. Notably, Adriatic does not go looking in Annex III — the product route settles it.

How Confir helps

Confir's rule-based engine runs the Article 3(14) test first — failure-consequence and safety-function limbs — then routes a qualifying component to Article 6(1) (with the right Annex I section, sector and 2 August 2028 deadline) or to Annex III point 2 (2 December 2027), and records which rule fired. The logic is deterministic: the same intake always returns the same route, deadline and conformity path, so the safety-component finding is auditable rather than a judgement call buried in a spreadsheet.

Frequently Asked Questions

What makes an AI component a "safety component" under the EU AI Act?

Article 3(14), as rewritten by the Digital Omnibus: a component is a safety component if it fulfils a safety function for a product or system, or if its failure or malfunctioning endangers the health and safety of persons or property. Either limb is enough. AI used solely for user assistance, performance optimisation, efficiency, automation, convenience or quality control is excluded (Article 6(1a)) — unless its malfunction can endanger health and safety, as with an efficiency module that can injure an operator (Article 6(1b)).

Does being a safety component automatically make my AI high-risk?

Not by itself. It makes you check two routes. If the component is in an Annex I product that needs third-party conformity assessment, it is high-risk under Article 6(1). If it is a safety component of critical infrastructure (road traffic, water, gas, heating, electricity, digital infrastructure), it is high-risk under Annex III point 2. A safety component outside both routes is not high-risk on that basis alone.

What deadline applies to a safety-component AI?

It depends on the route. A safety component embedded in an Annex I product follows the product route, with a deadline of 2 August 2028 (the statute originally read 2 August 2027). A safety component of critical infrastructure under Annex III point 2 follows the stand-alone track, deadline 2 December 2027 (statute originally read 2 August 2026). Both dates were set by the Digital Omnibus, Regulation (EU) 2026/1744, in force since 27 July 2026.

How is a safety-component AI in an Annex I product assessed for conformity?

Through the product's own sectoral legislation. For Annex I Section A products such as those under the MDR or IVDR, the AI Act's high-risk requirements are integrated into that existing conformity assessment (Article 43(3)). For Section B products — which since the Digital Omnibus include machinery — the requirements arrive through the sectoral law itself, for machinery via delegated acts amending the Machinery Regulation by 2 August 2028. Either way you do not run a separate stand-alone AI Act conformity assessment; the AI requirements are folded into the conformity or type-approval process the product already goes through.

Is every AI in critical infrastructure a safety component?

No. Annex III point 2 only catches AI used as a safety component of the infrastructure — where failure endangers persons or property or disrupts the supply itself. An analytics or billing tool whose malfunction merely produces wrong numbers is not a safety component and is not high-risk under point 2. Apply the Article 3(14) failure-consequence test before concluding.

Who is the provider and who is the deployer for a safety-component AI?

The party that develops the system and places it on the market under its own name is the provider (Article 16) — typically the manufacturer embedding the component. The organisation that uses the product in its operations is the deployer (Article 26). Note Article 25: a deployer that substantially modifies the system or rebrands it, or changes its intended purpose so that it becomes high-risk can flip into provider status. For Annex I Section B products such as machinery, these duties arrive through the sectoral legislation rather than applying directly.

Related guides

• Safety component
• How to classify your AI system
• Is my AI high-risk?
• The Article 6(3) exemption explained
• Article 6 — high-risk classification
• Annex III high-risk use cases