AI Governance for Manufacturers: One Model, Two Routes to High-Risk

A manufacturer's AI governance problem is not that it has a lot of AI. It is that its AI sits on two completely different legal tracks at once, and a governance model built for one will miss the other. Safety-function AI inside machinery routes through EU product-safety law as an Annex I product. The HR AI in the back office routes through Annex III point 4 as a stand-alone high-risk system. Different routes, different deadlines, different owners — and a single governance model has to hold both without confusing them.

This page sets out that model: how a manufacturer maps its AI across the two high-risk routes, where the bulk of factory AI actually lands (minimal-risk), and who owns what.

Start from the truth: most factory AI is minimal-risk

Before the high-risk routes, set expectations. Predictive maintenance, quality-control vision, demand forecasting, energy-optimisation, process tuning — these are operational tools. None of them appears in Annex III, and none performs a safety function inside a regulated product. Under the EU AI Act, Regulation (EU) 2024/1689, they are minimal-risk: no mandatory compliance stack. A governance model that treats every model on the shop floor as a compliance object will drown in paperwork and miss the two systems that actually matter.

So the governance job is triage, not blanket coverage. Two routes catch the high-risk systems; everything else is documented in the inventory and left alone.

Route 1 — Machinery safety-component AI: the Annex I product route

This is the route most specific to manufacturers, and the one most other industries do not have.

Under Article 6(1), an AI system is high-risk if it is a safety component of a product covered by EU harmonised legislation listed in Annex I of the AI Act, where that product must undergo third-party conformity assessment. The Machinery Regulation (EU) 2023/1230 — which replaced the old Machinery Directive and covers industrial machinery, robots and cobots — is an Annex I instrument. So AI that performs a safety function inside such machinery is high-risk via the product route.

A safety component is AI whose failure or malfunction could endanger the health or safety of persons. In a factory the clear cases are:

• AI enforcing safety zones between a collaborative robot and a human operator — detecting proximity, commanding speed reduction or stop
• AI controlling emergency-stop logic or guard interlocks
• AI governing pressure, temperature or chemical dosing where exceedance is hazardous

By contrast, AI that optimises a tool path, schedules maintenance, or tunes throughput is performing an operational function — minimal-risk. The same sensor data can feed a minimal-risk operational prediction and a high-risk safety command; classification follows the function, not the hardware.

The deadline for this route is 2 August 2028 — the Digital Omnibus date for high-risk AI that is a safety component of Annex I products, now formally adopted (deferred from the statutory 2 August 2027 for this track). The machine's OEM is the provider (Article 16); the factory operating the machine is the deployer (Article 26). The AI Act conformity evidence is integrated into the same Machinery Regulation technical file, but the two are separate legal requirements — CE marking under one does not discharge the other.

Route 2 — HR and workforce AI: Annex III point 4

The second route has nothing to do with the shop floor. Annex III point 4 makes AI used in employment and worker management high-risk: screening job applicants, driving promotion or termination decisions, allocating tasks through automated systems. A manufacturer that uses a CV-screening tool for factory hires is running an Annex III high-risk system regardless of what the rest of its AI does. Here the manufacturer is usually the deployer of a vendor tool (Article 26).

The deadline for stand-alone Annex III high-risk is 2 December 2027 under the Digital Omnibus deferral, now formally adopted (deferred from the statutory 2 August 2026). Note the two routes carry different dates — Annex I machinery AI is 2 August 2028; Annex III HR AI is 2 December 2027 — which is exactly the kind of distinction a single governance model has to keep straight.

One hard line on this route: Article 5(1)(f) prohibits emotion recognition in the workplace. A system inferring workers' emotional states on the line — sold as fatigue detection, attentiveness monitoring or wellbeing — is prohibited outright, not high-risk, in force since 2 February 2025, at the €35,000,000 or 7% tier (Article 99(3)). Catch it at intake, not after deployment.

The two routes side by side

A governance model that covers both

The model is one inventory, two classification paths, clear ownership.

1. One inventory. Every AI system — operational, safety, HR — in a single register, so nothing falls between the two tracks.
2. Route the classification. For each system ask first: is it a safety component inside Annex I machinery? If yes, Route 1. If no, is it an Annex III use (HR, critical-infrastructure safety)? If yes, Route 2. If neither, minimal-risk — document and move on.
3. Assign the role per system. Machinery AI you build = provider; machinery AI you only operate = deployer; vendor HR tools = deployer. The role drives the obligations.
4. Run the high-risk stack only where it bites. Article 9 risk management, Article 14 human oversight, Article 43 conformity, Article 72 post-market monitoring and Article 73 incident reporting apply to the high-risk systems — not the minimal-risk majority.
5. Track both deadlines distinctly. 2 August 2028 for machinery AI, 2 December 2027 for HR AI, both flagged "Digital Omnibus, adopted."

Worked example: Bergmann Antriebstechnik GmbH

Bergmann Antriebstechnik GmbH, a Baden-Württemberg drive-systems maker with about 320 employees and roughly €74M turnover, inventories 26 AI systems. Twenty-one are minimal-risk: predictive-maintenance models, weld-vision QC, an energy-optimiser. One is a Route 1 high-risk system — the safety-zone controller in a cobot cell Bergmann builds and sells, making Bergmann the provider with a 2 August 2028 conformity deadline, integrated into its Machinery Regulation file. Two are Route 2: a vendor CV-screening tool (Annex III point 4, Bergmann as deployer, 2 December 2027) and a vendor shift-allocation tool, also point 4. Two more are limited-risk customer-facing chatbots under Article 50.

The governance verdict is clean because the routes are kept apart: provider duties on the cobot safety AI, deployer duties on the two HR tools, transparency on the chatbots, and an explicit check that confirms no emotion-recognition "fatigue monitor" has crept into the line — which would be prohibited.

How Confir helps

Confir runs both routes in sequence from one inventory: it asks whether a system is a safety component within Annex I machinery, then whether it is an Annex III use, derives the provider/deployer role under Article 25's value-chain logic, and assembles the matching obligations and deadline for each. The classification is deterministic and rule-based — same intake, same finding, the firing rule visible — so a mixed portfolio of operational, machinery-safety and HR AI is classified consistently rather than case by case. No consultants, no multi-month rollout.

Related guides

• Manufacturing under the EU AI Act
• Article 6 high-risk classification
• Provider vs deployer
• Recruitment screening AI: is it high-risk?
• Emotion recognition in the workplace
• Conformity assessment routes