How to Classify Computer Vision AI Under the EU AI Act

"Is computer vision high-risk?" has no single answer, because computer vision is not a risk category — it is a technology that shows up in every risk tier. The EU AI Act classifies by use, never by the underlying technique (Article 6). The same convolutional model that inspects bottle caps on a production line, reads a chest X-ray, recognises a face in a crowd, and steers a car sits in four completely different regulatory places.

The short map: face recognition is high-risk under Annex III point 1 — or, in one real-time public-space configuration, prohibited under Article 5(1)(h). Defect and quality inspection is minimal-risk. Medical imaging is high-risk via the product route (Annex I, the Medical Device Regulation), with a 2 August 2028 deadline. Driver-assistance vision is also a product-route case (vehicle type-approval). One technology, four answers. This page is the decision walk-through that gets you to the right one.

---

The One Rule: Classify by Use, Not by the Model

The Act's classification machinery never asks "is this a vision model?" It asks what the system is used for. Article 6 sets two high-risk routes — 6(1) for AI that is a safety component of a product covered by the Annex I sectoral laws (medical devices, vehicles, machinery), and 6(2) for AI used in one of the Annex III areas. Anything in neither route, and not prohibited under Article 5, is limited- or minimal-risk.

So the classification question for any computer-vision system is a sequence, not a label:

1. Is the use prohibited under Article 5? (e.g. real-time remote biometric identification in public for law enforcement; sensitive-attribute categorisation; social scoring.) If yes — stop, it is banned.
2. Is it a safety component of an Annex I product? (medical device, vehicle, machinery) If yes — high-risk via the product route, 6(1), 2 August 2028.
3. Is the use named in Annex III? (biometrics, critical infrastructure, education, employment, essential services, law enforcement, migration, justice) If yes — high-risk, 6(2), 2 December 2027 (originally 2 August 2026; moved by the Digital Omnibus).
4. None of the above? Limited-risk if an Article 50 transparency duty applies; otherwise minimal-risk, with no mandatory obligations beyond Article 4 AI literacy.

Run any vision system through those four questions and the tier falls out.

---

Face Recognition: High-Risk, or Prohibited

Facial-image computer vision is the case people mean when they ask "is computer vision high-risk?" — and even here the answer splits.

Remote biometric identification (one-to-many) — matching a face against many records to establish who a person is — is high-risk under Annex III point 1(a), with the notified-body conformity route (Annex VII).

Real-time remote biometric identification in publicly accessible spaces for law-enforcement purposes is largely prohibited under Article 5(1)(h), subject to narrow authorised exceptions.

Sensitive-attribute categorisation from a face — inferring ethnicity, religion, sexual orientation — is prohibited under Article 5(1)(g).

One-to-one verification — face-unlock against your own enrolled template — is excluded from Annex III point 1 and is minimal-risk (GDPR Article 9 still applies). So even within "face recognition", you can be prohibited, high-risk, or minimal depending on one-to-many vs one-to-one and public-real-time vs not.

---

Defect Inspection: Minimal-Risk

The most common industrial computer-vision use is also the easiest call. A vision model that inspects products for defects — scratches on a panel, a missing weld, a mislabelled box — and routes rejects off a line, is minimal-risk. It is not in any Annex III area (it assesses objects, not people), and it is not a safety component of an Annex I product in the regulated sense; it is a quality-control tool. Since the Digital Omnibus, Article 6(1a) says so expressly: AI used solely for quality control is not a safety component, unless its failure would endanger health and safety (Article 6(1b)).

Minimal-risk means no mandatory AI Act obligations beyond Article 4 AI literacy — no conformity assessment, no registration, no FRIA. The sensible voluntary moves (an internal log, basic accuracy testing) are good practice, not legal duties. The line to watch: if the same camera is repurposed to monitor workers' productivity or behaviour, it leaves quality control and enters Annex III point 4 (employment / worker management) — high-risk. The object/person distinction is what keeps inspection minimal.

---

Medical Imaging: High-Risk via the Product Route

AI that reads medical images — detecting a tumour on a scan, triaging a chest X-ray — is high-risk, but not through Annex III. It reaches high-risk through Article 6(1): the AI is a safety component of, or is itself, a medical device regulated under the Medical Device Regulation (EU) 2017/745 (MDR) (or the IVDR), which is listed in Annex I.

That routing changes three things. First, the deadline is 2 August 2028, not the Annex III date. Second, the AI Act's high-risk requirements are integrated into the existing MDR conformity assessment — the notified-body review a medical-device manufacturer already runs — rather than a separate AI Act procedure. Third, the manufacturer is squarely the provider, with the hospital as deployer. A radiology-AI vendor does not pick between MDR and the AI Act; it satisfies the AI Act requirements within its MDR conformity route.

---

Driver-Assistance Vision (ADAS): The Product Route Again

The cameras in an advanced driver-assistance system — lane-keeping, pedestrian detection, automatic braking — are computer vision performing a safety function in a vehicle. Like medical imaging, this is the Article 6(1) product route: vehicles are type-approved under EU automotive legislation listed in Annex I, and the AI Act's high-risk requirements are integrated through that sectoral framework, with the 2 August 2028 deadline.

The practical consequence mirrors medical devices: the vehicle (or component) manufacturer is the provider; conformity runs through type-approval, not a stand-alone AI Act assessment; and the Annex III timeline does not apply. ADAS vision is a good reminder that "high-risk" does not always mean "Annex III" — the product route is a parallel high-risk track with its own deadline.

---

The Decision Table

Two date tracks live in that table: stand-alone Annex III high-risk (2 Dec 2027, originally 2 Aug 2026) and product-embedded high-risk (2 Aug 2028, originally 2 Aug 2027) — both set by the Digital Omnibus, Regulation (EU) 2026/1744, in force since 27 July 2026. Keep them distinct.

---

Worked Example: OptiSicht GmbH, a Machine-Vision Company

OptiSicht GmbH, a German machine-vision company, ~95 employees, ~€16M turnover, sells one model family across three products. Same core technology, three classifications.

• FactoryEye inspects automotive parts on a line for defects. It assesses objects. Minimal-risk — no AI Act obligations beyond Article 4 AI literacy, though OptiSicht keeps voluntary accuracy logs for its customers.
• MediScan flags suspected fractures on X-rays and is CE-marked as a medical device under the MDR. High-risk via the product route (Article 6(1)), deadline 2 August 2028; OptiSicht meets the AI Act requirements inside its MDR conformity assessment, as provider.
• GateGuard matches faces at a client's secure entrance against an enrolled gallery of staff to log who entered. Because it searches a gallery to establish identity, it is one-to-many remote identification — high-risk under Annex III point 1(a), requiring the Annex VII notified-body route and registration under Article 49. (Had it instead checked each person's face against their own badge template, it would be 1:1 verification — minimal-risk.)

OptiSicht's mistake-to-avoid is treating "we're a computer-vision company" as a single compliance posture. Three products, three programmes, two deadlines. A breach of the high-risk obligations on MediScan or GateGuard would sit at €15 million or 3% of worldwide turnover (Article 99(4)), with the SME cap under Article 99(6) taking the lower figure; FactoryEye carries no such exposure because it is minimal-risk.

---

How Confir Helps

Confir's rule-based engine runs each computer-vision system through the Article 6 sequence — prohibition test (Article 5), product route (Annex I / Article 6(1)), Annex III test (Article 6(2)), then transparency (Article 50) — and returns a tier per use, not per company. So a vision firm with an inspection tool, a medical-imaging device, and a face-recognition product sees three separate findings with three deadlines, each citing the rule that fired. The classification is deterministic and rule-based: the same intake produces the same tier every time, which is what lets you show an auditor why defect inspection is minimal while face identification is high-risk.

Related guides

• How to classify your AI system
• Biometric identification (Annex III)
• Medical imaging AI
• Article 6 classification